Fig 3 - available via license: Creative Commons Attribution 4.0 International
Content may be subject to copyright.
Beat-by-beat features extracted from an OW measurement and the filtered features using 2, 3, 4 and 5 points zero-phase moving average filters.
Source publication
In general, existing machine learning based approaches, developed for systolic and diastolic blood pressure (SBP and DBP) estimation from oscillometric waveforms (OWs), employ features extracted from the OW envelope (OWE) alone and ignore important beat-by-beat (BBB) features which represent fundamental physical properties of the entire non-invasiv...
Context in source publication
Citations
... These include no adjacent peaks within two adjacent cardiac beats and participation in an ensemble of signals separated by one cardiac period, generally increasing and then decreasing in amplitude. In due course this process can be fully automated using deep learning algorithms (Argha and Celler 2019; Argha, Celler, and Lovell 2020a; 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 A c c e p t e d M a n u s c r i p t Argha, Celler, and Lovell 2020b;Argha, J. Wu, et al. 2019;Celler, Le, Argha, et al. 2019). ...
Objectives.
In this study, we test the hypothesis that if, as demonstrated in a previous study, brachial arteries exhibit hysteresis as the occluding cuff is deflated and fail to open until cuff pressure (CP) is well below true intra-arterial blood pressure (IAPB), estimating systolic (SBP) and diastolic blood pressure (DBP) from the presence of Korotkoff sounds as CP increases may eliminate these errors and give more accurate estimates of SBP and DBP relative to IABP readings.
Approach. In 62 subjects of varying ages (45.1±19.8, range 20-6 - 75.8 years), including 44 men (45.3±19.4, range 20.6 – 75.8 years) and 18 women (44.4±21.4, range 20.9 - 75.3 years), we sequentially recorded SBP and DBP both during cuff inflation and cuff deflation using Korotkoff sounds.
Results. There was a significant (p<0.0001) increase in SBP from 122.8±13.2 to 127.6±13.0 mmHg and a significant (p=0.0001) increase in DBP from 70.0±9.0 to 77.5±9.7 mmHg. Of the 62 subjects, 51 showed a positive increase in SBP (0 to 14 mmHg) and 11 subjects showed a reduction (-0.3 to -7 mmHg). The average differences for SBP and DBP estimates derived as the cuff inflates and those derived as the cuff deflates were 4.8±4.6 mmHg and 2.5±4.6mmHg, not dissimilar to the differences reported between IABP and NIBP measurements. Although we could not develop multiparameter linear or non-linear models to explain this phenomenon we have clearly demonstrated through ANOVA tests that both body mass index (BMI) and pulse wave velocity (PWV) are implicated, supporting the hypothesis that the phenomenon is associated with age, higher BMI and stiffer arteries.
Significance. The implications of this study are that brachial sphygmomanometry carried out during cuff inflation could be more accurate than measurements carried out as the cuff deflates. Further research is required to validate these results with intra-arterial blood pressure measurements.
... AI-based oscillometric techniques include probabilistic methods (e.g., Gaussian Mixture Model), linear regression methods, support regression methods, feedforward neural networks, deep learning regression methods [31]. On the other hand, AI-based auscultatory techniques comprise feedforward neural networks, Hidden Markov Models, long-short-term recurrent neural networks, deep belief networks [5][6]8,31,[41][42][43][44][45][46][47]. Although AIbased NIBP methods outperform the accuracy of traditional techniques, the high complexity and computational cost, as well as the need to train the classifier, represent the major drawbacks of these approaches. ...
Objective:
The auscultatory technique is still considered the most accurate method for non-invasive blood pressure (NIBP) measurement, although its reliability depends on operator's skills. Various methods for automated Korotkoff sounds analysis have been proposed for reliable estimation of systolic (SBP) and diastolic (DBP) blood pressures. To this aim, very complex methodologies have been presented, including some based on artificial intelligence (AI). This study proposes a relatively simple methodology, named B3X, to estimate SBP and DBP by processing Korotkoff sounds recordings acquired during an auscultatory NIBP measurement.
Approach:
The beat-by-beat change in morphology of adjacent Korotkoff sounds is evaluated via their cross-correlation. The time series of the beat-by-beat cross-correlation and its first derivative are analyzed to locate the timings of SBP and DBP values. Extensive tests were performed on a public database of 350 annotated measurements, and the performance was evaluated according to the BHS, AAMI/ANSI, and ISO quality standards.
