Aim: Early identification of systolic heart failure (HF) from the electrocardiogram (ECG) could support scalable screening, but prior ECG models rely on labels derived from echocardiography (echo) that may be difficult to obtain at large scale. In this case study, we investigated whether systolic HF can be predicted from ECGs using weak labels derived from routine diagnosis codes.
Methods: We used the Harvard-Emory ECG Database (HEEDB) for training and derived a patient-level binary target from routinely collected ICD diagnosis codes, yielding a development cohort of 1,625,032 ECGs. A supervised 12-lead ECG classification model based on a 1D ResNet architecture was externally validated on 5,442 patients from EchoNext using echo-derived labels.
Results: Internal validation achieved an AUC of 0.962 (CI: 0.961–0.963) while external validation yielded an AUC of 0.797. Additionally, predicted probabilities showed an inverse relationship with continuous LVEF (Pearson's r = -0.467, p < 0.001), suggesting that model output may indirectly reflect left ventricular systolic function and thus the amount of blood ejected into the body with each heartbeat.
Conclusions: These findings suggest that ECG models for systolic HF can be developed using ICD-based labels from routine clinical data. The external validation results and the inverse correlation with LVEF suggest that the model output was related to systolic function. However, the performance decrease in external evaluation underscores the importance of careful interpretation and validation against imaging-derived endpoints.