Problem

Low left ventricular ejection fraction (LEF) is a critical indicator of left ventricular systolic dysfunction (LVSD), which often progresses to symptomatic heart failure if undetected. Early diagnosis is vital for timely treatment and improved prognosis. While echocardiography is the gold standard for assessing ventricular function, it is resource-intensive and not suitable for widespread population screening. Artificial intelligence-enabled electrocardiography (AI-ECG) offers a promising alternative for scalable screening, but current approaches either rely on uninterpretable black-box models or tabular systems dependent on commercial ECG measurement algorithms with suboptimal and non-generalizable performance. There is a clear need for a scalable, effective, and clinically interpretable framework for LEF detection from standard ECGs.

Method

The researchers developed the ECG-based Predictor-Driven LEF (ECGPD-LEF) framework, designed to detect LEF using standard 12-lead ECGs while ensuring interpretability. The framework consists of two main components: a predictor extractor and predictor-based inference models. The predictor extractor is a Transformer-based automatic ECG diagnosis model (ST-MEM), initially pre-trained on several ECG datasets (Chapman, Ningbo, CODE-15%) and then fine-tuned on the PTB-XL dataset to predict 71 conventional ECG diagnoses, outputting continuous probability estimates for each. These 71 probabilistic outputs serve as structured predictors. For LEF inference, two approaches were utilized: a single-predictor approach, where each predictor's value was independently evaluated for LEF detection in a zero-shot-like manner without additional training, and a multi-predictor approach, which jointly modeled all 71 predictors using lightweight tabular classifiers (logistic regression with an l2 penalty or XGBoost). The multi-predictor models were trained on the publicly available EchoNext dataset, comprising 72,475 ECG-echocardiogram pairs. The framework's performance was rigorously evaluated on both an independent internal test set from EchoNext (n=5,442) and a newly constructed external validation dataset called MIMIC-LEF (n=16,017), derived from MIMIC-IV, MIMIC-IV-ECG, and MIMIC-IV-Note. Model performance was quantified using AUROC, AUPRC, and F1 score, and interpretability was assessed using SHAP (SHapley Additive exPlanations) values to identify global and local feature contributions.

Results

The ECGPD-LEF framework demonstrated robust and superior performance compared to the official end-to-end Columbia mini model baseline. The optimal configuration, Transformer-PT-XGBoost, achieved an AUROC of 88.4% and an F1 score of 64.5% in the internal test set, and an AUROC of 86.8% and an F1 score of 53.6% in the external test set. These results significantly surpassed the Columbia mini model's performance (internal AUROC 85.2%, F1 57.9%; external AUROC 80.7%, F1 46.4%). Performance gains were consistently observed across various demographic and clinical subgroups, including those defined by age, sex, race/ethnicity, clinical context, and the presence or absence of structural or valvular heart disease. Interpretability analyses identified several high-impact predictors strongly associated with LEF detection. Notably, "normal ECG" (NORM), "incomplete left bundle branch block" (ILBBB), and "subendocardial injury in anterolateral leads" (INJAL) showed substantial standalone discriminative performance, with internal AUROC values ranging from 75.3% to 81.0% and external AUROC values from 71.6% to 78.6%. This indicates that subtle ventricular dysfunction signals are intrinsically encoded within these structured diagnostic probability representations. SHAP analyses further elucidated global and local predictor contributions, confirming that graded shifts in ECG-derived diagnostic probabilities meaningfully contribute to LEF risk estimation.

Implications

The ECGPD-LEF framework offers a powerful and clinically interpretable solution for scalable screening of low left ventricular ejection fraction from ECGs. By reconciling high predictive performance with mechanistic transparency, it addresses a key limitation of previous black-box AI-ECG models, potentially fostering greater clinician trust and facilitating broader clinical adoption. Its modular design allows for seamless integration with existing AI-ECG systems and future enhancements by incorporating additional clinically established ECG diagnostic features, ensuring scalability and extensibility. The framework's consistent and superior performance across diverse patient subgroups and clinical contexts supports its generalizability and real-world applicability. This approach represents a significant step towards practical and accessible early detection of heart failure, enabling timely therapeutic interventions and ultimately improving patient outcomes.