Main results:
The proposed approach achieved "A" scores for SBP and DBP in the BHS grading system, and passed the quality tests of AAMI/ANSI and ISO standards. The B3X algorithm outperformed two well-established algorithms for oscillometric NIBP measurement in both SBP and DBP estimation. It also outperformed four AI-based algorithms in DBP estimation, while providing comparable performance for SBP, at the cost of a much lower computational burden. The full code of the B3X algorithm is provided in a public repository.
Significance:
The very good performances ensured by the proposed B3X algorithm, at a low computational cost and without the need for parameter training, support its direct implementation into clinical blood pressure monitoring devices. The results of this study pave the way for solving/overcoming the trade-off between the accuracy of the auscultatory technique and the objectivity of oscillatory measurements, by bringing an automated auscultatory blood pressure measurement method in clinical practice.
... Another important factor that might influence the performance results of the DL approach is the data volume used in the learning phase. It is a well-known fact that a lot of data number is needed in the training phase for a DL method to work well [87,88]. However, as previously stated, most of the included studies used the SZ and MC data sets, which are considered small data sets. ...
Background:
Tuberculosis (TB) was the leading infectious cause of mortality globally prior to COVID-19 and chest radiography has an important role in the detection, and subsequent diagnosis, of patients with this disease. The conventional experts reading has substantial within- and between-observer variability, indicating poor reliability of human readers. Substantial efforts have been made in utilizing various artificial intelligence-based algorithms to address the limitations of human reading of chest radiographs for diagnosing TB.
Objective:
This systematic literature review (SLR) aims to assess the performance of machine learning (ML) and deep learning (DL) in the detection of TB using chest radiography (chest x-ray [CXR]).
Methods:
In conducting and reporting the SLR, we followed the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines. A total of 309 records were identified from Scopus, PubMed, and IEEE (Institute of Electrical and Electronics Engineers) databases. We independently screened, reviewed, and assessed all available records and included 47 studies that met the inclusion criteria in this SLR. We also performed the risk of bias assessment using Quality Assessment of Diagnostic Accuracy Studies version 2 (QUADAS-2) and meta-analysis of 10 included studies that provided confusion matrix results.
Results:
Various CXR data sets have been used in the included studies, with 2 of the most popular ones being Montgomery County (n=29) and Shenzhen (n=36) data sets. DL (n=34) was more commonly used than ML (n=7) in the included studies. Most studies used human radiologist's report as the reference standard. Support vector machine (n=5), k-nearest neighbors (n=3), and random forest (n=2) were the most popular ML approaches. Meanwhile, convolutional neural networks were the most commonly used DL techniques, with the 4 most popular applications being ResNet-50 (n=11), VGG-16 (n=8), VGG-19 (n=7), and AlexNet (n=6). Four performance metrics were popularly used, namely, accuracy (n=35), area under the curve (AUC; n=34), sensitivity (n=27), and specificity (n=23). In terms of the performance results, ML showed higher accuracy (mean ~93.71%) and sensitivity (mean ~92.55%), while on average DL models achieved better AUC (mean ~92.12%) and specificity (mean ~91.54%). Based on data from 10 studies that provided confusion matrix results, we estimated the pooled sensitivity and specificity of ML and DL methods to be 0.9857 (95% CI 0.9477-1.00) and 0.9805 (95% CI 0.9255-1.00), respectively. From the risk of bias assessment, 17 studies were regarded as having unclear risks for the reference standard aspect and 6 studies were regarded as having unclear risks for the flow and timing aspect. Only 2 included studies had built applications based on the proposed solutions.
Conclusions:
Findings from this SLR confirm the high potential of both ML and DL for TB detection using CXR. Future studies need to pay a close attention on 2 aspects of risk of bias, namely, the reference standard and the flow and timing aspects.
Trial registration:
PROSPERO CRD42021277155; https://www.crd.york.ac.uk/prospero/display_record.php?RecordID=277155.
... On the other hand, a multiple linear regression method used ten features used in 19 along with area under the oscillometric waveform before and after maximum altitude's position [151]. Beyond these, an improved technique using deep-belief network-deep neural network [152] to estimate the blood pressure more accurately was proposed, as it can outperform other methods such as maximum amplitude algorithm (MAA), feed-forward neural network (FFNN) and support vector regression (SVR) [153]. It takes advantage of the unsupervised learning step to initialize each layer and uses a backpropagation algorithm to continue the supervised learning steps, which fine-tune the values achieved in the first step [154]. ...
