No full-text available
To read the full-text of this research,
you can request a copy directly from the authors.
An artificial intelligence-based EEG algorithm for detection of epileptiform EEG discharges: Validation against the diagnostic gold standard
Abstract and Figures
Objective
To validate an artificial intelligence-based computer algorithm for detection of epileptiform EEG discharges (EDs) and subsequent identification of patients with epilepsy.
Methods
We developed an algorithm for automatic detection of EDs, based on a novel deep learning method that requires a low amount of labeled EEG data for training. Detected EDs are automatically grouped into clusters, consisting of the same type of EDs, for rapid visual inspection. We validated the algorithm on an independent dataset of 100 patients with sharp transients in their EEG recordings (54 with epilepsy and 46 with non-epileptic paroxysmal events). The diagnostic gold standard was derived from the video-EEG recordings of the patientś habitual events.
Results
The algorithm had a sensitivity of 89% for identifying EEGs with EDs recorded from patients with epilepsy, a specificity of 70%, and an overall accuracy of 80%.
Conclusions
Automated detection of EDs using an artificial intelligence-based computer algorithm had a high sensitivity. Human (expert) supervision is still necessary for confirming the clusters of detected EDs and for describing clinical correlations. Further studies on different patient populations will be needed to confirm our results.
Significance
The automated algorithm we describe here is a useful tool, assisting neurophysiologist in rapid assessment of EEG recordings.
Figures - uploaded by Franz Fürbass
Author content
All figure content in this area was uploaded by Franz Fürbass
Content may be subject to copyright.
... Deep learning (DL)-based methods have been widely developed and employed on EEG datasets for seizure and IED detection, achieving promising results [4][5][6][7] . However, most existing DL models are trained and validated on private datasets 5,6,8,9 , and their performance relies heavily on the quality, quantity, and distribution of the EEG signals in these datasets. ...
... Deep learning (DL)-based methods have been widely developed and employed on EEG datasets for seizure and IED detection, achieving promising results [4][5][6][7] . However, most existing DL models are trained and validated on private datasets 5,6,8,9 , and their performance relies heavily on the quality, quantity, and distribution of the EEG signals in these datasets. The use of private datasets restricts the reproducibility and independent verification of the results, making it challenging for other researchers to benchmark and compare different models. ...
... Many researchers have employed cross-validation to evaluate the performance of IED detection models 5,6,16 . In this scenario, the evaluation dataset contains a high ratio of IED recordings, which is different from the real-world EEG recordings that contain much more normal EEG data. ...
Interictal epileptiform discharge (IED) and its spatial distribution are critical for the diagnosis, classification, and treatment of epilepsy. Existing publicly available datasets suffer from limitations such as insufficient data amount and lack of spatial distribution information. In this paper, we present a comprehensive EEG dataset containing annotated interictal epileptic data from 84 patients, each contributing 20 minutes of continuous raw EEG recordings, totaling 28 hours. IEDs and states of consciousness (wake/sleep) were meticulously annotated by at least three EEG experts. The IEDs were categorized into five types based on occurrence regions: generalized, frontal, temporal, occipital, and centro-parietal. The dataset includes 2,516 IED epochs and 22,933 non-IED epochs, each 4 seconds long. We developed and validated a VGG-based model for IED detection using this dataset, achieving improved performance with the inclusion of consciousness and/or spatial distribution information. Additionally, our dataset serves as a reliable test set for evaluating and comparing existing IED detection models.
... 17 In EEG analysis, deep learning has been used to predict outcome in postanoxic coma, 18 detect and predict seizures, [19][20][21] or detect interictal discharges. [22][23][24] For IED detection algorithms to be used in clinical practice, performance should be similar to visual analysis by experts. 25,26 Assessing this is particularly challenging given the variability of IED detection between EEG experts. ...
... 5,27 In addition, the comparison of detection algorithms is hampered by the use of different data sets and reference standards. 22 For instance, in Fürbass et al., 22 the diagnostic gold standard was video-EEG with clinical episodes. In other works, the reference was defined as the agreement between experts using Cohen's or Fleiss' κ, depending on the number of raters, 28 whereas in Tveit et al., 24 for example, the reference was the majority consensus. ...
... 5,27 In addition, the comparison of detection algorithms is hampered by the use of different data sets and reference standards. 22 For instance, in Fürbass et al., 22 the diagnostic gold standard was video-EEG with clinical episodes. In other works, the reference was defined as the agreement between experts using Cohen's or Fleiss' κ, depending on the number of raters, 28 whereas in Tveit et al., 24 for example, the reference was the majority consensus. ...
Objective
Deep learning methods have shown potential in automating the detection of interictal epileptiform discharges (IEDs) in electroencephalography (EEG). We compared IED detection using our previously trained deep neural network with a group of experts to assess its potential applicability.
Methods
First, we performed clinical validation on an internal data set. Seven experts reviewed all EEG studies. Performance agreement between experts and the network was compared at both the EEG and IED levels. All EEG recordings were also processed with Persyst. Subsequently, we performed external validation, with data from four centers, using a hybrid approach, where detections by the deep neural network were reviewed by an expert. In case of disagreement with the original report, the EEG recording was annotated independently by five experts.
Results
For internal validation we included 22 EEG studies with IEDs and 28 EEG studies from controls. At the EEG level, our network showed performance similar to that of the experts. For individual IED detection, the sensitivities between experts ranged from 20.7%–86.4%, whereas the sensitivity of our network was 82.5% (confidence interval [CI]: 77.7%–87.4%) at 99% specificity and a false detection rate (FDR) of <.2/min, outperforming Persyst, with 64.6% sensitivity (CI: 61.4%–67.9%) at 98% specificity. External validation in 174 EEG studies demonstrated that all 85 EEG recordings classified as normal in the original report were classified correctly, with an FDR of .10/min. Of the 89 EEG studies with IEDs according to the report, 56 were correctly classified (Cohen's κ = .62). Visual analysis of the remaining 33 EEG recordings showed high interobserver variability among the five experts (Fleiss’ κ = .13).
Significance
Our deep neural network detects IEDs on par with clinical experts. The external validation in a hybrid approach showed substantial agreement with the original report. Disagreement was due mainly to high interobserver variability. Our deep neural network may support visual EEG analysis and assist in diagnostics, particularly when human resources are limited.
... DL's primary advantage lies in its ability to automatically extract features and achieve high classification accuracies, eliminating the need for the complex and time-intensive feature engineering associated with traditional ML algorithms. Numerous studies have explored the use of DL for seizure detection and prediction [30]- [43]. ...
... Convolutional Neural Networks (CNNs) have become the predominant DL approach in Epilepsy research for both seizure detection and prediction using EEG data [32]- [36], [40], [42], [43]. ...
... Continuous Wavelet Transform application for generating 2D time-frequency scalograms led to 93.6% accuracy in seizure classification in [42]. [43] employed the Fast R-CNN method for semi-supervised learning with various EEG data types, attaining an 80% accuracy in interpatient evaluation. ...
Epilepsy impacts around 6.38 per 1,000 people globally, presenting diagnostic challenges due to the complexity of seizures. Accurate classification of seizure types via Electroencephalogram (EEG) is critical for effective treatment and enhancing patient quality of life. However, the intricate characteristics of EEG data necessitate expert interpretation, a process that is both time-intensive and susceptible to human error. Recent advancements in Deep Learning (DL) have shown promise in EEG analysis, offering new avenues for seizure type classification. This study introduces two innovative DL architectures for seizure type classification: Network 1D Raw, which applies 1D Convolutions to Raw EEG signals, and Network 2D Conv, utilizing 2D convolutions on pre-computed spectrograms. Both architectures employ Separable Convolutions to enhance feature extraction efficiency, with the Network 1D Raw also incorporating a dilation rate technique for expanded analysis. Tested on the Temple University Hospital Seizure (TUSZ) dataset, these methods are evaluated using inter-patient 3-fold cross-validation. The Network 1D Raw achieved a weighted f1-score of 0.611 ± 0.037, while the Network 2D Conv reached 0.599 ± 0.052, both surpassing existing benchmarks. The Network 1D Raw demonstrated superior classification for Absence Seizure (ABSZ), Focal Non-specific Seizures (FNSZ), and Generalized Non-specific Seizures types (GNSZ), with Network 2D Conv excelling in FNSZ Seizures and GNSZ classes. Our DL models advance seizure type classification, blending efficiency with accuracy. Network 1D Raw’s compact design suits low-resource environments, aiding quick, precise diagnoses. Future work will focus on expanding dataset diversity, particularly for underrepresented seizure types, to further refine classification performance.
... There have been several approaches aiming to automate IED detection [13]. These range from thresholding of morphological or frequency features [14,15] to traditional machine learning methods [16,17] and, more recently, deep learning techniques [18][19][20][21][22][23]. The popularity of deep learning methods in the medical field has grown tremendously in the past years, with applications ranging from predicting mortality with echocardiographic video [24], interpretation of X-ray images [25,26], to skin cancer classification [27]. ...
... These considerations may also apply to deep learning approaches for IED detection, as networks are typically trained on a particular subset of EEG data. In particular, most studies discussing deep networks for IED detection use routine EEG recordings for training, with an average duration of 20-30 min [20][21][22][23] while in other works data length and acquisition conditions are not clearly specified [18]. ...
... One study used EEGs with a median length of 53 min to train a convolutional neural network (SpikeNet) [20]. Other algorithms do not specify the length of the training recordings [18,30,42]. ...
Objective: Deep learning methods have shown potential in automating interictal epileptiform discharge (IED) detection in electroencephalograms (EEGs). While it is known that these algorithms are dependent on the type of data used for training, this has not been explicitly explored in EEG analysis applications. We study the difference in performance of deep learning algorithms on routine and ambulatory EEGs. Methods: For training we used three datasets: i) 166 routine EEGs; ii) 75 ambulatory EEGs and iii) a combination of the two data types (241 EEGs). Routine EEGs were recorded in the hospital, ambulatory EEGs in the home environment, and included sleep. We trained a deep neural network (VGGC), on all three datasets, resulting in three deep nets, a VGGC-R for the routine EEGs, a VGGC-A for the ambulatory EEGs and a net that was trained on all data, VGGC-C. All three networks were subsequently tested using a test set that was comprised of 34 routine EEGs and 33 ambulatory recordings. For the evaluation, all 2 s non-overlapping epochs were labeled with a probability that expressed the likelihood of containing an epileptiform discharge. Performance was quantified as sensitivity, specificity and the rate of false positive detections (FPR). Results: The VGGC-R, had the best performance for routine EEGs, with 84% sensitivity at 99% specificity, however the sensitivity of this model was only 53% on ambulatory EEGs, with a specificity of 95% and FPR >3 FP/min. The networks that had been trained using only ambulatory data or all data, the VGGC-A and VGGC-C, yielded sensitivities in the test set comprised of ambulatory data of 60% and 79%, respectively, at 99% specificity, with a FPR of
... In a recent study, two neural networks trained for IED detection were tested in a fully automated and using a hybrid approach (Kural et al., 2022). In the fully automated approach, SpikeNet (Thomas et al., 2020) yielded 67% sensitivity at 63% specificity and Encevis (Fürbass et al., 2020) led to 97% sensitivity at 17% specificity. In the hybrid approach, experts looked at the clustered detections from the network and classified those as IEDs or non-IEDs. ...
... 20.6 ± 1. deep learning algorithms, clustered into IED types (Kural et al., 2022). This study included three methods (two deep learning approaches, SpikeNet (Thomas et al., 2020) and Encevis (Fürbass et al., 2020), as well as Persyst (Scheuer et al., 2017)) but it did not compare time reduction between algorithms. We have previously shown that algorithm performance is not necessarily maintained between routine and ambulatory data, so it is not clear how the results of the algorithms, as well as the time-burden reduction, would compare in ambulatory EEGs. ...
... Our aim was to assess the presence or absence of IEDs in EEG recordings based on automated detections, which is a task with high clinical relevance when developing an assistive tool to support decisions and streamline the visual analysis process of ambulatory recordings. This differs from the topics of previous works (Scheuer et al., 2017;da Silva Lourenço et al., 2021, 2023Fürbass et al., 2020;Scherg et al., 2012), which try to optimize performance for individual IED detection. While it is necessary for an automated approach to show sufficient performance for single IEDs (as several recently published approaches do), in the clinic, these methods will likely be applied to full EEG recordings. ...
Objective:
Interictal epileptiform discharges (IED) are hallmark biomarkers of epilepsy which are typically detected through visual analysis. Deep learning has shown potential in automating IED detection, which could reduce the burden of visual analysis in clinical practice. This is particularly relevant for ambulatory electroencephalograms (EEGs), as these entail longer review times.
