Using domain adaptation for classification of healthy and disease conditions from mobile-captured images of standard 12-lead electrocardiograms.

12-lead electrocardiogram (ECG) recordings can be collected in any clinic and the interpretation is performed by a clinician. Modern machine learning tools may make them automatable. However, a large fraction of 12-lead ECG data is still available in printed paper or image only and comes in various formats. To digitize the data, smartphone cameras can be used. Nevertheless, this approach may introduce various artifacts and occlusions into the obtained images. Here we overcome the challenges of automating 12-lead ECG analysis using mobile-captured images and a deep neural network that is trained using a domain adversarial approach. The net achieved an average 0.91 receiver operating characteristic curve on tested images captured by a mobile device. Assessment on image from unseen 12-lead ECG formats that the network was not trained on achieved high accuracy. We further show that the network accuracy can be improved by including a small number of unlabeled samples from unknown formats in the training data. Finally, our models also achieve high accuracy using signals as input rather than images. Using a domain adaptation approach, we successfully classified cardiac conditions on images acquired by a mobile device and showed the generalizability of the classification using various unseen image formats.

Automatic classification of healthy and disease conditions from images or digital standard 12-lead electrocardiograms.

Standard 12-lead electrocardiography (ECG) is used as the primary clinical tool to diagnose changes in heart function. The value of automated 12-lead ECG diagnostic approaches lies in their ability to screen the general population and to provide a second opinion for doctors. Yet, the clinical utility of automated ECG interpretations remains limited. We introduce a two-way approach to an automated cardiac disease identification system using standard digital or image 12-lead ECG recordings. Two different network architectures, one trained using digital signals (CNN-dig) and one trained using images (CNN-ima), were generated. An open-source dataset of 41,830 classified standard ECG recordings from patients and volunteers was generated. CNN-ima was trained to identify atrial fibrillation (AF) using 12-lead ECG digital signals and images that were also transformed to mimic mobile device camera-acquired ECG plot snapshots. CNN-dig accurately (92.9-100%) identified every possible combination of the eight most-common cardiac conditions. Both CNN-dig and CNN-ima accurately (98%) detected AF from standard 12-lead ECG digital signals and images, respectively. Similar classification accuracy was achieved with images containing smartphone camera acquisition artifacts. Automated detection of cardiac conditions in standard digital or image 12-lead ECG signals is feasible and may improve current diagnostic methods.

An SVM approach for identifying atrial fibrillation.

OBJECTIVES: We designed an automated algorithm to classify short electrocardiogram (ECG) strips into four categories: normal rhythm, atrial fibrillation, noisy segment, or other rhythm disturbances. APPROACH: The algorithm is based on identification of the R peak and recognition of the other ECG waves. Time-frequency domain features, the average and variability of the intra-beat temporal interval, and the average beat morphology were also calculated. These features (61 features at all) were the input to a support vector machine (SVM) with and without a feed-forward 2-layer neural network consisting of 20 neurons trained on an annotated database. Data were drawn from the PhysioNet Challenge 2017 dataset, consisting of 8528 recordings, of which 60.43% are normal, 0.54% are noisy, 9.04% are AF, and 30% are other rhythm disturbances. The results were validated on 3658 ECG recordings of similar length and percent from each of the four groups. MAIN RESULTS: We used a quadratic SVM classifier with a combination of 61 features to classify the short ECG recordings into one of the four categories mentioned above. The use of an additional neural network to improve the identification of 'other' rhythms that were misclassified as 'normal' did not statistically affect the results. Our algorithm obtained a total score (F1) of 0.80 on the hidden dataset (placing 18th-24th out of all the algorithms participating in the challenge; places 18-24 received the same score). SIGNIFICANCE: Our algorithm was also able to classify AF versus non-AF and normal versus abnormal (arrhythmia or noise) records.

Novel Method to Efficiently Create an mHealth App: Implementation of a Real-Time Electrocardiogram R Peak Detector.

BACKGROUND: In parallel to the introduction of mobile communication devices with high computational power and internet connectivity, high-quality and low-cost health sensors have also become available. However, although the technology does exist, no clinical mobile system has been developed to monitor the R peaks from electrocardiogram recordings in real time with low false positive and low false negative detection. Implementation of a robust electrocardiogram R peak detector for various arrhythmogenic events has been hampered by the lack of an efficient design that will conserve battery power without reducing algorithm complexity or ease of implementation. OBJECTIVE: Our goals in this paper are (1) to evaluate the suitability of the MATLAB Mobile platform for mHealth apps and whether it can run on any phone system, and (2) to embed in the MATLAB Mobile platform a real-time electrocardiogram R peak detector with low false positive and low false negative detection in the presence of the most frequent arrhythmia, atrial fibrillation. METHODS: We implemented an innovative R peak detection algorithm that deals with motion artifacts, electrical drift, breathing oscillations, electrical spikes, and environmental noise by low-pass filtering. It also fixes the signal polarity and deals with premature beats by heuristic filtering. The algorithm was trained on the annotated non-atrial fibrillation MIT-BIH Arrhythmia Database and tested on the atrial fibrillation MIT-BIH Arrhythmia Database. Finally, the algorithm was implemented on mobile phones connected to a mobile electrocardiogram device using the MATLAB Mobile platform. RESULTS: Our algorithm precisely detected the R peaks with a sensitivity of 99.7% and positive prediction of 99.4%. These results are superior to some state-of-the-art algorithms. The algorithm performs similarly on atrial fibrillation and non-atrial fibrillation patient data. Using MATLAB Mobile, we ran our algorithm in less than an hour on both the iOS and Android system. Our app can accurately analyze 1 minute of real-time electrocardiogram signals in less than 1 second on a mobile phone. CONCLUSIONS: Accurate real-time identification of heart rate on a beat-to-beat basis in the presence of noise and atrial fibrillation events using a mobile phone is feasible.