... A vocational technical school has recently used speech recordings consisting of vowels with their corresponding various BP measurements of each participant and used Convolutional Neural Networks-Regression (CNN-R), Support Vector Machines-Regression (SVMs-R) and Multi Linear Regression (MLR) and provided results of 93.65%, 92.15%, 89.43% respectively [164]. A number of machine learning based methods in the existing literature have attempted to some extent overcome the lack of enough training data by data augmentation [152,165]. These use bootstrap technique to augment the training data, while others used elegant deep learning-based methods, including auto-encoders (AEs) and generative adversarial networks (GANs) [147]. ...
The measurement of physiologic pressure helps diagnose and prevent associated health complications. From typical conventional methods to more complicated modalities, such as the estimation of intracranial pressures, numerous invasive and noninvasive tools that provide us with insight into daily physiology and aid in understanding pathology are within our grasp. Currently, our standards for estimating vital pressures, including continuous BP measurements, pulmonary capillary wedge pressures, and hepatic portal gradients, involve the use of invasive modalities. As an emerging field in medical technology, artificial intelligence (AI) has been incorporated into analyzing and predicting patterns of physiologic pressures. AI has been used to construct models that have clinical applicability both in hospital settings and at-home settings for ease of use for patients. Studies applying AI to each of these compartmental pressures were searched and shortlisted for thorough assessment and review. There are several AI-based innovations in noninvasive blood pressure estimation based on imaging, auscultation, oscillometry and wearable technology employing biosignals. The purpose of this review is to provide an in-depth assessment of the involved physiologies, prevailing methodologies and emerging technologies incorporating AI in clinical practice for each type of compartmental pressure measurement. We also bring to the forefront AI-based noninvasive estimation techniques for physiologic pressure based on microwave systems that have promising potential for clinical practice.
... With advances in machine learning and deep learning, biomedical signal processing technology has recently made significant technological advancements. Many machine learning techniques have been recently established to estimate blood pressure (BP), such as support vector machine [6], random forest [7], and neural networks [8]. Reconstruction of mathematical models of BP estimation using different machine learning approaches is based on electrocardiogram (ECG) and photoplethysmography (PPG) signals [9], [10]. ...
A novel temporal convolutional network (TCN) model is utilized to reconstruct the central aortic blood pressure (aBP) waveform from the radial blood pressure waveform. The method does not need manual feature extraction as traditional transfer function approaches. The data acquired by the SphygmoCor CVMS device in 1,032 participants as a measured database and a public database of 4,374 virtual healthy subjects were used to compare the accuracy and computational cost of the TCN model with the published convolutional neural network and bi-directional long short-term memory (CNN-BiLSTM) model. The TCN model was compared with CNN-BiLSTM in the root mean square error (RMSE). The TCN model generally outperformed the existing CNN-BiLSTM model in terms of accuracy and computational cost. For the measured and public databases, the RMSE of the waveform using the TCN model was 0.55 ± 0.40 mmHg and 0.84 ± 0.29 mmHg, respectively. The training time of the TCN model was 9.63 min and 25.51 min for the entire training set; the average test time was around 1.79 ms and 8.58 ms per test pulse signal from the measured and public databases, respectively. The TCN model is accurate and fast for processing long input signals, and provides a novel method for measuring the aBP waveform. This method may contribute to the early monitoring and prevention of cardiovascular disease.
... Unlike those methods that use data samples of the same individuals in training and test sets [12,18], we ensured that our training-validation and test sets are consisted of data from different individuals and therefore our reported results better demonstrate the generalizability of our approach to unseen data. This is especially important when evaluating medical systems that work based on the analysis of physiological signals, as physiological characteristics may vary significantly between different individuals. ...
... Unlike most of the existing techniques that are based on the analysis of expert-engineered features extracted from the oscillometric waveform [3,7,9,11,12,18], our method provides an end-to-end deep learning framework that does not rely on expert-engineered features. Our proposed method automatically extracts the best representative features from the time pattern of the oscillometric waveforms and therefore, can capture complex input-output relationships that may not be visually observable. ...