Methods:
We applied a previously trained neural network to an independent dataset of 100 ambulatory EEGs (average duration 20.6 h). From these, 42 EEGs contained IEDs, 25 were abnormal without IEDs and 33 were normal. The algorithm flagged 2 second epochs that it considered IEDs. The EEGs were provided to an expert, who used NeuroCenter EEG to review the recordings. The expert concluded if each recording contained IEDs, and was timed during the process.
Results:
The conclusion of the reviewer was the same as the EEG report in 97% of the recordings. Three EEGs contained IEDs that were not detected based on the flagged epochs. Review time for the 100 EEGs was approximately 4 h, with half of the recordings taking <2 minutes to review.
Conclusions:
Our network can be used to reduce time spent on visual analysis in the clinic by 50-75 times with high reliability.
Significance:
Given the large time reduction potential and high success rate, this algorithm can be used in the clinic to aid in visual analysis.
... Two of them were using convolutional neural networks: SpikeNet 14 and the commercially available software package Encevis using the DeepSpike algorithm for IED detection. 15 The spike detector in the commercially available software program Persyst uses extracted features and a feed-forward neural network. 16 In the hybrid approach, clusters of IED candidate waves, detected by the algorithms, were visually evaluated by experts using the criteria in the operational definition of the International Federation of Clinical Neurophysiology (IFCN) for IEDs, 17 which were shown previously to provide high specificity, essential in clinical EEG reading. ...
... DeepSpike was developed using the Fast Region-based Convolutional Network method (Fast R-CNN). 15 It uses deep regression for estimating the position of EDs (negative peaks) 21,22 followed by classification of EDs. 15 Supervised and unsupervised learning was used to train DeepSpike. For supervised learning, 447 000 labeled EEG epochs from 166 patients and synthetic data sets were used. ...
... 15 It uses deep regression for estimating the position of EDs (negative peaks) 21,22 followed by classification of EDs. 15 Supervised and unsupervised learning was used to train DeepSpike. For supervised learning, 447 000 labeled EEG epochs from 166 patients and synthetic data sets were used. ...
... In a recent study, two neural networks trained for IED detection were tested in a fully automated and using a hybrid approach (Kural et al., 2022). In the fully automated approach, SpikeNet (Thomas et al., 2020) yielded 67% sensitivity at 63% specificity and Encevis (Fürbass et al., 2020) led to 97% sensitivity at 17% specificity. In the hybrid approach, experts looked at the clustered detections from the network and classified those as IEDs or non-IEDs. ...
... 20.6 ± 1. deep learning algorithms, clustered into IED types (Kural et al., 2022). This study included three methods (two deep learning approaches, SpikeNet (Thomas et al., 2020) and Encevis (Fürbass et al., 2020), as well as Persyst (Scheuer et al., 2017)) but it did not compare time reduction between algorithms. We have previously shown that algorithm performance is not necessarily maintained between routine and ambulatory data, so it is not clear how the results of the algorithms, as well as the time-burden reduction, would compare in ambulatory EEGs. ...
... Our aim was to assess the presence or absence of IEDs in EEG recordings based on automated detections, which is a task with high clinical relevance when developing an assistive tool to support decisions and streamline the visual analysis process of ambulatory recordings. This differs from the topics of previous works (Scheuer et al., 2017;da Silva Lourenço et al., 2021, 2023Fürbass et al., 2020;Scherg et al., 2012), which try to optimize performance for individual IED detection. While it is necessary for an automated approach to show sufficient performance for single IEDs (as several recently published approaches do), in the clinic, these methods will likely be applied to full EEG recordings. ...
... The average sensitivity of the network, 82.5%, was on the higher end of the sensitivities calculated between the experts (assuming one expert as classifying and one as ground truth). It was also signi cantly higher than that of Persyst, one of the industry standards, trained with hand-crafted features [24]. The average speci city of our network was > 99%, which implies a false detection rate < 10/hour (with epoch lengths of 2 s), which is most satisfactory. ...
... Validation of algorithms trained for IED detection with multiple human experts has been previously reported in the literature [24][25][26][27][28][29][30][31] DeepSpike, a deep network used in Encevis, a commercially available software, was tested against eight experts [24]. The algorithm reported a sensitivity of 81.6% at 46.4% speci city for detecting IEDs against non-epileptic paroxysmal events. ...
... Validation of algorithms trained for IED detection with multiple human experts has been previously reported in the literature [24][25][26][27][28][29][30][31] DeepSpike, a deep network used in Encevis, a commercially available software, was tested against eight experts [24]. The algorithm reported a sensitivity of 81.6% at 46.4% speci city for detecting IEDs against non-epileptic paroxysmal events. ...
Deep learning methods have shown potential in automating interictal epileptiform discharge (IED) detection in electroencephalograms (EEGs). To implement this in a clinical setting, it needs to have similar performance to visual assessment. We aim to compare a neural network trained for IED detection with a group of experts for validation and assessment of its potential applicability in a clinical setting.
We processed EEGs from 20 patients with focal and generalized epilepsy and 30 controls with a neural network. Seven experts labeled the IEDs in the dataset. Kappa scores, sensitivity and specificity were calculated. Network performance was compared to the experts at EEG and at IED level, as well as with an industry standard, Persyst.
For EEG level classification, the neural network showed a similar trend to the experts. For individual IED detection, sensitivities between experts ranged from 20.7–86.4%. The sensitivity of our network was 82.5% at 99.0% specificity, and it outperformed Persyst.
The network can detect IEDs in agreement with the experts with a high sensitivity and specificity. This indicates that the algorithm can potentially be used in the clinic to support visual EEG analysis or provide access to diagnostics if human resources are limited.
... There have been several approaches aiming to automate IED detection [10]. These range from thresholding of morphological or frequency features [11,12] to traditional machine learning methods [13,14] and, more recently, deep learning techniques [15][16][17][18][19][20]. The popularity of deep learning methods in the medical field has grown in the past years, with applications from predicting mortality from echocardiographic video [21] to skin cancer classification [22]. ...
... This limitation may also apply to deep learning approaches for IED detection, but this has been left unexplored in current EEG literature . Most studies that study deep networks for IED detection use routine EEG recordings for training, with an average duration of 20-30 minutes [17][18][19][20], while in other works data length and acquisition conditions are not clearly specified [16]. ...
... Other algorithms do not specify the length of the training recordings [16,25,33]. DeepSpike [16], a convolutional neural network embedded in the Encevis software, reported a sensitivity of 81.6% at 46.4% specificity for IED detection. ...
Objective
Deep learning methods have shown potential in automating interictal epileptiform discharge (IED) detection in electroencephalograms (EEGs). However, it is known that these algorithms are dependent on the type of data used for training and this is not currently explored in EEG analysis applications. We aim to explore the difference in performance of artificial neural networks on routine and ambulatory EEG data.
Methods
We trained the same neural network on three datasets: 166 routine EEGs (VGGC–R), 75 ambulatory EEGs (VGGC–R) and a combination of the two data types (VGGC-C, 241 EEGs total). These networks were tested on 34 routine EEGs and 33 ambulatory recordings. Sensitivity, specificity and false positive rate (FPR) were calculated at a 0.99 probability threshold.
Results
The VGGC-R led to 84% sensitivity at 99% specificity on the routine EEGs, but its sensitivity was only 53% on ambulatory EEGs, with FPR > 3 FP/min. The VGGC-Cand VGGC-A yielded sensitivities of 79% and 60%, respectively, at 99% specificity on ambulatory data, but their sensitivity was under 60% for routine EEGs.
Conclusion
We show that the VGGC-R should be used for routine recordings and the VGGC-C should be used for ambulatory recordings for IED detection.
Significance
As different networks work better for different types of data, algorithms should be trained with the same type of EEG data they will be applied to, either routine or ambulatory.
... The median number of EEGs with IEDs is 156 (IQR: 26 -344). There were 12 studies [11,87,88,90,90,91,93,[95][96][97][98][99] including normal EEG recordings (without IEDs) (median: 106; IQR: 67 -496). Of all studies, there were only 2 studies with more than 1,000 recordings in total [11,91]. ...
... Low pass filterings of 50 Hz (N=2) and 49 Hz (N=1) were also employed to remove the power line noise. Apart from bandpass filtering, the PureEEG algorithm [102] was used in 1 study [98] to remove artifacts based on a stochastic, spatio-temporal model using a linear minimum mean square error estimator. There were 5 choices of montage in studies: temporal central parasagittal (TCP), common average (CA), longitudinal bipolar (LB), source derivation (SD), and laplacian. ...
... The 10-20 system was the most popular recording setting (N=16). The 25-electrode array was used in 1 study [98]. We did not observe any significant channel selection methods apart from excluding 2 ear electrodes in 10 studies. ...
Epilepsy is the most common chronic neurological disorder. Clinical neurologists use Electroencephalography (EEG) to record the voltage fluctuations within the brain via surface scalp electrodes. The EEG recording of an epileptic patient may show interictal epileptiform discharges (IEDs), which are intermittent electrophysiological events occurring between 2 seizures. Visual analysis of the EEG recording for IED is a time-consuming process that might take a neurologist up to 4 hours. This research leverages clinical data from patients with epilepsy collected at the Alfred Health and Royal Melbourne over the past 10 years and investigated different deep learning architectures to automate this process.
... Most deep learning-based approaches use convolutional neural networks (CNNs) [15], [16], [17], [18], [19], [20], This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ ...
... For more information, see https://creativecommons.org/licenses/by/4.0/ [21], [22], [23], [24], recurrent neural networks (RNNs) [21], [25], [26], and hybrid methods [16], [21], [27] to identify IEDs on the scalp [15], [16], [17], [18], [19], [20], [21], [25], [27] and intracranial [22], [23], [24], [26] EEG recordings. To enhance IED detection ability, some studies further considered modification of conventional CNN architectures to extract spatial EEG features effectively from multiple scalp regions [7], [9]; multi-level morphological features and multichannel co-occurrences of IEDs in deep neural networks [10]; augmentation of synthetic IEDs using a generative adversarial network [19]; a combination of template-matching and CNN approaches [18]; and visualization techniques to examine important areas of input data in time [13] and time-frequency [18] domains. ...
... Recent studies on deep learning-based automated IED detection have reported high IED and non-IED binary classification performance for scalp [15], [16], [17], [18], [19], [20], [21], [25], [27] and intracranial [22], [23], [24], [26] EEG recordings. Some studies further evaluated their approaches for distinguishing between EEG recordings with and without IEDs. ...
Centrotemporal spike-waves (CTSWs) are typical interictal epileptiform discharges (IEDs) observed in centrotemporal regions in self-limited epilepsy with centrotemporal spikes (SLECTS). This study aims to develop a deep learning-based approach for automated detection of CTSWs in scalp electroencephalography (EEG) recordings of patients with SLECTS. To lower the substantial burden of IED annotation on clinicians, we simplified it by limiting IEDs to CTSWs because electroencephalographic patterns of CTSWs are known to be highly consistent. Two neurologists annotated 1672 CTSWs of 20 patients with SLECTS. Thereafter, we performed a two-level CTSW detection procedure: epoch-level and EEG-level. In the epoch-level detection, we constructed convolutional neural network-based classification models for CTSW and non-CTSW binary classification using the recordings of 20 patients and 20 controls. We then set the thresholds of the classification models for 100% specificity. In the EEG-level detection, we applied the threshold-adjusted classification models to the recordings of 50 patients and 50 controls that were not used in the epoch-level detection to distinguish between CTSW-positive (with one or more CTSWs) and CTSW-negative (with no CTSW) recordings based on the detection of CTSW presence. We obtained an average sensitivity, specificity, and accuracy of 99.8%, 98.4%, and 99.1%, respectively, with an average false detection rate of 0.19/hr for the controls. Our approach showed high detectability for CTSWs despite the simplified annotation process. We expect that the proposed CTSW detectors have potential clinical usefulness for efficiently reading EEGs and diagnosing SLECTS, and can significantly reduce the burden of IED annotation on clinicians.
... Recent investigations explored the use of quantitative scalp EEG analysis to assist the diagnosis of epilepsy, mainly based on the use of ML. [54][55][56][57][58][59] For example, SpikeNet, a deep neural network, was trained on a total of 9571 scalp EEG records (with and without spikes) to perform spike detection and showed performances compared to those achieved by fellowship-trained neurophysiology experts. 54 On the other side, DeepSpike was developed for the detection of epileptiform discharges based on multiple instance object detection and required a relatively low number of labeled training data. ...
