Figure 1 - uploaded by Igor Matias
Content may be subject to copyright.
-Flowchart of clinical implications of AF [22].

-Flowchart of clinical implications of AF [22].

Source publication
Thesis
Full-text available
This dissertation presents research on the construction of predictive models for health conditions through the application of Artificial Intelligence methods. The work is thus focused on the prediction, in the short and long term, of Atrial Fibrillation conditions through the analysis of Electrocardiography exams, with the use of several techniques...

Contexts in source publication

Context 1
... clinical implications, AF can lead to rapid ventricular response, decreased diastolic filling or blood stasis and atrial clot formation, which, in the end, represent an increased morbidity and mortality rates. It's most common clinical implications are shown in Fig. 1, as stated by [22]. A rapid ventricular response is characterised as "a heart rhythm disorder (arrhythmia) caused by abnormal electrical signals in the lower chambers of the heart (ventricles)" [26], and is related to a shorter diastolic fill time, which can lead to sudden death, according to [27], thus decreasing the cardiac output, ...
Context 2
... original signals were sliced, which were 30 minutes long with which scaled our data by 60 times, that is, 6000 samples of "normal" type, 3000 of "PAF-overlap of samples condition, that is, there are no data samples repeated in these new data files of 30 seconds, providing each one of the data samples (at 128 Hz) only one time to the models. The Fig. 10 shows the drawing of the first 30 seconds of the database record "p01", which were cut from the original 30 minutes long and stored ...
Context 3
... Fig. 11 shows the produced spectrogram for the wave presented in Fig. 10, that is, the first 30 seconds from the signal "p01" from the database. In it, it is possible to have a better understanding of the signal in need to analyse, by plotting the power of the signal depending on the frequency at each instance of time. The colour bar used in ...
Context 4
... Fig. 11 shows the produced spectrogram for the wave presented in Fig. 10, that is, the first 30 seconds from the signal "p01" from the database. In it, it is possible to have a better understanding of the signal in need to analyse, by plotting the power of the signal depending on the frequency at each instance of time. The colour bar used in this approach represented the lowest power with a dark blue, and ...
Context 5
... in Fig. 12 and Fig. 13, it is possible to compare the spectrogram images of the last 30 seconds of the records "n01" and "p01", respectively, that is, the last seconds before a normal condition ECG and a PAF episode. It is easily noticeable the difference between the yellow regions between both images, with the majority of signal power on low ...
Context 6
... in Fig. 12 and Fig. 13, it is possible to compare the spectrogram images of the last 30 seconds of the records "n01" and "p01", respectively, that is, the last seconds before a normal condition ECG and a PAF episode. It is easily noticeable the difference between the yellow regions between both images, with the majority of signal power on low frequencies for ...
Context 7
... 30 seconds of the records "n01" and "p01", respectively, that is, the last seconds before a normal condition ECG and a PAF episode. It is easily noticeable the difference between the yellow regions between both images, with the majority of signal power on low frequencies for a pre-PAF episode and medium-high frequencies for a normal ECG signal. Fig. 14 and 15 also follow this trend, representing the first 30 seconds of the records "n01" and "p01", respectively. All the figures represent the data as it is at the end of step 2 of the pre-processing phase, that is, ready to be fed to the Convolutional Neural Network described at step 3. For this reason, there are no axis or white spaces ...
Context 8
... all approaches, all the input data had the same image size of 496 by 369, and all were trained and tested 10 times with a batch size of 130 and 100 fixed epochs. The obtained results are described in Section 5. A diagram of the network used in all the three approaches is presented by Fig. 19 for the SPSA and by Fig. 20 for both the SPHA and ...
Context 9
... the final step, and as on the simple approach, it was implemented a fully connected layer with only one neuron, and applied a Sigmoid function as activation, thus returning a probability of a spectrogram being "pre-PAF" type or not, assuming a true case when the probability goes above 0.5. The proposed model diagram is represented in Fig. ...
Context 10
... "pre-PAF" and "PAF-distant" types as positive), this allowed us to classify an ECG section spectrogram between positive and negative for a preceding PAF episode with an event horizon of 45.0 minutes or more, that is, being able to predict the onset of a PAF event in the next 45.0 minutes, at least. The proposed model diagram is represented in Fig. 18. ...
Context 11
... the beginning of the search, it yielded 375 unique records, after the removal of duplicates. After the review of the title and abstract and following the inclusion and exclusion criteria presented in Table 1, 293 records were excluded; 82 full-text publications were assessed for eligibility and after full-text review, of which 72 records were excluded (Figure 1). The excluded records can be described as follows. ...
Context 12
... database that conducted all the three most accurate models achieved was the Atrial Fibrillation Prediction Database. As shown by Figure 1, between 2009 and 2019 (the period this systematic literature review covers), more than 80% of the total published studies were performed from 2016 ahead, 50% belonging to the last two years (2018 and 2019). The amount of work on the prediction of AF episodes is rapidly increasing and showing promising results. ...
Context 13
... this paper, we propose an accurate PAF onset prediction method applying Machine Learning techniques combined with ECG signal spectrograms and noise reduction methods and present early results. Fig. 1 presents a comparison between the common approach and the proposed approach to this prediction problem. The article is organised as follows. Section 2 presents relevant related work, and section 3 describes the used dataset for the training and evaluation of the algorithm. Section 4 elaborates the construction, designing and applied ...
Context 14
... both approaches, all the input data had the same image size of 496 by 369, and both were trained and tested 10 times with a batch size of 130 and 100 fixed epochs. The obtained results are described in Section 5. A diagram of the network used in both approaches is presented by Fig. 12 and Fig. 13, for the simple and the hybrid approach, ...
Context 15
... layer with only one neuron, and applied a Sigmoid function as activation, thus returning a probability of a spectrogram being "pre-PAF" type or not, so that values above 0.5 represent a binary true classification for ECG signals preceding a PAF episode in the next 29 minutes and 30 seconds or less. The proposed model diagram is represented in Fig. 10. For compiling this training model, we used the Binary Crossentropy function as loss function and rmsprop as optimiser parameter. On the train and test data sets, we applied a rescale of 1/255, and a shear range of 0.2, a zoom range of 0.3 and a true horizontal flip for the training data. This algorithm was fed with 6704 data samples ...
Context 16
... the final step, as on the simple approach, we implemented a fully connected layer with only one neuron, and applied a Sigmoid function as activation, thus returning a probability of a spectrogram being "pre-PAF" type or not, assuming a true case when the probability goes above 0.5. The proposed model diagram is represented in Fig. 11. The first algorithm was fed with 8954 data samples for training (4454 of "normal" type and 4500 of "pre-PAF" and "PAF-distant" types combined) and 2985 samples for testing (1485 "normal" and 1500 of "pre-PAF" and "PAF-distant" types combined), that is, with 75.00% of the total data samples for the training phase and the remaining ...

Similar publications

Article
Full-text available
Conventional blood pressure (BP) measurement methods have different drawbacks such as being invasive, cuff-based or requiring manual operations. There is significant interest in the development of non-invasive, cuff-less and continual BP measurement based on physiological measurement. However, in these methods, extracting features from signals is c...