Repurposed ECG Foundation Model Diagnostic Scores Accurately Detect Left Ventricular Dysfunction

A new computational framework demonstrates that diagnostic scores of conventional electrocardiographic findings from foundation models can be repurposed to identify left ventricular systolic dysfunction while outperforming end-to-end deep learning. Heart failure affects over 64 million people worldwide, with prevalence rising across diverse populations. In many individuals, left ventricular systolic dysfunction develops asymptomatically before progressing into symptomatic heart failure with poor prognosis. Low left ventricular ejection fraction, defined as 45 percent or lower, marks this deterioration toward symptomatic failure. Although transthoracic echocardiography remains the clinical gold standard for ventricular assessment, it is resource-intensive, requires specialized operators, and cannot be readily scaled for broad screening. To address this limitation, researchers developed the predictor-driven framework ECGPD-LVEF, which extracts structured diagnostic representations from standard twelve-lead electrocardiograms to evaluate ventricular dysfunction transparently.

Electrocardiography has tracked cardiac electrical activity since Willem Einthoven introduced the technology, earning the Nobel Prize in 1924, but artificial intelligence has expanded its diagnostic scope. Deep learning models now match cardiologists in standard rhythm interpretation and detect subtle patterns linked to structural heart disease imperceptible to human observers. However, clinical adoption for detecting low ejection fraction remains limited due to the opaque black-box nature of deep neural networks. A prior study proposed a random forest model using 555 discrete measurements from commercial software, which outperformed blood biomarker testing but trailed deep learning. Furthermore, that approach was constrained by poor feature portability and inter-device measurement variations. To overcome these deployment barriers, the authors investigated whether diagnostic scores produced by an electrocardiogram foundation model could serve as portable, interpretable predictors without relying on vendor-dependent measurement extraction pipelines.

Developing the predictor extractor relied on four electrocardiogram-only datasets comprising 413,167 recordings. The architecture adopted ST-MEM, a Transformer-based foundation model whose pre-trained weights incorporated representations from the Chapman dataset of 10,646 recordings, the Ningbo dataset of 34,905 examinations, and the CODE-15% cohort, a stratified sample of 345,779 examinations from 233,770 patients. Signals originally sampled at 500 Hertz were resampled to a standardized frequency of 250 Hertz across twelve leads with ten-second durations. To configure the predictor extractor, the model underwent supervised post-training on the PTB-XL database, which contains 21,837 clinical records from 18,885 patients recorded between 1989 and 1996 on Schiller AG devices. Equipped with a final sigmoid activation function, the resulting post-trained network outputs 71 continuous diagnostic scores between zero and one, each representing a distinct conventional clinical electrocardiographic interpretation encompassing rhythm disturbances, conduction blocks, ischemic alterations, and chamber hypertrophies.

Validation on the PTB-XL test set confirmed that the post-trained Transformer, denoted Transformer-PT, achieved strong discriminative capacity across 71 diagnostic categories. Transformer-PT recorded a macro area under the receiver operating characteristic curve of 94.5 percent, a macro precision-recall curve of 41.4 percent, and an F1 score of 38.3 percent. These results surpassed the baseline Transformer without post-training, which attained an area under the curve of 89.8 percent and precision-recall curve of 30.7 percent. Although precision-recall and F1 scores varied across categories because of statistical class imbalance among rare clinical diagnoses, the model maintained consistent ranking accuracy across all classes. Crucially, the downstream detection framework uses continuous predictor scores rather than threshold-dependent binary decisions. This ranking precision ensures that reliable diagnostic information transfers into subsequent modeling stages, even for ultra-rare diagnostic findings that do not cross standard binary classification boundaries.

Downstream model development and internal validation utilized EchoNext, a paired dataset collected at Columbia University Irving Medical Center between 2008 and 2022. EchoNext contains 82,543 paired electrocardiogram and echocardiogram records from 36,286 adult patients, acquired using GE MUSE systems at 250 Hertz. Following official patient-level separation protocols, the cohort was divided into a training set of 72,475 examinations with 16,962 low ejection fraction cases, a validation set of 4,626 examinations with 866 cases, and an internal test set of 5,442 examinations with 962 cases. Reference ejection fraction values were extracted from Siemens Syngo Dynamics and Philips Xcelera databases, defining low ejection fraction as 45 percent or lower. An electrocardiogram was labeled positive if recorded within one year prior to an echocardiogram showing an ejection fraction of 45 percent or less, whereas negative labels required an echocardiogram exceeding 45 percent.

To evaluate external transportability, researchers constructed an independent test cohort from MIMIC-IV, MIMIC-IV-ECG, and MIMIC-IV-Note databases, representing patients evaluated at Beth Israel Deaconess Medical Center between 2008 and 2019. From an initial pool of 1,113,547 pairs derived from 521,654 electrocardiograms and 247,313 clinical notes across 103,505 patients, exclusions removed notes lacking ejection fraction keywords and waveforms with missing numerical values. A large language model extracted documented ejection fraction numbers from clinical discharge summaries, which were matched against MIMIC-ECG-LVEF records to confirm concordance. Retaining the single pair with the shortest interval per patient produced a final external test set of 16,017 examinations, including 2,517 low ejection fraction cases. This cohort exhibited notable demographic differences from EchoNext, comprising 54.9 percent male individuals, 73.0 percent White patients, and a high proportion of emergency encounters at 57.3 percent, providing a diverse real-world evaluation setting.