Oscillometric monitors are the most common automated blood pressure (BP) measurement devices used in non-specialist settings. However, their accuracy and reliability vary under different settings and for different age groups and health conditions. A main limitation of the existing oscillometric monitors is their underlying analysis algorithms that are unable to fully capture the BP information encoded in the pattern of the recorded oscillometric pulses. In this paper, we propose a new 2D oscillometric data representation that enables a full characterization of arterial system and empowers the application of deep learning to extract the most informative features correlated with BP. A hybrid convolutional-recurrent neural network was developed to capture the oscillometric pulses morphological information as well as their temporal evolution over the cuff deflation period from the 2D structure, and estimate BP. The performance of the proposed method was verified on three oscillometric databases collected from the wrist and upper arms of 245 individuals. It was found that it achieves a mean error and a standard deviation of error of as low as 0.08 mmHg and 2.4 mmHg in the estimation of systolic BP, and 0.04 mmHg and 2.2 mmHg in the estimation of diastolic BP, respectively. Our proposed method outperformed the state-of-the-art techniques and satisfied the current international standards for BP monitors by a wide margin. The proposed method shows promise toward robust and objective BP estimation in a variety of patients and monitoring situations.
... Another important factor that might influence the performance results of the DL approach is the data volume used in the learning phase. It is a well-known fact that a lot of data number is needed in the training phase for a DL method to work well [87,88]. However, as previously stated, most of the included studies used the SZ and MC data sets, which are considered small data sets. ...
BACKGROUND
Tuberculosis (TB) was the leading infectious cause of mortality globally prior to COVID-19 and chest radiography has an important role in the detection, and subsequent diagnosis, of patients with this disease. The conventional experts reading has substantial within- and between-observer variability, indicating poor reliability of human readers. Substantial efforts have been made in utilizing various artificial intelligence–based algorithms to address the limitations of human reading of chest radiographs for diagnosing TB.
OBJECTIVE
This systematic literature review (SLR) aims to assess the performance of machine learning (ML) and deep learning (DL) in the detection of TB using chest radiography (chest x-ray [CXR]).
METHODS
In conducting and reporting the SLR, we followed the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines. A total of 309 records were identified from Scopus, PubMed, and IEEE (Institute of Electrical and Electronics Engineers) databases. We independently screened, reviewed, and assessed all available records and included 47 studies that met the inclusion criteria in this SLR. We also performed the risk of bias assessment using Quality Assessment of Diagnostic Accuracy Studies version 2 (QUADAS-2) and meta-analysis of 10 included studies that provided confusion matrix results.
RESULTS
Various CXR data sets have been used in the included studies, with 2 of the most popular ones being Montgomery County (n=29) and Shenzhen (n=36) data sets. DL (n=34) was more commonly used than ML (n=7) in the included studies. Most studies used human radiologist’s report as the reference standard. Support vector machine (n=5), k-nearest neighbors (n=3), and random forest (n=2) were the most popular ML approaches. Meanwhile, convolutional neural networks were the most commonly used DL techniques, with the 4 most popular applications being ResNet-50 (n=11), VGG-16 (n=8), VGG-19 (n=7), and AlexNet (n=6). Four performance metrics were popularly used, namely, accuracy (n=35), area under the curve (AUC; n=34), sensitivity (n=27), and specificity (n=23). In terms of the performance results, ML showed higher accuracy (mean ~93.71%) and sensitivity (mean ~92.55%), while on average DL models achieved better AUC (mean ~92.12%) and specificity (mean ~91.54%). Based on data from 10 studies that provided confusion matrix results, we estimated the pooled sensitivity and specificity of ML and DL methods to be 0.9857 (95% CI 0.9477-1.00) and 0.9805 (95% CI 0.9255-1.00), respectively. From the risk of bias assessment, 17 studies were regarded as having unclear risks for the reference standard aspect and 6 studies were regarded as having unclear risks for the flow and timing aspect. Only 2 included studies had built applications based on the proposed solutions.
CONCLUSIONS
Findings from this SLR confirm the high potential of both ML and DL for TB detection using CXR. Future studies need to pay a close attention on 2 aspects of risk of bias, namely, the reference standard and the flow and timing aspects.
CLINICALTRIAL
PROSPERO CRD42021277155; https://www.crd.york.ac.uk/prospero/display_record.php?RecordID=277155
... It is also known as Einthoven's Triangle, and the body's bipolar leads are shown in Figs. 6a-6c [23,24]. ...