... 54 On the other side, DeepSpike was developed for the detection of epileptiform discharges based on multiple instance object detection and required a relatively low number of labeled training data. 55 For a review regarding the use of DL for the detection of epileptiform discharges from scalp EEG, readers are referred to Ref. 59 In line with these investigations, Nadalin et al. (2021) trained a CNN for spike ripple detection based on recordings from a total of 34 subjects. 60 Matos et al. (2022) proposed a classifier for supporting the diagnosis of epilepsy, based on functional connectivity features of EEG in patients who had a first seizure. ...
Collaborative efforts between basic scientists, engineers, and clinicians are enabling translational epileptology. In this article, we summarize the recent advances presented at the International Conference for Technology and Analysis of Seizures (ICTALS 2022): (1) novel developments of structural magnetic resonance imaging; (2) latest electroencephalography signal‐processing applications; (3) big data for the development of clinical tools; (4) the emerging field of hyperdimensional computing; (5) the new generation of artificial intelligence (AI)–enabled neuroprostheses; and (6) the use of collaborative platforms to facilitate epilepsy research translation. We highlight the promise of AI reported in recent investigations and the need for multicenter data‐sharing initiatives.
... In their study, Franz Fürbass and colleagues developed an algorithm using deep learning and could accurately predict whether an individual had epilepsy or not with 80% success, based on EEG data [11]. ...
... 1 2 x Precision x Recall Precision + Recall (11) h) The ROC curve graphically shows the relationship between expressions of sensitivity and specificity for a data test. The ROC curve is obtained by plotting the false positive rates (1-specificity) values versus the sensitivity value. ...
This study uses machine learning to perform the hearing test (audiometry) processes autonomously with EEG signals. Sounds with different amplitudes and wavelengths given to the person tested in standard hearing tests are assigned randomly with the interface designed with MATLAB GUI. The person stated that he heard the random size sounds he listened to with head-phones but did not take action if he did not hear them. Simultaneously, EEG (electro-encephalog-raphy) signals were followed, and the waves created in the brain by the sounds that the person attended and did not hear were recorded. EEG data generated at the end of the test were pre-processed , and then feature extraction was performed. The heard and unheard information received from the MATLAB interface was combined with the EEG signals, and it was determined which sounds the person heard and which they did not hear. During the waiting period between the sounds given via the interface, no sound was given to the person. Therefore, these times are marked as not heard in EEG signals. In this study, brain signals were measured with Brain Products Vamp 16 EEG device, and then EEG raw data were created using the Brain Vision Recorder program and MATLAB. After the data set was created from the signal data produced by the heard and unheard sounds in the brain, machine learning processes were carried out with the PYTHON programming language. The raw data created with MATLAB was taken with the Python programming language, and after the pre-processing steps were completed, machine learning methods were applied to the classification algorithms. Each raw EEG data has been detected by the Count Vectorizer method. The importance of each EEG signal in all EEG data has been calculated using the TF-IDF (Term Frequency-Inverse Document Frequency) method. The obtained dataset has been classified according to whether people can hear the sound. Naïve Bayes, Light Gradient Strengthening Machine (LGBM), support vector machine (SVM), decision tree, k-NN, logistic regression, and random forest classifier algorithms have been applied in the analysis. The algorithms selected in our study were preferred because they showed superior performance in ML and succeeded in analyzing EEG signals. Selected classification algorithms also have features of being used online. Naïve Bayes, Light Gradient Strengthening Machine (LGBM), support vector machine (SVM), decision tree, k-NN, logistic regression, and random forest classifier algorithms were used. In the analysis of EEG signals, Light Gradient Strengthening Machine (LGBM) was obtained as the best method. It was determined that the most successful algorithm in prediction was the prediction of the LGBM classification algorithm , with a success rate of 84%. This study has revealed that hearing tests can also be performed using brain waves detected by an EEG device. Although a completely independent hearing test can be created, an audiologist or doctor may be needed to evaluate the results. Citation: Küçükakarsu, M.; Kavsaoğlu, A.R.; Alenezi, F.; Alhudhaif, A.; Alwadie, R.; Polat, K.
... Some of these algorithms have been validated in routine and critical care EEGs, achieving human expert level performance with improved efficiency and consistency. [16][17][18][19][20][21][22][23][24] As such, it is more than plausible that the future will involve appreciable adoption of these technologies. However, we need to consider the larger implications of implementation and have the foresight to address them in a meaningful way through proper education. ...
... Subsequently, the EEG files underwent automated conversion into DICOM EEG format based on a conversion routine by Sigma [3]. The DICOM EEG was then processed by the encevis AI-based automated spike detector [4], whose results were stored as a DICOM structured report (DICOM waveform annotation SR, currently in process of standardization). Both, the recorded EEG and spike detections were then made available in the encevis EEG viewer for review by clinicians. ...
Background:
Self-recorded EEG by patients at home might present a viable alternative to inpatient epilepsy evaluations.
Objectives and methods:
We developed a novel telemonitoring system comprising seamlessly integrated hard- and software with automated AI-based EEG analysis.
Results:
The first complete study participation results demonstrate feasibility and clinical utility.
Conclusion:
Our telemonitoring solution potentially improves treatment of patients with epilepsy and moreover might help to better distribute resources in the healthcare system.
... These results indicate that AI-powered computer-assisted EEG analysis could significantly improve the speed and precision of EEG assessments, thereby potentially enhancing treatment outcomes for epilepsy patients. Fürbass et al. [171] employed the Fast R-CNN method for object detection, using deep regression for localization estimation of EDs (negative peaks) and the UDA training process to handle noise and artefacts in EEG. The authors used EEG data from 590,000 epochs of 289 patients for unsupervised training and tested it against 100 proprietary datasets. ...
Disease diagnosis represents a critical and arduous endeavor within the medical field. Artificial intelligence (AI) techniques, spanning from machine learning and deep learning to large model paradigms, stand poised to significantly augment physicians in rendering more evidence-based decisions, thus presenting a pioneering solution for clinical practice. Traditionally, the amalgamation of diverse medical data modalities (e.g., image, text, speech, genetic data, physiological signals) is imperative to facilitate a comprehensive disease analysis, a topic of burgeoning interest among both researchers and clinicians in recent times. Hence, there exists a pressing need to synthesize the latest strides in multi-modal data and AI technologies in the realm of medical diagnosis. In this paper, we narrow our focus to five specific disorders (Alzheimer’s disease, breast cancer, depression, heart disease, epilepsy), elucidating advanced endeavors in their diagnosis and treatment through the lens of artificial intelligence. Our survey not only delineates detailed diagnostic methodologies across varying modalities but also underscores commonly utilized public datasets, the intricacies of feature engineering, prevalent classification models, and envisaged challenges for future endeavors. In essence, our research endeavors to contribute to the advancement of diagnostic methodologies, furnishing invaluable insights for clinical decision making.
... These extracted features were used in the classification of emotions by SVM and k-nearest neighbor algorithm (KNN) networks and the highest network accuracy was reported to be 81.45%. Yin et al. (2016) developed a suitable method for classifying emotions based on a deep learning model in stacked autoencoder networks to improve the deep learning network performance (Fürbass, 2020). They reported acceptable results with the highest accuracy of the dimensions of arousal and valence equal to 84.18% and 83.04%, respectively. ...
Introduction
The study explores the use of Electroencephalograph (EEG) signals as a means to uncover various states of the human brain, with a specific focus on emotion classification. Despite the potential of EEG signals in this domain, existing methods face challenges. Features extracted from EEG signals may not accurately represent an individual’s emotional patterns due to interference from time-varying factors and noise. Additionally, higher-level cognitive factors, such as personality, mood, and past experiences, further complicate emotion recognition. The dynamic nature of EEG data in terms of time series introduces variability in feature distribution and interclass discrimination across different time stages.
Methods
To address these challenges, the paper proposes a novel adaptive ensemble classification method. The study introduces a new method for providing emotional stimuli, categorizing them into three groups (sadness, neutral, and happiness) based on their valence-arousal (VA) scores. The experiment involved 60 participants aged 19–30 years, and the proposed method aimed to mitigate the limitations associated with conventional classifiers.
Results
The results demonstrate a significant improvement in the performance of emotion classifiers compared to conventional methods. The classification accuracy achieved by the proposed adaptive ensemble classification method is reported at 87.96%. This suggests a promising advancement in the ability to accurately classify emotions using EEG signals, overcoming the limitations outlined in the introduction.
Conclusion
In conclusion, the paper introduces an innovative approach to emotion classification based on EEG signals, addressing key challenges associated with existing methods. By employing a new adaptive ensemble classification method and refining the process of providing emotional stimuli, the study achieves a noteworthy improvement in classification accuracy. This advancement is crucial for enhancing our understanding of the complexities of emotion recognition through EEG signals, paving the way for more effective applications in fields such as neuroinformatics and affective computing.
... It is, furthermore, primarily reported during sleep and mostly in deeper layers of the cortex (Lam et al. 2017(Lam et al. , 2020Liedorp et al. 2010;Vossel et al. 2013). Besides the development of more accurate automatic spike detection algorithms (Furbass et al. 2020;Wilson and Emerson 2002), there is a need for alternative, quantitative, and more sensitive measures to timely capture abnormal E-I ratios in short noninvasive recordings. ...
A novel network version of permutation entropy, the inverted joint permutation entropy (JPEinv), holds potential as non-invasive biomarker of abnormal excitation-inhibition (E-I) ratio in Alzheimer’s disease (AD). In this computational modelling study, we test the hypotheses that this metric, and related measures of signal variability and functional connectivity, are sensitive to altered E-I ratios. The E-I ratio in each neural mass of a whole-brain computational network model was systematically varied. We evaluated whether JPEinv, local signal variability (by permutation entropy) and functional connectivity (by weighted symbolic mutual information (wsMI)) were related to E-I ratio, on whole-brain and regional level. The hub disruption index can identify regions primarily affected in terms of functional connectivity strength (or: degree) by the altered E-I ratios. Analyses were performed for a range of coupling strengths, filter and time-delay settings. On whole-brain level, higher E-I ratios were associated with higher functional connectivity (by JPEinv and wsMI) and lower local signal variability. These relationships were nonlinear and depended on the coupling strength, filter and time-delay settings. On regional level, hub-like regions showed a selective decrease in functional degree (by JPEinv and wsMI) upon a lower E-I ratio, and non-hub-like regions showed a selective increase in degree upon a higher E-I ratio. These results suggest that abnormal functional connectivity and signal variability, as previously reported in patients across the AD continuum, can inform us about altered E-I ratios.
Supplementary Information
The online version contains supplementary material available at 10.1007/s11571-023-10003-x.
... Affective computing is a representative field, which aims to give computer systems the ability to automatically recognize, understand and respond to human emotions, so as to realize intelligent human-computer interaction. As the core and important component of affective computing, emotion recognition has a wide range of applications in psychology, emotional computing, artificial intelligence, computer vision, medical, and other fields (Ramirez et al., 2001;Hu et al., 2019;Fürbass et al., 2020). ...
In this paper, a novel EEG emotion recognition method based on residual graph attention neural network is proposed. The method constructs a three-dimensional sparse feature matrix according to the relative position of electrode channels, and inputs it into the residual network to extract high-level abstract features containing electrode spatial position information. At the same time, the adjacency matrix representing the connection relationship of electrode channels is constructed, and the time-domain features of multi-channel EEG are modeled using graph. Then, the graph attention neural network is utilized to learn the intrinsic connection relationship between EEG channels located in different brain regions from the adjacency matrix and the constructed graph structure data. Finally, the high-level abstract features extracted from the two networks are fused to judge the emotional state. The experiment is carried out on DEAP data set. The experimental results show that the spatial domain information of electrode channels and the intrinsic connection relationship between different channels contain salient information related to emotional state, and the proposed model can effectively fuse these information to improve the performance of multi-channel EEG emotion recognition.
... Although (subclinical) epileptiform abnormalities are the most overt sign of neuronal hyperexcitability, which in turn is a result of E-I ratio changes, they occur infrequently, locally and predominantly during sleep [13][14][15] , and, therefore, challenges a correct detection. In addition, the vast number of automatic spike detection algorithms developed for neurophysiological data are not widely used in the clinic, yet, because their accuracy needs improvement [16][17][18] . Novel, quantitative surrogate markers of E-I balance are, thus, required. ...
