Abstract 11784: A Deep-Learning-Based Multi-Modal Phonocardiogram and Electrocardiogram Processing Framework for Screening Left Ventricular Systolic Dysfunction Using a Wearable Cardiac Patch

Abstract

Introduction: Left ventricular systolic dysfunction (LVSD) is associated with poor health outcomes. Previous study has demonstrated the effectiveness of electromechanical activation time (EMAT) derived from phonocardiography (PCG) and electrocardiography (ECG) signals in identifying LVSD. However, high throughput and well-performed algorithm is needed in automatically detecting LVSD. Hypothesis: An artificial intelligence (AI) algorithm was developed and validated in detecting LVSD. Methods: Using ECG and PCG data collected by wearable patch from 1020 admitted patients in Ruijin hospital, we trained an AI algorithm to detect the LVSD, which was determined by ejection fraction <50% from echocardiogram. A separate 590 patients were taken as independent test set. The AI algorithm followed a two-step detection structure, where the first step is to use LSTM learning to quantify heartbeat and heartbeat interval in ECG, and the second step is to apply convolution neural network (CNN) in estimating optimal EMAT cutoff points in detecting LVSD. To overcome the challenge of multi-modal signal training, we develop an alignment-based contrastive learning approach that generates shared feature space embeddings for ECG and PCG signals. Results: Among training and test sets, 18.2% and 26.1% were LVSD, respectively. The AI algorithm yields an AUROC of 0.92 in detecting LVSD with sensitivity of 85.63%, and specificity of 79.39%. Among the test set, model yielded an AUROC of 0.86, sensitivity of 81.29%, and specificity of 77.13%. Conclusions: A wearable device incorporated with a high performed AI algorithm can be a viable, convenient, cost-effective tool in screening LVSD.