A multilabel deep learning model for the detection of conduction and rhythm disorders from PDF outputs of a widely available portable ECG Device

Abstract Background Wearable and portable devices with ECG capabilities can provide broad access to screening for cardiovascular disorders. However, most are approved for the detection of only atrial fibrillation. Most algorithms for detecting rhythm and conduction disorders are trained and tested on 12-lead ECGs or the lead I of clinical ECGs, with an unclear role for extending the diagnostic range of wearable and portable devices. Moreover, no current algorithms enable cardiovascular diagnosis on data available to end users, who largely have access to the PDF outputs of their data. Purpose To validate a novel artificial intelligence (AI) model trained to identify conduction disorders and arrhythmias on synthetic ECG outputs from lead I data of clinical ECGs, to detect the same disorders on cardiologist-annotated PDF outputs from the Alivecor KardiaMobile single-lead ECGs (Kardia ECGs). Methods We extracted lead I data from 3,076,352 clinical ECGs across 5 hospitals of a large U.S.-based hospital system from 2000-2024 and simulated wearable PDF outputs from commercial devices using a novel plotting approach that replicated the variations in plotted formats. Gold-standard cardiologist written diagnosis statements were processed to generate 8 clinical labels spanning conduction and rhythm disorders, and we developed a multilabel model to predict these labels using an EfficientNet-B3 convolutional neural network. We examined the performance of the model on 24,808 Kardia ECGs annotated by cardiologists. Results The model achieved a high performance on clinical ECGs in a held-out subset of the synthetic lead I ECG printouts resembling wearable outputs (AUC > 0.9 for all 8 labels). In the independent 24,808 Kardia ECGs, annotated by experts, the model achieved high performance across all labels. For the rhythm disorders atrial fibrillation or atrial flutter, premature atrial contraction, and premature ventricular contraction, the model had AUROCs of 0.99, 0.91, and 0.95, respectively. For the conduction disorders 1st-degree AV block and 2nd or 3rd-degree AV blocks, the model had AUROCs of 0.94 and 0.99, respectively. Finally, for supraventricular tachycardias, wide QRS, and paced ECGs, the model had AUROCs of 0.998, 0.98, and 0.98, respectively. Conclusions We demonstrate that a deep learning model trained to identify rhythm and conduction disorders based on synthetic ECGs where lead I data from clinical 12 lead ECGs are used to replicate real-world PDF outputs of wearable ECGs, generalizes to real-world wearable and handheld single-lead devices.