The Taub Faculty of Computer Science Events and Talks

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, I will present 5 papers (3 published) that describe our journey toward a clinical trial. In our first paper, we designed an automated algorithm to classify short ECG signal strips into 4 categories: normal rhythm, atrial fibrillation, noisy segment, or other rhythm disturbances. We used a feature-based classification to classify the short ECG recordings. Our algorithm obtained a total score (F1) of 0.80 on the hidden dataset. Our algorithm was able to classify AF vs. non-AF and normal vs. abnormal (arrhythmia or noise) records. In our second paper, 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. 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. In our third paper, 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. The network accuracy can be improved by including a small number of unlabeled samples from unknown formats in the training data. Finally, our models achieve high accuracy using signals as input rather than images.In our fourth paper, we aim to introduce a high-resolution ECG interpretation tool designed for real clinical images captured by mobile devices. Our approach capitalizes on the sensitivity of the Jacobian matrix for input images. We showcase interpretability for both morphological and arrhythmogenic cardiac conditions in images captured in clinical environments. The interpretability tool accentuates key signal features with high resolution, aligning with known clinical signs.

In our fifth paper, we present BeatBox AI—an automated diagnostic platform for 12 lead ECGs. This innovative system is crafted as a continually self-optimizing solution, offering scalability and adaptability to diverse populations and various 12-lead ECG layouts. BeatBox AI's diagnostic capabilities were assessed using 4,060 ECGs gathered during a 9-month prospective clinical study. This study involved collaboration with 14 cardiologists from five distinct hospitals. In the comparative analysis between the automatic interpretation and the cardiologists' assessments, the system demonstrated improvement across all operational dimensions. At the concluding evaluation stage, the MCC significantly improved to 0.56, and the system successfully identified a total of 54 distinct cardiac conditions.