In initial evaluations, investigators tested whether individual diagnostic scores could identify low ejection fraction in a zero-shot-like manner without task-specific training. Several individual predictors demonstrated notable standalone discriminative capacity. Eight predictors achieved areas under the receiver operating characteristic curve between 71.0 percent and 81.0 percent on the internal test set, with five maintaining values between 70.7 percent and 78.8 percent externally. The normal electrocardiogram score, termed NORM, was the strongest standalone predictor, yielding an internal area under the curve of 81.0 percent and an external area of 78.8 percent. Incomplete left bundle branch block and anterolateral subendocardial injury also displayed strong internal discrimination, reaching areas under the curve of 80.0 percent and 75.3 percent. Crucially, the optimal low ejection fraction threshold on NORM was 0.003641, substantially lower than the 0.370 threshold used for classifying normal rhythms.

For composite predictive modeling, the multi-predictor framework combined all 71 diagnostic scores using lightweight tabular models, comparing regularized logistic regression against XGBoost. On the internal EchoNext test set, the primary configuration pairing Transformer-PT with XGBoost achieved an area under the receiver operating characteristic curve of 88.4 percent, a precision-recall curve of 69.2 percent, and an F1 score of 64.5 percent. This configuration significantly outperformed the official Columbia mini benchmark, an end-to-end convolutional neural network combining waveforms with seven tabular features, which achieved an area under the curve of 85.2 percent, precision-recall curve of 59.9 percent, and F1 score of 57.9 percent. Controlled ablation experiments confirmed that both framework components contributed to these gains. Holding the extractor constant, XGBoost consistently exceeded logistic regression. Similarly, holding the tabular model constant, post-training the Transformer extractor produced sustained performance increases across all evaluation metrics.

When applied to the external MIMIC test cohort, the Transformer-PT and XGBoost pipeline maintained robust discrimination, achieving an area under the receiver operating characteristic curve of 86.9 percent, precision-recall curve of 57.6 percent, and F1 score of 53.8 percent. These metrics remained numerically superior to the Columbia mini model, which attained an area under the curve of 80.8 percent and F1 score of 47.7 percent, though external baseline comparisons were considered secondary because several baseline input features required approximate reconstruction from fiducial points. Subgroup analyses confirmed consistent performance across diverse clinical strata. On the internal test set, areas under the curve ranged from 85.8 percent to 92.5 percent across age, sex, race, and encounter settings, while external subgroup discrimination spanned from 79.4 percent to 90.2 percent. The model also maintained high discrimination in patients with and without concurrent structural or valvular heart disease, consistently outperforming the convolutional benchmark.

In addition to strong predictive discrimination, the tabular architecture provided model interpretability through SHAP analysis, generating both global population summaries and case-specific explanations. Global feature importance revealed that cumulative SHAP contributions increased as more diagnostic predictors entered the model, with NORM, incomplete left bundle branch block, and lateral ischemic changes contributing the most. Because conventional diagnostic categories share electrophysiological correlations, these metrics represent conditional feature attributions rather than independent causal influences. At the individual level, SHAP waterfall plots demonstrated how specific diagnostic representations drove predictions for individual positive and negative cases. Notably, substantial variation in SHAP attributions occurred at low predictor values far below conventional diagnostic thresholds. This demonstrates that continuous foundation model outputs capture subclinical electrophysiological deviations associated with ventricular remodeling before abnormalities trigger categorical diagnostic labels, explaining how artificial intelligence detects ventricular impairment from subtle waveform features.

The authors outlined several methodological limitations that temper immediate clinical translation. Reference labels relied on clinical echocardiography reports, which are subject to known inter-observer variability in image acquisition and quantitative interpretation. Additionally, because both EchoNext and MIMIC datasets involved patients evaluated in clinical settings who underwent echocardiography, selection bias may lead to higher ventricular dysfunction prevalence than in general asymptomatic populations. Model calibration also showed noticeable deterioration in the external test set, where the expected calibration error rose to 0.120 compared with 0.021 internally, accompanied by risk overestimation that requires cautious interpretation in screening environments. Furthermore, the foundation model was post-trained on a moderately sized dataset of 21,837 recordings due to limited public repositories, suggesting larger training corpora could yield further improvements. Finally, the retrospective data lacked longitudinal follow-up to evaluate whether predictions provide independent prognostic value. The researchers emphasized that prospective clinical validation in broader international cohorts and population-based screening settings remains necessary.

Subject of Research: Detection of low left ventricular ejection fraction using electrocardiogram foundation model diagnostic scores

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

Article References: Zhou, Y., Hao, T., Cai, Z., Zhu, H., He, K., Liu, J., Fan, X., & Yuan, J. (2026). Detecting low left ventricular ejection fraction from ECG using an interpretable and scalable predictor-driven framework: a model development and validation study. eClinicalMedicine, 102, Article 104331. https://doi.org/10.1016/j.eclinm.2026.104331

Image Credits: AI Generated

DOI: 10.1016/j.eclinm.2026.104331

Keywords: electrocardiography, artificial intelligence, left ventricular ejection fraction, heart failure, echocardiography, foundation models, machine learning, XGBoost, SHAP analysis, PTB-XL dataset, EchoNext benchmark, MIMIC-IV, structural heart disease, left ventricular systolic dysfunction, ST-MEM architecture, model interpretability, clinical decision support, diagnostic classification