Work

Detecting low left ventricular ejection fraction from ECG using an interpretable and scalable predictor-driven framework: a model development and validation study

Year: 2026
Type: article
Abstract: Background Low left ventricular ejection fraction (LVEF) frequently remains undetected until progression to symptomatic heart failure, underscoring the need for effective approaches to earlier detection of low LVEF. Although artificial intelligence-enabled electrocardiography (AI-ECG) has shown promise, existing approaches rely solely on end-to-end black-box models with limited interpretability or on tabular models with limited applicability. Methods In this retrospective multicohort study, we developed and validated a predictor-driven framework that reuses diagnostic scores of conventional ECG findings derived from an ECG foundation model for the detection of low LVEF. The foundation model was developed using 413,167 ECGs from four independent publicly available ECG-only datasets collected across multiple countries, locations, and time periods. The official test set of the EchoNext benchmark, comprising ECG–echocardiography (ECHO) pairs collected at Columbia University Irving Medical Center between 2008 and 2022, was used as the internal test set (n = 5442). We constructed an external test set (n = 16,017) from MIMIC-IV-ECG and related data, comprising ECG–clinical note (Note) pairs collected at Beth Israel Deaconess Medical Center between 2008 and 2019. The primary outcome was low LVEF (LVEF ≤ 45%), assessed by ECHO or Note documentation within one year after the ECG. The model-derived diagnostic scores for each ECG finding were evaluated for low LVEF discrimination in the internal and external test sets, neither of which was used to develop the ECG foundation model. We further trained tree-based models using these diagnostic representations on the official EchoNext training set (72,475 paired ECG–ECHO records), followed by evaluation on the internal and external test sets. Explainability analyses were performed to identify key predictive features and characterize ECG-related patterns associated with low LVEF. Findings Diagnostic scores for conventional ECG findings showed discriminative performance for low LVEF across cohorts, with area under the receiver operating characteristic curve (AUROC) values up to 81.0% for individual predictors. The tree-based model achieved an AUROC of 88.4% (95% confidence interval (CI), 87.1–89.5) in the internal cohort, outperforming the official end-to-end baseline on the EchoNext benchmark, and maintained discrimination in the external cohort, with an AUROC of 86.9% (86.2–87.7). Normal ECG (NORM), incomplete left bundle branch block (ILBBB), and ischemic changes in lateral leads (ISCLA) ranked 1–3 by internal single-predictor AUROC and had the same ranking in the SHapley Additive exPlanations (SHAP)–based analysis. Interpretation Diagnostic scores of conventional ECG findings from ECG foundation models contain information relevant to the detection of ventricular dysfunction. These representations can be directly reused for low LVEF detection and further integrated into interpretable predictive models, providing a scalable and extensible computational framework for existing AI-ECG systems. Future research should validate the framework in population-based screening settings and larger international cohorts. Funding This work was partially supported by the CAMS Innovation Fund for Medical Sciences and the National Natural Science Foundation of China.
Language: English
Field: Medicine
Open Access status: gold
Estimated APC paid: $5,780
OpenAlex ID: W7146966859
FWCI:
Cites: 40
Cited by:
Related to: 10