An early disruption of neuronal excitation–inhibition (E–I) balance in preclinical animal models of Alzheimer’s disease (AD) has been frequently reported, but is difficult to measure directly and non-invasively in humans. Here, we examined known and novel neurophysiological measures sensitive to E–I in patients across the AD continuum. Resting-state magnetoencephalography (MEG) data of 86 amyloid-biomarker-confirmed subjects across the AD continuum (17 patients diagnosed with subjective cognitive decline, 18 with mild cognitive impairment (MCI) and 51 with dementia due to probable AD (AD dementia)), 46 healthy elderly and 20 young control subjects were reconstructed to source-space. E–I balance was investigated by detrended fluctuation analysis ( DFA ), a functional E/I ( fE/I ) algorithm, and the aperiodic exponent of the power spectrum. We found a disrupted E–I ratio in AD dementia patients specifically, by a lower DFA , and a shift towards higher excitation, by a higher fE/I and a lower aperiodic exponent. Healthy subjects showed lower fE/I ratios ( < 1.0) than reported in previous literature, not explained by age or choice of an arbitrary threshold parameter, which warrants caution in interpretation of fE/I results. Correlation analyses showed that a lower DFA (E–I imbalance) and a lower aperiodic exponent (more excitation) was associated with a worse cognitive score in AD dementia patients. In contrast, a higher DFA in the hippocampi of MCI patients was associated with a worse cognitive score. This MEG-study showed E–I imbalance, likely due to increased excitation, in AD dementia, but not in early stage AD patients. To accurately determine the direction of shift in E–I balance, validations of the currently used markers and additional in vivo markers of E–I are required.
... Ongoing investigations have endeavored to develop an automated technique that could efficiently detect IEDs in EEG recordings at an acceptable accuracy 8 . Recently, deep learning techniques have been widely accepted as the main strategy for building automated IED detectors for scalp [9][10][11][12][13][14][15][16][17] and intracranial [18][19][20][21] Epilepsy, a chronic disorder of the brain that causes recurrent spontaneous seizures, is categorized as focal or generalized. Recurrent seizures originating within a neuronal network limited to one hemisphere, unifocal or multifocal, are core features of focal epilepsy. ...
Detection and spatial distribution analyses of interictal epileptiform discharges (IEDs) are important for diagnosing, classifying, and treating focal epilepsy. This study proposes deep learning-based models to detect focal IEDs in electroencephalography (EEG) recordings of the frontal, temporal, and occipital scalp regions. This study included 38 patients with frontal (n = 15), temporal (n = 13), and occipital (n = 10) IEDs and 232 controls without IEDs from a single tertiary center. All the EEG recordings were segmented into 1.5-s epochs and fed into 1- or 2-dimensional convolutional neural networks to construct binary classification models to detect IEDs in each focal region and multiclass classification models to categorize IEDs into frontal, temporal, and occipital regions. The binary classification models exhibited accuracies of 79.3–86.4%, 93.3–94.2%, and 95.5–97.2% for frontal, temporal, and occipital IEDs, respectively. The three- and four-class models exhibited accuracies of 87.0–88.7% and 74.6–74.9%, respectively, with temporal, occipital, and non-IEDs F1-scores of 89.9–92.3%, 84.9–90.6%, and 84.3–86.0%; and 86.6–86.7%, 86.8–87.2%, and 67.8–69.2% for the three- and four-class (frontal, 50.3–58.2%) models, respectively. The deep learning-based models could help enhance EEG interpretation. Although they performed well, the resolution of region-specific focal IED misinterpretations and further model improvement are needed.
... These specificities are higher than reported in previous studies [14,15,16,17], probably due to the fact that we did not review single IEDs, but measured whether 10-second epochs and 30-minute selections contained IEDs. ...
Objective:
Automated detection of spikes and seizures has been a subject of research for several decades now. There have been important advances, yet automated detection in EMU (Epilepsy Monitoring Unit) settings has not been accepted as standard practice. We intend to implement this software at our EMU and so carried out a qualitative study to identify factors that hinder ('barriers') and facilitate ('enablers') implementation.
Method:
Twenty-two semi-structured interviews were conducted with 14 technicians and neurologists involved in recording and reporting EEGs and eight neurologists who receive EEG reports in the outpatient department. The study was reported according to the Consolidated Criteria for Reporting Qualitative Studies (COREQ).
Results:
We identified 14 barriers and 14 enablers for future implementation. Most barriers were reported by technicians. The most prominent barrier was lack of trust in the software, especially regarding seizure detection and false positive results. Additionally, technicians feared losing their EEG review skills or their jobs. Most commonly reported enablers included potential efficiency in the EEG workflow, the opportunity for quantification of EEG findings and the willingness to try the software.
Conclusions:
This study provides insight into the perspectives of users and offers recommendations for implementing automated spike and seizure detection in EMUs.
... Accurately identifying emotions is crucial in many contexts. Emotion detection research has recently found widespread use in fields as diverse as medicine, computer vision, artificial intelligence, psychology, and neuroscience [3]. For instance, diagnosing mental depression and schizophrenia can benefit from emotion recognition technology. ...
... 22 The average level of pairwise spike marking sensitivities of three skilled human EEG readers was about 45%, and comparable to a sensitivity of 43.9% of Persyst. Very recently, an artificial intelligence-based computer algorithm had a sensitivity of 89% for identifying EEGs with IEDs recorded from patients with epilepsy, but human expert supervision was still necessary for confirming the clusters of detected IEDs. 29 In previous studies, the reported average pairwise reader spike sensitivities were comparably high with 52% F I G U R E 4 ESI results of an MRI positive patient with 256-channel hdEEG. The patient was Engel class 1A at 1-year follow-up after anterior temporal lobectomy. ...
Objective
Presurgical high‐density electric source imaging (hdESI) of interictal epileptic discharges (IEDs) is only used by few epilepsy centers. One obstacle is the time‐consuming workflow both for recording as well as for visual review. Therefore, we analyzed the effect of (a) an automated IED detection and (b) the number of IEDs on the accuracy of hdESI and time‐effectiveness.
Methods
In 22 patients with pharmacoresistant focal epilepsy receiving epilepsy surgery (Engel 1) we retrospectively detected IEDs both visually and semi‐automatically using the EEG analysis software Persyst in 256‐channel EEGs. The amount of IEDs, the Euclidean distance between hdESI maximum and resection zone, and the operator time were compared. Additionally, we evaluated the intra‐individual effect of IED quantity on the distance between hdESI maximum of all IEDs and hdESI maximum when only a reduced amount of IEDs were included.
Results
There was no significant difference in the number of IEDs between visually versus semi‐automatically marked IEDs (74 ± 56 IEDs/patient vs 116 ± 115 IEDs/patient). The detection method of the IEDs had no significant effect on the mean distances between resection zone and hdESI maximum (visual: 26.07 ± 31.12 mm vs semi‐automated: 33.6 ± 34.75 mm). However, the mean time needed to review the full datasets semi‐automatically was shorter by 275 ± 46 min (305 ± 72 min vs 30 ± 26 min, P < 0.001).
The distance between hdESI of the full versus reduced amount of IEDs of the same patient was smaller than 1 cm when at least a mean of 33 IEDs were analyzed. There was a significantly shorter intraindividual distance between resection zone and hdESI maximum when 30 IEDs were analyzed as compared to the analysis of only 10 IEDs (P < 0.001).
Significance
Semi‐automatized processing and limiting the amount of IEDs analyzed (~30–40 IEDs per cluster) appear to be time‐saving clinical tools to increase the practicability of hdESI in the presurgical work‐up.
... Sophisticated algorithms detect spikes as accurately as clinicians at speeds impossible for humans. [10][11][12] These advances allow us to do what the human eye cannot: we can summarize the relative frequency of spikes in different brain regions, and how they change over time 13 ; we can measure the precise timing of when a spike reaches different electrodes (Figure 1) 14-17 ; and we can compare spike morphology across space and time. 4 Applying these tools has taught us that not all spikes are equal as surgical biomarkers. ...
Electroencephalography (EEG) has been the primary diagnostic tool in clinical epilepsy for nearly a century. Its review is performed using qualitative clinical methods that have changed little over time. However, the intersection of higher resolution digital EEG and analytical tools developed in the past decade invites a re‐exploration of relevant methodology. In addition to the established spatial and temporal markers of spikes and high‐frequency oscillations, novel markers involving advanced postprocessing and active probing of the interictal EEG are gaining ground. This review provides an overview of the EEG‐based passive and active markers of cortical excitability in epilepsy and of the techniques developed to facilitate their identification. Several different emerging tools are discussed in the context of specific EEG applications and the barriers we must overcome to translate these tools into clinical practice.
... Motivated by its success in computer vision and natural language processing, employing deep learning techniques for IED detection has become popular and demonstrates promising performance [12]. Recent proposals cover Convolutional Neural Networks (CNN) [1,15,24,25,53,76,85], Long Short-Term Memory (LSTM) [56] and Generative Adversarial Networks (GAN) [26,84]. According to recent studies on data series classication, CNN models (e.g., ResNet [21] and Inception-Time [22]) generally achieve the SOTA performance [21]. ...
Epilepsy is a chronic neurological disease, ranked as the second most burdensome neurological disorder worldwide. Detecting Interictal Epileptiform Discharges (IEDs) is among the most important clinician operations to support epilepsy diagnosis, rendering automatic IED detection based on electroencephalography (EEG) signals an important topic. However, most existing solutions were designed and evaluated upon artificially balanced IED datasets, which do not conform to the real-world highly imbalanced scenarios. In this work, we propose the iEDeaL framework for automatic IED detection in challenging real-world use cases. The main components of iEDeaL are the new SC neural network architecture, to efficiently detect IEDs on raw EEG series instead of extracted features, and SaSu, a novel loss function to train SC by optimizing the F β -score. Experiments on two real-world imbalanced IED datasets verify the advantages of iEDeaL in offering more accurate and efficient IED detection when compared with other state-of-the-art deep learning-based and spectrogram feature-based solutions.
... Ongoing investigations have endeavored to develop an automated technique that could e ciently detect IEDs in EEG recordings at an acceptable accuracy [8] . Recently, deep learning techniques have been widely accepted as the main strategy for building automated IED detectors for scalp [9][10][11][12][13][14][15][16][17] and intracranial [18][19][20][21] EEG recordings. ...
Detection and spatial distribution analysis of interictal epileptiform discharges (IEDs) are important for diagnosing, classifying, and treating focal epilepsy. This study proposes deep learning-based models to detect focal IEDs in electroencephalography (EEG) recordings of the frontal, temporal, and occipital scalp regions. This study included 38 patients with frontal ( n = 15), temporal ( n = 13), or occipital ( n = 10) IEDs and 232 controls without IEDs from a single tertiary center. All EEG recordings were segmented into 1.5-s epochs and fed into 1- or 2-dimensional convolutional neural networks to construct binary models to detect IEDs in each focal region and multiclass models to categorize IEDs into frontal, temporal, and occipital regions. The binary models exhibited accuracies of 79.3–86.4%, 93.3–94.2%, and 95.5–97.2% for frontal, temporal, and occipital IEDs, respectively. The three and four multiclass models exhibited an accuracy of 87.0–88.7% and 74.6–74.9%, respectively, with temporal, occipital, and non-IEDs F1-scores of 89.9–92.3%, 84.9–90.6%, and 84.3–86.0% and 86.6–86.7%, 86.8–87.2%, and 67.8–69.2% for the three- and four-class (frontal, 50.3–58.2%) models, respectively. The constructed deep learning-based models could help enhance EEG interpretation. Although they performed well, the resolution of region-specific focal IED misinterpretations and further model improvement are needed.
Epilepsy, a neurological condition affecting approximately 50 million individuals globally, is among the most common nervous system disorders. Electroencephalography (EEG) is vital for evaluating epilepsy, yet its intricate nature often restricts its application to specialized clinical environments. OptiEEG, a remote monitoring system incorporating OpenBCI’s EEG technology, addresses challenges by integrating a communication gateway and a mobile application for user-friendly operation. This study benchmarks OptiEEG’s performance against the clinically validated Natus NicoletOne EEG System through three routine EEG tests: Eye Open Close, Hyperventilation, and Photic Stimulation. Signal quality, component analysis, and reliability were evaluated using error metrics, time-frequency analysis, Bland-Altman plots, repeatability, Pearson Correlation Coefficient (PCC) and also EEG characteristics analysis of individual channels. OptiEEG demonstrated comparable signal quality to Natus, with average standard deviations for signal-to-noise ratio (Natus: 3.27 vs. OptiEEG: 2.95), peak signal-to-noise ratio (Natus: 2.76 vs. OptiEEG: 2.16), and mean squared error (Natus: 0.01 vs. OptiEEG: 0.04). Time-frequency analysis revealed less than 10% differences across alpha, theta, and delta bands. Reliability tests confirmed repeatability, with intra-system differences lower than inter-system differences, and Bland-Altman plots meeting 83% agreement criteria. PCC analysis highlighted moderate signal alignment, confirming similar EEG patterns across systems. Channel-specific analysis showed median differences as low as 0.80%, validating OptiEEG’s ability to capture critical EEG features. The results establish OptiEEG as a reliable alternative to traditional systems, combining clinically comparable performance with a portable design. These findings highlight its potential as a robust remote monitoring tool for epilepsy, enabling broader access to EEG diagnostics and management.
