Identifying high-risk patients using artificial intelligence-enabled electrocardiogram: early detection of rehospitalization in post-discharge heart failure patients

Abstract Background Preventing heart failure (HF) rehospitalizations remains a major challenge, and current monitoring strategies have notable limitations. There is limited evidence on the role of artificial intelligence-enabled electrocardiogram (AI-ECG) as a digital biomarker for remote patient monitoring after hospital discharge in patients with HF. We hypothesize that AI-ECG models can predict HF rehospitalization.1 Purpose This study aims to develop and characterize an AI-ECG algorithm for predicting rehospitalization in patients previously hospitalized for HF. Methods This retrospective single-center study analyzed patients hospitalized for acute heart failure for the second time between March 2017 and January 2025 and had available ECG data. AI-ECG was used to analyze ECGs recorded from two months before admission until hospitalization, assessing scores for left ventricular systolic dysfunction (LVSD), left ventricular diastolic dysfunction (LVDD), and myocardial infarction (MI). Based on these AI-ECG scores, we identified changes in ECG patterns associated with HF readmission, developed an algorithm to detect these changes, and characterized its features. Results A total of 891 ECGs from 81 patients and 115 hospitalizations were analyzed. AI-ECG identified distinct patterns for LVSD, LVDD, and MI, with rapid deterioration observed in 40, 15, and 15 of the 115 hospitalizations, respectively (Figure 1). In the LVSD score, the decline group exhibited relatively lower values, approaching the cutoff by approximately 0.1 compared to the stable group. Each decline group showed a significant increase in AI-ECG scores within one month before admission compared to the stable group. There was minimal overlap between the decline groups, as illustrated in Figure 2a. Only a small number of patients exhibited overlapping patterns across these groups: two cases shared both LVDD and MI decline characteristics, two cases demonstrated overlap between LVSD and LVDD, and a single case was common among all three decline groups. The LVSD and LVDD probability scores demonstrated moderate correlation (r = 0.65), indicating that while these dysfunctions often co-occur, they do not fully overlap. (Figure 2b) In contrast, the MI probability score showed no meaningful correlation with either LVSD (r = -0.01) or LVDD (r = 0.05) Conclusion AI-ECG demonstrates potential as a digital biomarker for predicting hospital readmission in patients discharged after HF hospitalization. We provided insights into the characteristics of each model in predicting HF rehospitalization. This approach represents a non-invasive and rapid method that can potentially transform the current paradigm of HF care.