In this study, a multisensory physiological measurement system was built with wireless transmission technology, using a DSPIC30F4011 as the master control center and equipped with physiological signal acquisition modules such as an electrocardiogram module, blood pressure module, blood oxygen concentration module, and respiratory rate module. The physiological data were transmitted wirelessly to Android-based mobile applications via the TCP/IP or Bluetooth seri-al ports of Wi-Fi. The Android applications displayed the acquired physiological signals in real time and performed a preliminary abnormity diagnosis based on the measured physiological data and built-in index diagnostic data provided by doc-tors, such as blood oxygen concentration, systolic pressure, and diastolic pressure. In addition, in this study, the R waves, which are the highest peaks of the PQRST waves in electrocardiograms, were analyzed and detected using the heart rate variability (HRV) for time-frequency analysis and calculation of the RR interval time series. In this study, the heart rate data in the same age group were collected, and the optimal value of the standard deviation of normal to normal (SDNN) during the time domain of a normal heartbeat was found using the particle swarm optimization (PSO) algorithm and set as the risk level of the SDNN time domain analysis. A spectral analysis on the activity of the autonomic nervous system (ANS) was performed and the preliminary analysis results were displayed on the Android handheld devices for comprehensive physiological data analysis and HRV time-frequency analysis. Healthcare needs are distributed at all levels, therefore user-friendly software interfaces have been written to meet the health-care needs at all ages.
... Therefore, Argha and Celler considered seven features extracted from each pulse of OMW and utilized these features for a long short-term memory recurrent neural network (LSTM-RNN)based method [13]. The study was further developed, and improved results were reported by the author using the DBN-DNN [14]. These studies attempted to identify useful features that contain information on SBP and DBP. ...
Blood pressure measurement is required to monitor the cardiovascular state of a person, and it is commonly conducted in a noninvasive way using oscillometry-based blood pressure monitors (BPM). Blood pressure can be estimated by analyzing the oscillometric waveform (OMW) in the BPM, and many methods have been examined to increase their estimation accuracy. In this study, we proposed a new method that enhances estimation accuracy and requires no external user information, such as age and gender, in the test phase. In the method, the entire OMW was considered as an input to reduce information loss via feature extraction, and convolutional neural networks were utilized to effectively analyze the high-dimensional input. Additionally, the proposed method included a novel ensemble method to further increase the estimation accuracy. The performance of the proposed method was evaluated and compared with other studies via subject-independent tests considering real situations in which it is difficult to obtain preliminary information on a test subject. Data from 64 subjects were used in the test. The mean absolute error of the proposed method was 3.12 and 3.98 mmHg for systolic and diastolic blood pressure, respectively, which was superior to those reported in other studies conducted in similar conditions. Individuals can measure their blood pressure with higher precision using the proposed method with improved estimation performance. This can aid in reducing the risk of cardiovascular diseases.
... These classification-based methods can be divided into two major categories based on their ability to take into account the dependencies between the neighboring pulses; e.g. long short-term memory recurrent neural networks (LSTM-RNNs) [14] and Hidden Markov Models (HMMs) [15] versus deep-belief-network deep-neural-networks (DBN-DNNs) and feedforward neural networks (FFNNs)-based classification models [16]. ...
While measurement of blood pressure (BP) is now widely carried out by automated non-invasive BP (NIBP) monitoring devices, as they do not require skilled clinicians and do not carry risk of complications, their accuracy is in doubt. A novel end-to-end deep learning-based algorithm was developed in this study that estimates NIBP directly from sequences of Korotkoff sounds (KSs) rather than oscillometric waveforms. First, sequences of segments of KSs were formed using different signal segmentation techniques, i.e., segmentation using sliding window with or without overlap and segmentation using the cardiac period estimation. Each segment within each sequence was then labeled as (i) after-systolic and before-diastolic (AB), or (ii) before-systolic or after-diastolic (BA) such that a binary sequence-to-sequence classification problem was achieved. To deal with the resultant sequence-to-sequence classification problem, an algorithm was developed by combining one-dimensional (1D) convolutional neural networks (CNNs) and recurrent neural networks (RNNs). The segments associated with systolic and diastolic blood pressure (SBP and DBP) are then identified as the segments at which the output target sequence switches from class BA to class AB and later from class AB to class BA. Lastly, the values of SBP and DBP are obtained by mapping the center point of the switching segments to the deflation curve. To evaluate the performance of the proposed NIBP estimation method, we used a database of 350 NIBP samples collected from 155 participants (87 male, age: 23-97 years, arm circumference: 10-35 cm, SBP: 81-104 mmHg, and DBP: 37-104 mmHg), and the achieved estimation errors for SBP and DBP, relative to the reference values, using a 5-fold cross validation approach, were 1.6±3.9 mmHg (mean absolute error ± standard deviation of error) and 2.5±4.0 mmHg, respectively. We finally conclude that the proposed end-to-end deep learning-based NIBP estimation algorithm from sequences of KSs is a novel technique that requires modest preprocessing steps and can measure BP accurately.