The capacity of AI systems to recognize and interpret human emotions, actions, and gestures is reshaping numerous sectors, from entertainment to security. This chapter offers a comprehensive review of the current state-of-the-art technologies in this domain, shedding light on their strengths, potential limitations, and avenues for improvement. In recent years, systems capable of recognizing and understanding human emotions, actions, and gestures have shown remarkable progress. They are deployed in diverse applications, including virtual reality, healthcare, and human-computer interaction. However, understanding their capabilities and limitations is crucial to harness their potential fully. Our research provides a deep dive into the strengths of existing AI systems, showcasing their ability to accurately decipher complex human expressions, movements, and emotional states. We also critically examine potential limitations, such as bias in training data or challenges in recognizing subtle cultural nuances.
A growing number of studies apply deep neural networks (DNNs) to recordings of human electroencephalography (EEG) to identify a range of disorders. In many studies, EEG recordings are split into segments, and each segment is randomly assigned to the training or test set. As a consequence, data from individual subjects appears in both the training and the test set. Could high test-set accuracy reflect data leakage from subject-specific patterns in the data, rather than patterns that identify a disease? We address this question by testing the performance of DNN classifiers using segment-based holdout (in which segments from one subject can appear in both the training and test set), and comparing this to their performance using subject-based holdout (where all segments from one subject appear exclusively in either the training set or the test set). In two datasets (one classifying Alzheimer's disease, and the other classifying epileptic seizures), we find that performance on previously-unseen subjects is strongly overestimated when models are trained using segment-based holdout. Finally, we survey the literature and find that the majority of translational DNN-EEG studies use segment-based holdout. Most published DNN-EEG studies may dramatically overestimate their classification performance on new subjects.
A core aim of neurocritical care is to prevent secondary brain injury. Spreading depolarizations (SDs) have been identified as an important independent cause of secondary brain injury. SDs are usually detected using invasive electrocorticography recorded at high sampling frequency. Recent pilot studies suggest a possible utility of scalp electrodes generated electroencephalogram (EEG) for non-invasive SD detection. However, noise and attenuation of EEG signals makes this detection task extremely challenging. Previous methods focus on detecting temporal power change of EEG over a fixed high-density map of scalp electrodes, which is not always clinically feasible. Having a specialized spectrogram as an input to the automatic SD detection model, this study is the first to transform SD identification problem from a detection task on a 1-D time-series wave to a task on a sequential 2-D rendered imaging. This study presented a novel ultra-light-weight multi-modal deep-learning network to fuse EEG spectrogram imaging and temporal power vectors to enhance SD identification accuracy over each single electrode, allowing flexible EEG map and paving the way for SD detection on ultra-low-density EEG with variable electrode positioning. Our proposed model has an ultra-fast processing speed (<0.3 sec). Compared to the conventional methods (2 hours), this is a huge advancement towards early SD detection and to facilitate instant brain injury prognosis. Seeing SDs with a new dimension – frequency on spectrograms, we demonstrated that such additional dimension could improve SD detection accuracy, providing preliminary evidence to support the hypothesis that SDs may show implicit features over the frequency profile.
A growing number of studies use deep neural networks (DNNs) to identify diseases from recordings of brain activity. DNN studies of electroencephalography (EEG) typically use cross-validation to test how accurately a model can predict the disease state of held-out test data. In these studies, segments of EEG data are often randomly assigned to the training or test sets. As a consequence, data from individual subjects appears in both training and test data. Could high test-set accuracy reflect leakage from subject-specific representations, rather than patterns that identify a disease? We address this question by testing the performance of DNN classifiers using segment-based holdout (where EEG segments from one subject can appear in both the training and test sets), and comparing this to their performance using subject-based holdout (where individual subjects' data appears exclusively in either the training set or the test set). We compare segment-based and subject-based holdout in two EEG datasets: one classifying Alzheimer's disease, and the other classifying epileptic seizures. In both datasets, we find that performance on previously-unseen subjects is strongly overestimated when models are trained using segment-based holdout. Next, we survey the literature and find that the majority of translational DNN-EEG studies use segment-based holdout, and therefore overestimate model performance on new subjects. In a hospital or doctor's office, clinicians need to diagnose new patients whose data was not used in training the model; segment-based holdout, therefore, does not reflect the real-world performance of a translational DNN model. When evaluating how DNNs could be used for medical diagnosis, models must be tested on subjects whose data was not included in the training set.
Brain interictal epileptiform discharges (IEDs), as one of the hallmarks of epileptic brain, are transient events captured by electroencephalogram (EEG). IEDs are generated by seizure networks, and they occur between seizures (interictal periods). The development of a robust method for IED detection could be highly informative for clinical treatment procedures and epileptic patient management. Since 1972, different machine learning techniques, from template matching to deep learning, have been developed to automatically detect IEDs from scalp EEG (scEEG) and intracranial EEG (iEEG). While the scEEG signals suffer from low information details and high attenuation of IEDs due to the high skull electrical impedance, the iEEG signals recorded using implanted electrodes enjoy higher details and are more suitable for identifying the IEDs. In this review paper, we group IED detection techniques into six categories: (1) template matching, (2) feature representation (mimetic, time-frequency, and nonlinear features), (3) matrix decomposition, (4) tensor factorization, (5) neural networks, and (6) estimation of the iEEG from the concurrent scEEG followed by detection and classification. The methods are compared quantitatively (e.g., in terms of accuracy, sensitivity, and specificity), and their general advantages and limitations are described. Finally, current limitations and possible future research paths related to this field are mentioned.
The objective of this work was to develop a deep learning-based automatic system with reliable performance in detecting interictal epileptiform discharges (IEDs) from scalp electroencephalograms (EEGs). For the present study, 484 raw scalp EEG recordings were included, standardized, and split into 406 for training and 78 for testing. Two neurophysiologists individually annotated the recordings for training in channel-wise manner. Annotations were divided into segments, on which nine deep neural networks (DNNs) were trained for the multiclassification of IED, artifact, and background. The fitted IED detectors were then evaluated on 78 EEG recordings with IED events fully annotated by three experts independently (majority agreement). A two montage-based decision mechanism (TMDM) was designed to determine whether an IED event occurred at a single time instant. Area under the precision–recall curve (AUPRC), as well as false-positive rates, F1 scores, and kappa agreement scores for sensitivity = 0.8 were estimated. In multitype classification, five DNNs provided one-versus-rest AUPRC mean value >0.993 using fivefold cross-validation. In IED detection, the system that had integrated the temporal convolutional network (TCN)-based IED detector and the TMDM rule achieved an AUPRC of 0.811. The false positive was 0.194/min (11.64/h), and the F1 score was 0.745. The agreement score between the system and the experts was 0.905. The proposed framework provides a TCN-based IED detector and a novel two montage-based determining mechanism that combined to make an automatic IED detection system. The system would be useful in aiding clinic EEG interpretation.
Importance:
Electroencephalograms (EEGs) are a fundamental evaluation in neurology but require special expertise unavailable in many regions of the world. Artificial intelligence (AI) has a potential for addressing these unmet needs. Previous AI models address only limited aspects of EEG interpretation such as distinguishing abnormal from normal or identifying epileptiform activity. A comprehensive, fully automated interpretation of routine EEG based on AI suitable for clinical practice is needed.
Objective:
To develop and validate an AI model (Standardized Computer-based Organized Reporting of EEG-Artificial Intelligence [SCORE-AI]) with the ability to distinguish abnormal from normal EEG recordings and to classify abnormal EEG recordings into categories relevant for clinical decision-making: epileptiform-focal, epileptiform-generalized, nonepileptiform-focal, and nonepileptiform-diffuse.
Design, setting, and participants:
In this multicenter diagnostic accuracy study, a convolutional neural network model, SCORE-AI, was developed and validated using EEGs recorded between 2014 and 2020. Data were analyzed from January 17, 2022, until November 14, 2022. A total of 30 493 recordings of patients referred for EEG were included into the development data set annotated by 17 experts. Patients aged more than 3 months and not critically ill were eligible. The SCORE-AI was validated using 3 independent test data sets: a multicenter data set of 100 representative EEGs evaluated by 11 experts, a single-center data set of 9785 EEGs evaluated by 14 experts, and for benchmarking with previously published AI models, a data set of 60 EEGs with external reference standard. No patients who met eligibility criteria were excluded.
Main outcomes and measures:
Diagnostic accuracy, sensitivity, and specificity compared with the experts and the external reference standard of patients' habitual clinical episodes obtained during video-EEG recording.
Results:
The characteristics of the EEG data sets include development data set (N = 30 493; 14 980 men; median age, 25.3 years [95% CI, 1.3-76.2 years]), multicenter test data set (N = 100; 61 men, median age, 25.8 years [95% CI, 4.1-85.5 years]), single-center test data set (N = 9785; 5168 men; median age, 35.4 years [95% CI, 0.6-87.4 years]), and test data set with external reference standard (N = 60; 27 men; median age, 36 years [95% CI, 3-75 years]). The SCORE-AI achieved high accuracy, with an area under the receiver operating characteristic curve between 0.89 and 0.96 for the different categories of EEG abnormalities, and performance similar to human experts. Benchmarking against 3 previously published AI models was limited to comparing detection of epileptiform abnormalities. The accuracy of SCORE-AI (88.3%; 95% CI, 79.2%-94.9%) was significantly higher than the 3 previously published models (P < .001) and similar to human experts.
Conclusions and relevance:
In this study, SCORE-AI achieved human expert level performance in fully automated interpretation of routine EEGs. Application of SCORE-AI may improve diagnosis and patient care in underserved areas and improve efficiency and consistency in specialized epilepsy centers.
Backgraund: Active implementation of artificial intelligence technologies in the healthcare in recent years promotes increasing amount of medical data for the development of machine learning models, including radiology and instrumental diagnostics data. To solve various problems of digital medical technologies, new datasets are being created through machine learning algorithms, therefore, the problems of their systematization and standardization, storage, access, rational and safe use become actual. A i m : development of an approach to systematization and standardization of information about datasets to represent, store, apply and optimize the use of datasets and ensure the safety and transparency of the development and testing of medical devices using artificial intelligence. M a t e r i a l s a n d m e t h o d s : analysis of own and international experience in the creation and use of medical datasets, medical reference books searching and analysis, registry structure development and justification, scientific publications search with the keywords “datasets”, “registry of medical data”, placed in the databases of the RSCI, Scopus, Web of Science. R e s u l t s . The register of medical instrumental diagnostics datasets structure has been developed in accordance with stages of datasets lifecycle: 7 parameters at the initiation stage, 8 – at the planning stage, 70 – dataset card, 1 – version change, 14 – at the use stage, total – 100 parameters. We propose datasets classification according to the purpose of their creation, a classification of data verification methods, as well as the principles of forming names for standardization and datasets presentation clarity. In addition, the main features of the organization of maintaining this registry are highlighted: management, data quality, confidentiality and security. C o n c l u s i o n s . For the first time, an original technology of medical datasets for instrumental diagnostics structuring and systematization is proposed. It is based on the developed terminology and principles of information classification. This makes it possible to standardize the structure of information about datasets for machine learning, and ensures the storage centralization. It also allows to get quick access to all information about the dataset, and ensure transparency, reliability and reproducibility of artificial intelligence developments. Creating a registry makes it possible to quickly form visual data libraries. This allows a wide range of researchers, developers and companies to choose data sets for their tasks. This approach ensures their widespread use, resource optimization and contributes to the rapid development and implementation of artificial intelligence.
The Scalp Electroencephalogram (EEG) signal of epileptic patients often contains Interictal Epileptiform Discharges (IED) during the period of seizures. Detection of IEDs is significant for the diagnosis of epilepsy and the prediction of seizures. In this paper, we proposed a graph convolutional network and bi-directional LSTM co-embedded broad learning system to detect IEDS. Here, we represent EEG signal as a graph and utilize Graph Convolutional Networks (GCN) to extract contextual features. In addition, bi-directional LSTM is also adopted for extracting the temporal feature from signals. Then these features are incorporated into Broad Learning System (BLS) to automatically detect epileptiform events. Experimental results indicate the proposed approach can achieve superior accuracy in the classification of IEDs than other commonly used time series processing models and reach a consensus with neurologists in predicting the lead of an EEG recording.
Epilepsy is a chronic brain disorder characterized by recurrent unprovoked seizures. It is caused by alterations in normal electrical activity in the brain, leading to various clinical manifestations depending on the regions that are affected. Scalp electroencephalography (EEG) is an important tool in the diagnosis of epilepsy. It provides data that can pinpoint the foci of epileptiform disturbances and can characterize the epilepsy syndrome. However, providing a timely review of EEG data by clinical experts is a tedious and error-prone exercise. Moreover, there is a disparity in the global and national distribution of EEG experts. In order to assist EEG experts in reading EEGs, machine learning techniques can serve as valuable clinical tools to analyze EEG data in an objective and computationally efficient manner. Such methods have been developed mainly for two purposes in the context of epilepsy: for the detection of interictal epileptiform discharges (IED) and for the detection of electrographical epileptic seizures. Our aim is to concisely review state-of-the-art machine learning methods for IED and seizure detection, to elaborate on existing drawbacks and challenges for such approaches, and to provide guidance to physicians and researchers when designing an automated algorithm for the annotation of epileptic EEG. Furthermore, this chapter will outline potential future directions and opportunities for research in the diagnosis and monitoring of epilepsy from EEG recordings.
Machine learning leverages statistical and computer science principles to develop algorithms capable of improving performance through interpretation of data rather than through explicit instructions. Alongside widespread use in image recognition, language processing, and data mining, machine learning techniques have received increasing attention in medical applications, ranging from automated imaging analysis to disease forecasting. This review examines the parallel progress made in epilepsy, highlighting applications in automated seizure detection from electroencephalography (EEG), video, and kinetic data, automated imaging analysis and pre‐surgical planning, prediction of medication response, and prediction of medical and surgical outcomes using a wide variety of data sources. A brief overview of commonly used machine learning approaches, as well as challenges in further application of machine learning techniques in epilepsy, is also presented. With increasing computational capabilities, availability of effective machine learning algorithms, and accumulation of larger datasets, clinicians and researchers will increasingly benefit from familiarity with these techniques and the significant progress already made in their application in epilepsy.
The rapid development of deep learning, a family of machine learning techniques, has spurred much interest in its application to medical imaging problems. Here, we develop a deep learning algorithm that can accurately detect breast cancer on screening mammograms using an “end-to-end” training approach that efficiently leverages training datasets with either complete clinical annotation or only the cancer status (label) of the whole image. In this approach, lesion annotations are required only in the initial training stage, and subsequent stages require only image-level labels, eliminating the reliance on rarely available lesion annotations. Our all convolutional network method for classifying screening mammograms attained excellent performance in comparison with previous methods. On an independent test set of digitized film mammograms from the Digital Database for Screening Mammography (CBIS-DDSM), the best single model achieved a per-image AUC of 0.88, and four-model averaging improved the AUC to 0.91 (sensitivity: 86.1%, specificity: 80.1%). On an independent test set of full-field digital mammography (FFDM) images from the INbreast database, the best single model achieved a per-image AUC of 0.95, and four-model averaging improved the AUC to 0.98 (sensitivity: 86.7%, specificity: 96.1%). We also demonstrate that a whole image classifier trained using our end-to-end approach on the CBIS-DDSM digitized film mammograms can be transferred to INbreast FFDM images using only a subset of the INbreast data for fine-tuning and without further reliance on the availability of lesion annotations. These findings show that automatic deep learning methods can be readily trained to attain high accuracy on heterogeneous mammography platforms, and hold tremendous promise for improving clinical tools to reduce false positive and false negative screening mammography results. Code and model available at: https://github.com/lishen/end2end-all-conv.
The development of decision support systems for pathology and their deployment in clinical practice have been hindered by the need for large manually annotated datasets. To overcome this problem, we present a multiple instance learning-based deep learning system that uses only the reported diagnoses as labels for training, thereby avoiding expensive and time-consuming pixel-wise manual annotations. We evaluated this framework at scale on a dataset of 44,732 whole slide images from 15,187 patients without any form of data curation. Tests on prostate cancer, basal cell carcinoma and breast cancer metastases to axillary lymph nodes resulted in areas under the curve above 0.98 for all cancer types. Its clinical application would allow pathologists to exclude 65–75% of slides while retaining 100% sensitivity. Our results show that this system has the ability to train accurate classification models at unprecedented scale, laying the foundation for the deployment of computational decision support systems in clinical practice.
Objectives:
Electroencephalography (EEG) is a central part of the medical evaluation for patients with neurological disorders. Training an algorithm to label the EEG normal vs abnormal seems challenging, because of EEG heterogeneity and dependence of contextual factors, including age and sleep stage. Our objectives were to validate prior work on an independent data set suggesting that deep learning methods can discriminate between normal vs abnormal EEGs, to understand whether age and sleep stage information can improve discrimination, and to understand what factors lead to errors.
Methods:
We train a deep convolutional neural network on a heterogeneous set of 8522 routine EEGs from the Massachusetts General Hospital. We explore several strategies for optimizing model performance, including accounting for age and sleep stage.
Results:
The area under the receiver operating characteristic curve (AUC) on an independent test set (n = 851) is 0.917 marginally improved by including age (AUC = 0.924), and both age and sleep stages (AUC = 0.925), though not statistically significant.
Conclusions:
The model architecture generalizes well to an independent dataset. Adding age and sleep stage to the model does not significantly improve performance.
Significance:
Insights learned from misclassified examples, and minimal improvement by adding sleep stage and age suggest fruitful directions for further research.
Automatic computer-based seizure detection and warning devices are important for objective seizure documentation, for SUDEP prevention, to avoid seizure related injuries and social embarrassments as a consequence of seizures, and to develop on demand epilepsy therapies. Automatic seizure detection systems can be based on direct analysis of epileptiform discharges on scalp-EEG or intracranial EEG, on the detection of motor manifestations of epileptic seizures using surface electromyography (sEMG), accelerometry (ACM), video detection systems and mattress sensors and finally on the assessment of changes of physiologic parameters accompanying epileptic seizures measured by electrocardiography (ECG), respiratory monitors, pulse oximetry, surface temperature sensors, and electrodermal activity. Here we review automatic seizure detection based on scalp-EEG, ECG, and sEMG. Different seizure types affect preferentially different measurement parameters. While EEG changes accompany all types of seizures, sEMG and ACM are suitable mainly for detection of seizures with major motor manifestations. Therefore, seizure detection can be optimized by multimodal systems combining several measurement parameters. While most systems provide sensitivities over 70%, specificity expressed as false alarm rates still needs to be improved. Patients' acceptance and comfort of a specific device are of critical importance for its long-term application in a meaningful clinical way.
This work deals with computer aided monitoring of electrical brain signals to detect disease-related
patterns. Severe neurological disorders can trigger unusually strong firing of brain cells that
distinguishes clearly from normal brain activity. A well-known example of such a disease is epilepsy.
During an epileptic seizure, fast repeating electrical discharges on the head surface are often
measurable. Epileptic seizures usually occur rarely or unnoticed by patients during night wherefore
considerable effort is needed to properly evaluate and treat patients. Besides epilepsy, inflammatory
brain diseases such as encephalopathies or traumatic brain injuries can trigger different types of
patterns with repetitive discharges and seizures. The severity of these diseases and injuries often
require intensive medical treatment and continuous monitoring of neurological activity. The
automatic detection of epileptic seizures and repetitive patterns in measurements of electrical brain
signals is a central part of this work.
Currently, diagnostic in neurological patients involves a wide spectrum of methods and tools. In
addition to clinical observations the objective quantification of the brain status is the primary step
in diagnosis. Imaging methods such as magnetic resonance imaging (MRI) are able to generate a
snapshot of the brain morphology. To evaluate the brain activity over time the
electroencephalography (EEG) is used. The EEG is able to continuously record the electrical activity
of the cortex which makes this method to a central element in the diagnosis of neurological patients.
Manual evaluation of EEG is done by visual inspection of a graphical representation of these signals.
Specially trained medical staff interpret the EEG in 10 to 15 second sections throughout the whole
recording period. With a usual recording time of 4 days more than 34,000 EEG pages need to be
evaluated, which requires a considerably amount of time for EEG interpretation. The correct
evaluation of the EEG curves requires a high degree of experience in order to avoid
misinterpretations. Despite intensive training, different interpretations of an EEG by different
reviewers and differences in the evaluation of an EEG by the same reviewer at different time points
are fundamental problems. Further, distortions from electrical signals of muscles (electromyography,
EMG) can overlay with the EEG, which may lead to misinterpretations. In order to make medical
findings comparable in the context of international studies, standardized descriptions of EEG
patterns have been developed which are, however, require experienced staff to be assigned
correctly. If the EEG is used for real-time monitoring of patients with serious illnesses, the evaluation
of the EEGs must be carried out promptly in order to achieve an improved treatment. However, this
places extremely high demands on the staff. Many of the mentioned problems in manual EEG
analysis can be addressed by using computer-aided evaluation methods. A computer algorithm is
able to reduce the cost and time of analysis and provides perfect repeatability of the result. Despite
these obvious advantages, correct automatic analysis of the highly complex EEG signal is an
unsolved problem that prevents the widespread use of such computer algorithms.
In this work novel computer algorithms for the automatic interpretation and monitoring of EEG
signals are presented that were published in eight papers of highly-ranked peer reviewed journals.
The overall aim is to make EEG evaluation considerably easier by automatically marking important
time points in real-time. The focus is on the detection of epileptic seizures and patterns with
repetitive discharges. Although the EEG is the primary data source for the algorithms, EMG
interferences have to be treated adequately in order to achieve the highest possible precision. To
raise sensitivity of the automatic seizure detection algorithm even further the electrocardiography
(ECG) signal was additionally evaluated to find seizure related activity. The use of computer
algorithms for real-time monitoring of EEG activity is intended to improve treatment of patients and
to increase patient safety.
A fundamentally new approach for the detection of EEG discharges was developed in this work that
can be applied to a wide range of pathological patterns. By combining individual discharges into
groups that are extended spatially and over time, different types of patterns are modelled. Important
measures such as frequency and amplitude can then be found by simply averaging the group
elements. Furthermore, the temporal progression of patterns is used to quantify changes. The timedomain
algorithm therefore creates the basis for analysis of seizures and other EEG patterns. The
7
classification algorithms that utilize this information then allow the detection of seizures as well as
the quantification of EEG patterns in intensive care patients. The results of the computer algorithms
can be read and interpreted efficiently by means of a newly developed graphical visualization.
The clinical validation of computer algorithms is an essential part of this work. The quality of the
algorithms can only be determined with statistical significance by diagnostic studies including a high
number of patients. Results of the algorithms were compared to manual annotations from experts
to measure sensitivity and specificity. In this work, four multi-center studies and some smaller
preliminary investigations were carried out for different medical questions and algorithms. In total,
EEGs of 621 patients from 6 centers in Europe and the USA were used for validation. The results
show that seizure alarming is possible with a sensitivity of 81% and a false alarm rate of 7 false alarms
per day. A time delay of only 3 seconds was measured from the seizure pattern to the alarm. In the
detection of seizures based on existing EEG files, the algorithms achieved a high sensitivity of 86%
which is required for efficient evaluation. Special epilepsy types such as temporal lobe epilepsy
showed a sensitivity of 94%. The detection of different patterns in the EEG of intensive care patients
yielded in sensitivities between 85% and 93% and specificities in the range of 90% and 96%.
Improved treatment of patients as well as a reduction in workload for medical staff are thus possible.
In the future, mobile EEG systems in the outpatient setting can represent a further significant
improvement in diagnostic. Patients with rarely occurring seizures can save themselves from
protracted hospital stays. Moreover, the use of mobile sleep diagnostic including EEG can increase
the quality of life and save costs. At present, mobile EEG systems are still suffering from problems
such as high time expenditure for the attachment of the electrodes and complex wiring. This results
in low patient acceptance and prevents their widespread use. As soon as these difficulties are solved
computer algorithms can be utilized to evaluate such mobile EEG systems. A large number of medical
applications are then conceivable in which the quality of the algorithms will play a central role.
Objective
This study investigated sensitivity and false detection rate of a multimodal automatic seizure detection algorithm and the applicability to reduced electrode montages for long-term seizure documentation in epilepsy patients.
Methods
An automatic seizure detection algorithm based on EEG, EMG, and ECG signals was developed. EEG/ECG recordings of 92 patients from two epilepsy monitoring units including 494 seizures were used to assess detection performance. EMG data were extracted by bandpass filtering of EEG signals. Sensitivity and false detection rate were evaluated for each signal modality and for reduced electrode montages.
Results
All focal seizures evolving to bilateral tonic-clonic (BTCS, n=50) and 89% of focal seizures (FS, n=139) were detected. Average sensitivity in temporal lobe epilepsy (TLE) patients was 94% and 74% in extratemporal lobe epilepsy (XTLE) patients. Overall detection sensitivity was 86%. Average false detection rate was 12.8 false detections in 24 hours (FD/24h) for TLE and 22 FD/24h in XTLE patients. Utilization of 8 frontal and temporal electrodes reduced average sensitivity by only 5% to 81%.
Conclusion
Our automatic multimodal seizure detection algorithm shows high sensitivity with full and reduced electrode montages.
Significance
Evaluation of different signal modalities and electrode montages paces the way for semi-automatic seizure documentation systems.
Objective:
Compare the spike detection performance of three skilled humans and three computer algorithms.
Methods:
40 prolonged EEGs, 35 containing reported spikes, were evaluated. Spikes and sharp waves were marked by the humans and algorithms. Pairwise sensitivity and false positive rates were calculated for each human-human and algorithm-human pair. Differences in human pairwise performance were calculated and compared to the range of algorithm versus human performance differences as a type of statistical Turing test.
Results:
5474 individual spike events were marked by the humans. Mean, pairwise human sensitivities and false positive rates were 40.0%, 42.1%, and 51.5%, and 0.80, 0.97, and 1.99/min. Only the Persyst 13 (P13) algorithm was comparable to humans - 43.9% and 1.65/min. Evaluation of pairwise differences in sensitivity and false positive rate demonstrated that P13 met statistical noninferiority criteria compared to the humans.
Conclusion:
Humans had only a fair level of agreement in spike marking. The P13 algorithm was statistically noninferior to the humans.
Significance:
This was the first time that a spike detection algorithm and humans performed similarly. The performance comparison methodology utilized here is generally applicable to problems in which skilled human performance is the desired standard and no external gold standard exists.
Importance
Interictal epileptiform discharges (IEDs) in electroencephalograms (EEGs) are a biomarker of epilepsy, seizure risk, and clinical decline. However, there is a scarcity of experts qualified to interpret EEG results. Prior attempts to automate IED detection have been limited by small samples and have not demonstrated expert-level performance. There is a need for a validated automated method to detect IEDs with expert-level reliability.
Objective
To develop and validate a computer algorithm with the ability to identify IEDs as reliably as experts and classify an EEG recording as containing IEDs vs no IEDs.
Design, Setting, and Participants
A total of 9571 scalp EEG records with and without IEDs were used to train a deep neural network (SpikeNet) to perform IED detection. Independent training and testing data sets were generated from 13 262 IED candidates, independently annotated by 8 fellowship-trained clinical neurophysiologists, and 8520 EEG records containing no IEDs based on clinical EEG reports. Using the estimated spike probability, a classifier designating the whole EEG recording as positive or negative was also built.
Main Outcomes and Measures
SpikeNet accuracy, sensitivity, and specificity compared with fellowship-trained neurophysiology experts for identifying IEDs and classifying EEGs as positive or negative or negative for IEDs. Statistical performance was assessed via calibration error and area under the receiver operating characteristic curve (AUC). All performance statistics were estimated using 10-fold cross-validation.
Results
SpikeNet surpassed both expert interpretation and an industry standard commercial IED detector, based on calibration error (SpikeNet, 0.041; 95% CI, 0.033-0.049; vs industry standard, 0.066; 95% CI, 0.060-0.078; vs experts, mean, 0.183; range, 0.081-0.364) and binary classification performance based on AUC (SpikeNet, 0.980; 95% CI, 0.977-0.984; vs industry standard, 0.882; 95% CI, 0.872-0.893). Whole EEG classification had a mean calibration error of 0.126 (range, 0.109-0.1444) vs experts (mean, 0.197; range, 0.099-0.372) and AUC of 0.847 (95% CI, 0.830-0.865).
Conclusions and Relevance
In this study, SpikeNet automatically detected IEDs and classified whole EEGs as IED-positive or IED-negative. This may be the first time an algorithm has been shown to exceed expert performance for IED detection in a representative sample of EEGs and may thus be a valuable tool for expedited review of EEGs.
Background:
Atrial fibrillation is frequently asymptomatic and thus underdetected but is associated with stroke, heart failure, and death. Existing screening methods require prolonged monitoring and are limited by cost and low yield. We aimed to develop a rapid, inexpensive, point-of-care means of identifying patients with atrial fibrillation using machine learning.
Methods:
We developed an artificial intelligence (AI)-enabled electrocardiograph (ECG) using a convolutional neural network to detect the electrocardiographic signature of atrial fibrillation present during normal sinus rhythm using standard 10-second, 12-lead ECGs. We included all patients aged 18 years or older with at least one digital, normal sinus rhythm, standard 10-second, 12-lead ECG acquired in the supine position at the Mayo Clinic ECG laboratory between Dec 31, 1993, and July 21, 2017, with rhythm labels validated by trained personnel under cardiologist supervision. We classified patients with at least one ECG with a rhythm of atrial fibrillation or atrial flutter as positive for atrial fibrillation. We allocated ECGs to the training, internal validation, and testing datasets in a 7:1:2 ratio. We calculated the area under the curve (AUC) of the receiver operatoring characteristic curve for the internal validation dataset to select a probability threshold, which we applied to the testing dataset. We evaluated model performance on the testing dataset by calculating the AUC and the accuracy, sensitivity, specificity, and F1 score with two-sided 95% CIs.
Findings:
We included 180 922 patients with 649 931 normal sinus rhythm ECGs for analysis: 454 789 ECGs recorded from 126 526 patients in the training dataset, 64 340 ECGs from 18 116 patients in the internal validation dataset, and 130 802 ECGs from 36 280 patients in the testing dataset. 3051 (8·4%) patients in the testing dataset had verified atrial fibrillation before the normal sinus rhythm ECG tested by the model. A single AI-enabled ECG identified atrial fibrillation with an AUC of 0·87 (95% CI 0·86-0·88), sensitivity of 79·0% (77·5-80·4), specificity of 79·5% (79·0-79·9), F1 score of 39·2% (38·1-40·3), and overall accuracy of 79·4% (79·0-79·9). Including all ECGs acquired during the first month of each patient's window of interest (ie, the study start date or 31 days before the first recorded atrial fibrillation ECG) increased the AUC to 0·90 (0·90-0·91), sensitivity to 82·3% (80·9-83·6), specificity to 83·4% (83·0-83·8), F1 score to 45·4% (44·2-46·5), and overall accuracy to 83·3% (83·0-83·7).
Interpretation:
An AI-enabled ECG acquired during normal sinus rhythm permits identification at point of care of individuals with atrial fibrillation.
Funding:
None.
Background:
Finding interictal epileptiform discharges (IEDs) in the EEG is a part of diagnosing epilepsy. Automated software for annotating EEGs of patients with suspected epilepsy can therefore help with reaching a diagnosis. A large amount of data is required for training and evaluating an effective IED detection system. IEDs occur infrequently in the most patients' EEG, therefore, interictal EEG recordings contain mostly background waveforms.
New method:
As the first step to detect IEDs, we propose a machine learning technique eliminating most EEG background data using an ensemble of simple fast classifiers based on several EEG features. This could save computation time for an IED detection method, allowing the remaining waveforms to be classified by more computationally intensive methods. We consider several efficient features and reject background by applying thresholds on them in consecutive steps.
Results:
We applied the proposed algorithm on a dataset of 156 EEGs (93 and 63 with and without IEDs, respectively). We were able to eliminate 78% of background waveforms while retaining 97% of IEDs on our cross-validated dataset.
Comparison with existing methods:
We applied support vector machine, k-nearest neighbours, and random forest classifiers to detect IEDs with and without initial background rejection. Results show that rejecting background by our proposed method speeds up the overall classification by a factor ranging from 1.8 to 4.7 for the considered classifiers.
Conclusions:
The proposed method successfully reduces computation time of an IED detection system. Therefore, it is beneficial in speeding up IED detection especially when utilizing large EEG datasets.
Objective:
Visual assessment of the EEG still outperforms current computer algorithms in detecting epileptiform discharges. Deep learning is a promising novel approach, being able to learn from large datasets. Here, we show pilot results of detecting epileptiform discharges using deep neural networks.
Methods:
We selected 50 EEGs from focal epilepsy patients. All epileptiform discharges (n = 1815) were annotated by an experienced neurophysiologist and extracted as 2 s epochs. In addition, 50 normal EEGs were divided into 2 s epochs. All epochs were divided into a training (n = 41,381) and test (n = 8775) set. We implemented several combinations of convolutional and recurrent neural networks, providing the probability for the presence of epileptiform discharges. The network with the largest area under the ROC curve (AUC) in the test set was validated on seven independent EEGs with focal epileptiform discharges and twelve normal EEGs.
Results:
The final network had an AUC of 0.94 for the test set. Validation allowed detection of epileptiform discharges with 47.4% sensitivity and 98.0% specificity (FPR: 0.6/min). For the normal EEGs in the validation set, the specificity was 99.9% (FPR: 0.03/min).
Conclusions:
Deep neural networks can accurately detect epileptiform discharges from scalp EEG recordings.
Significance:
Deep learning may result in a fundamental shift in clinical EEG analysis.
Electroencephalography (EEG) remains an essential diagnostic tool for people with epilepsy (PWE). The International Federation of Clinical Neurophysiology produces new guidelines as an educational service for clinicians to address gaps in knowledge in clinical neurophysiology. The current guideline was prepared in response to gaps present in epilepsy-related neurophysiological assessment and is not intended to replace sound clinical judgement in the care of PWE. Furthermore, addressing specific pathophysiological conditions of the brain that produce epilepsy is of primary importance though is beyond the scope of this guideline. Instead, our goal is to summarize the scientific evidence for the utility of EEG when diagnosing and monitoring PWE.
State-of-the-art object detection networks depend on region proposal algorithms to hypothesize object locations. Advances like SPPnet [7] and Fast R-CNN [5] have reduced the running time of these detection networks, exposing region pro-posal computation as a bottleneck. In this work, we introduce a Region Proposal Network (RPN) that shares full-image convolutional features with the detection network, thus enabling nearly cost-free region proposals. An RPN is a fully-convolutional network that simultaneously predicts object bounds and objectness scores at each position. RPNs are trained end-to-end to generate high-quality region proposals, which are used by Fast R-CNN for detection. With a simple alternating optimization, RPN and Fast R-CNN can be trained to share convolu-tional features. For the very deep VGG-16 model [18], our detection system has a frame rate of 5fps (including all steps) on a GPU, while achieving state-of-the-art object detection accuracy on PASCAL VOC 2007 (73.2% mAP) and 2012 (70.4% mAP) using 300 proposals per image. The code will be released.
Standardized EEG electrode positions are essential for both clinical applications and research. The aim of this guideline is to update and expand the unifying nomenclature and standardized positioning for EEG scalp electrodes. Electrode positions were based on 20% and 10% of standardized measurements from anatomical landmarks on the skull. However, standard recordings do not cover the anterior and basal temporal lobes, which is the most frequent source of epileptogenic activity. Here, we propose a basic array of 25 electrodes including the inferior temporal chain, which should be used for all standard clinical recordings. The nomenclature in the basic array is consistent with the 10-10-system. High-density scalp EEG arrays (64-256 electrodes) allow source imaging with even sub-lobar precision. This supplementary exam should be requested whenever necessary, e.g. search for epileptogenic activity in negative standard EEG or for presurgical evaluation. In the near future, nomenclature for high density electrodes arrays beyond the 10-10 system needs to be defined, to allow comparison and standardized recordings across centers. Contrary to the established belief that smaller heads needs less electrodes, in young children at least as many electrodes as in adults should be applied due to smaller skull thickness and the risk of spatial aliasing.
Aim of the study:
A novel method for removal of artifacts from long-term EEGs was developed and evaluated. The method targets most types of artifacts and works without user interaction.
Materials and methods:
The method is based on a neurophysiological model and utilizes an iterative Bayesian estimation scheme. The performance was evaluated by two independent reviewers. From 48 consecutive epilepsy patients, 102 twenty-second seizure onset EEGs were used to evaluate artifacts before and after artifact removal and regarding the erroneous attenuation of true EEG patterns.
Results:
The two reviewers found "major improvements" in 59% and 49% of the EEG epochs respectively, and "minor improvements" in 38% and 47% of the epochs, respectively. The answer "similar or worse" was chosen only in 0% and 4%, respectively. Neither of the reviewers found "major attenuations", i.e., a significant attenuation of significant EEG patterns. Most EEG epochs were found to be either "mostly preserved" or "all preserved". A "minor attenuation" was found only in 0% and 17%, respectively.
Conclusions:
The proposed artifact removal algorithm effectively removes artifacts from EEGs and improves the readability of EEGs impaired by artifacts. Only in rare cases did the algorithm slightly attenuate EEG patterns, but the clear visibility of significant patterns was preserved in all cases of this study. Current artifact removal methods work either semi-automatically or with insufficient reliability for clinical use, whereas the "PureEEG" method works fully automatically and leaves true EEG patterns unchanged with a high reliability.
Purpose
To describe the development and implementation of video EEG telemetry (VT) in the patient's home (home video telemetry, HVT) in a single centre.
Methods
HVT met the UK Medical Research Council definition of a complex intervention, and we used its guidance to evaluate the process of piloting, evaluating, developing and implementing this new clinical service. The first phase was a feasibility study, comparing inpatient VT (IVT) with HVT in a test-retest design, to assess data quality and yield of clinically-relevant events (n = 5). The second phase was a pre-implementation study, to examine acceptability and satisfaction (n = 8) as well as the costs of IVT and HVT. Subsequently, we implemented the service, and reviewed the outcomes of the first 34 patients.
Results
The feasibility study found no difference in the quality of recording or clinical yield between IVT and HVT. The pre-implementation study showed excellent patient satisfaction. We also discuss the findings of the main stakeholder survey (consultants and technicians). Our economic modelling demonstrates a clear financial superiority of HVT over IVT.
Conclusion
Our findings show that diagnostic HVT for seizure classification and polysomnographies can be carried out safely in the patients’ home and poses no security risks for staff. HVT can be effectively integrated into an existing tertiary care service as a routine home or community-based procedure. We hope to encourage other clinical neurophysiology departments and epilepsy centres to take advantage of our experience and consider adopting and implementing HVT, with the aim of a nationwide coverage.
In a previous study we proposed a robust method for automatic seizure detection in scalp EEG recordings. The goal of the current study was to validate an improved algorithm in a much larger group of patients in order to show its general applicability in clinical routine.
For the detection of seizures we developed an algorithm based on Short Time Fourier Transform, calculating the integrated power in the frequency band 2.5-12Hz for a multi-channel seizure detection montage referenced against the average of Fz-Cz-Pz. For identification of seizures an adaptive thresholding technique was applied. Complete data sets of each patient were used for analyses for a fixed set of parameters.
159 patients (117 temporal-lobe epilepsies (TLE), 35 extra-temporal lobe epilepsies (ETLE), 7 other) were included with a total of 25,278h of EEG data, 794 seizures were analyzed. The sensitivity was 87.3% and number of false detections per hour (FpH) was 0.22/h. The sensitivity for TLE patients was 89.9% and FpH=0.19/h; for ETLE patients sensitivity was 77.4% and FpH=0.25/h.
The seizure detection algorithm provided high values for sensitivity and selectivity for unselected large EEG data sets without a priori assumptions of seizure patterns.
The algorithm is a valuable tool for fast and effective screening of long-term scalp EEG recordings.
The burden of reviewing long-term scalp electroencephalography (EEG) is not much alleviated by automated spike detection if thousands of events need to be inspected and mentally classified by the reviewer. This study investigated a novel technique of clustering and 24-h hyper-clustering on top of automated detection to assess whether fast review of focal interictal spike types was feasible and comparable to the spikes types observed during routine EEG review in epilepsy monitoring.
Spike detection used a transformation of scalp EEG into 29 regional source activities and adaptive thresholds to increase sensitivity. Our rule-based algorithm estimated 18 parameters around each detected peak and combined multichannel detections into one event. Similarity measures were derived from equivalent location, scalp topography, and source waveform of each event to form clusters over 2-h epochs using a density-based algorithm. Similar measures were applied to all 2-h clusters to form 24-h hyper-clusters. Independent raters evaluated electroencephalography data of 50 patients with epilepsy (25 children) using traditional visual spike review and optimized hyper-cluster inspection. Congruence between visual spike types and epileptiform hyper-clusters was assessed on a sublobar level using three-dimensional (3D) peak topographies.
Visual rating found 126 different epileptiform spike types (2.5 per patient). Independently, 129 hyper-clusters were classified as epileptiform and originating in separate sublobar regions (2.6 per patient). Ninety-one percent of visual spike types matched with hyper-clusters (temporal lobe spikes 94%, extratemporal 89%). Conversely, 11% of hyper-clusters rated epileptiform had no corresponding visual spike type. Numbers were comparable in adults and children. On average, 15 hyper-clusters had to be inspected and rated per patient with an evaluation time of around 5 min.
Hyper-clustering over 24 h provides an independent tool for rapid daily evaluation of interictal spikes in long-term video-EEG monitoring. If used in addition to routine review of 2-5 min EEG per hour, sensitivity and reliability in noninvasive diagnosis of focal epilepsy increases.
With the advent of noninvasive neuroimaging, a plethora of digital human neuroanatomical atlases has been developed. The accuracy of these atlases is constrained by the resolution and signal-gathering powers of available imaging equipment. In an attempt to circumvent these limitations and to produce a high resolution in vivo human neuroanatomy, we investigated the usefulness of intrasubject registration for post hoc MR signal averaging.
Twenty-seven high resolution (7 x 0.78 and 20 x 1.0 mm3) T1-weighted volumes were acquired from a single subject, along with 12 double echo T2/proton density-weighted volumes. These volumes were automatically registered to a common stereotaxic space in which they were subsampled and intensity averaged. The resulting images were examined for anatomical quality and usefulness for other analytical techniques.
The quality of the resulting image from the combination of as few as five T1 volumes was visibly enhanced. The signal-to-noise ratio was expected to increase as the root of the number of contributing scans to 5.2, n = 27. The improvement in the n = 27 average was great enough that fine anatomical details, such as thalamic subnuclei and the gray bridges between the caudate and putamen, became crisply defined. The gray/white matter boundaries were also enhanced, as was the visibility of any finer structure that was surrounded by tissue of varying T1 intensity. The T2 and proton density average images were also of higher quality than single scans, but the improvement was not as dramatic as that of the T1 volumes.
Overall, the enhanced signal in the averaged images resulted in higher quality anatomical images, and the data lent themselves to several postprocessing techniques. The high quality of the enhanced images permits novel uses of the data and extends the possibilities for in vivo human neuroanatomy.
The purpose of this study was to evaluate and validate an offline, automated scalp EEG-based seizure detection system and to compare its performance to commercially available seizure detection software.
The test seizure detection system, IdentEvent™, was developed to enhance the efficiency of post-hoc long-term EEG review in epilepsy monitoring units. It translates multi-channel scalp EEG signals into multiple EEG descriptors and recognizes ictal EEG patterns. Detection criteria and thresholds were optimized in 47 long-term scalp EEG recordings selected for training (47 subjects, ∼3653h with 141 seizures). The detection performance of IdentEvent was evaluated using a separate test dataset consisting of 436 EEG segments obtained from 55 subjects (∼1200h with 146 seizures). Each of the test EEG segments was reviewed by three independent epileptologists and the presence or absence of seizures in each epoch was determined by majority rule. Seizure detection sensitivity and false detection rate were calculated for IdentEvent as well as for the comparable detection software (Persyst's Reveal®, version 2008.03.13, with three parameter settings). Bootstrap re-sampling was applied to establish the 95% confidence intervals of the estimates and for the performance comparison between two detection algorithms.
The overall detection sensitivity of IdentEvent was 79.5% with a false detection rate (FDR) of 2 per 24h, whereas the comparison system had 80.8%, 76%, and 74% sensitivity using its three detection thresholds (perception score) with FDRs of 13, 8, and 6 per 24h, respectively. Bootstrap 95% confidence intervals of the performance difference revealed that the two detection systems had comparable detection sensitivity, but IdentEvent generated a significantly (p<0.05) smaller FDR.
The study validates the performance of the IdentEvent™ seizure detection system.
With comparable detection sensitivity, an improved false detection rate makes the automated seizure detection software more useful in clinical practice.
Computerized detection of epileptiform transients (ETs), also called spikes and sharp waves, in the electroencephalogram (EEG) has been a research goal for the last 40years. A reliable method for detecting ETs could improve efficiency in reviewing long EEG recordings and assist physicians in interpreting routine EEGs. Computer algorithms developed so far for detecting ETs are not as reliable as human expert interpreters, mostly due to the large number of false positive detections. Typical methods for ET detection include measuring waveform morphology, detecting signal non-stationarity, and power spectrum analysis. Some progress has been made by using more advanced algorithmic approaches including wavelet analysis, artificial neural networks, and dipole analysis. Comparing the performance of different algorithms is difficult since each study uses its own EEG test dataset. In order to overcome this problem, European researchers in the field of computerized electrocardiogram interpretation organized a large multi-center research workgroup to create a standardized dataset of ECG recordings which were interpreted by a large group of cardiologists. EEG researchers need to follow this as a model and seek funding for the creation of a standardized EEG research dataset to develop ET detection algorithms and certify commercial software.
Most epilepsy centers obtain ictal EEG recordings to localize the epileptogenic zone during presurgical evaluations. Inpatient monitoring is standard practice but is expensive and can be inconvenient. The authors sought to determine whether outpatient monitoring can be safe and effective as the sole method of recording seizures in the presurgical evaluation of patients with refractory temporal lobe epilepsy. They reviewed the data of seven temporal lobectomy patients whose presurgical monitoring was performed entirely outside the hospital. Mean baseline seizure frequency was at least 9.1 seizures per week. An average of 7.4 seizures was recorded over 9.4 days of monitoring. Only one patient had any antiepileptic drug taper; none suffered any complications. After temporal lobectomy on the side of demonstrated ictal onset, postoperative follow-up averaged 5.5 years. At the most recent follow-up, all patients were either seizure free or had only rare disabling or nocturnal seizures (four patients had outcomes in Engel's class I and three patients in Engel's class II). A comparison group who underwent standard inpatient monitoring was similar in average seizure frequency, monitoring duration, number of seizures recorded, and postoperative outcome, although all but one had antiepileptic drugs tapered during monitoring. The authors conclude that there is a subset of patients for whom solely outpatient presurgical EEG monitoring can be used to help plan successful temporal lobectomy.
The interictal EEG provides information that aids in diagnosis and management of epilepsy. One must remember that the EEG is merely a tool, and its usefulness depends largely upon the skill of the individual who wields it. Like all diagnostic tests, it has significant limitations and cannot substitute for a careful history and exercise of good judgment. Nonetheless, in skilled hands, it provides unique and vital information in many patients, and enhances our understanding of their condition.
Clinical-grade computational pathology using weakly supervised deep learning on whole slide images
- G Campanella
- M G Hanna
- L Geneslaw
- A Miraflor
- Vwk Silva
- K J Busam
Campanella G, Hanna MG, Geneslaw L, Miraflor A, Silva VWK, Busam KJ, et al.
Clinical-grade computational pathology using weakly supervised deep learning
on whole slide images. Nat Med 2019;25:1301-9.
A comparison of rule-based and machine learning methods for classification of Spikes in EEG
- W Ganglberger
- G Gritsch
- M M Hartmann
- F Fürbass
- H Perko
- A Skupch
Ganglberger W, Gritsch G, Hartmann MM, Fürbass F, Perko H, Skupch A, et al. A
comparison of rule-based and machine learning methods for classification of
Spikes in EEG. J Commun 2017. https://doi.org/10.12720/jcm.12.10.589-595:589-95.
Atlas und Klassifikation der Elektroenzephalographie: Einführung in die EEG-Auswertung von Lüders, Hans O. ; Soheyl Noachtar: Ciba-Geigy-Verl
- H Lüders
- S Noachtar
Lüders H, Noachtar S. Atlas und Klassifikation der Elektroenzephalographie:
Einführung in die EEG-Auswertung von Lüders, Hans O. ; Soheyl Noachtar:
Ciba-Geigy-Verl., Wehr 9783929126075 -Versandantiquariat Wolter. CIBA-GEIGY Verlag; 1994.
- L Shen
- L R Margolies
- J H Rothstein
- E Fluder
- R Mcbride
- W Sieh
Shen L, Margolies LR, Rothstein JH, Fluder E, McBride R, Sieh W. Deep Learning to Improve Breast
Cancer Detection on Screening Mammography. Sci Rep. 2019;9:12495.
Unsupervised Data Augmentation for Consistency Training. ArXiv E-Prints
- Q Xie
- Z Dai
- E Hovy
- M-T Luong
- Q V Le
Xie Q, Dai Z, Hovy E, Luong M-T, Le QV. Unsupervised Data Augmentation for Consistency
Training. ArXiv E-Prints. 2019 Apr 1;1904:arXiv:1904.12848.
Unsupervised data augmentation for consistency training
- Q Xie
- Z Dai
- E Hovy
- M-T Luong
- Q V Le
Xie Q, Dai Z, Hovy E, Luong M-T, Le QV. Unsupervised data augmentation for
consistency training. ArXiv E-Prints; 2019 Apr 1;1904:arXiv:1904.12848.