Foundation ECG embeddings enable data-efficient detection of hypertrophic cardiomyopathy in a limited labeled cohort

Background

Supervised deep learning (DL) for ECG analysis typically requires very large labelled datasets, which are rarely available for uncommon conditions such as hypertrophic cardiomyopathy (HCM). Foundation models trained using self-supervised learning (SSL) enable transfer of pre-learned cardiac representations and may allow accurate classification in data scarce settings.

Purpose

To evaluate whether ECG embeddings derived from a large foundational model can support detection of HCM and discriminate it from general controls and left ventricular hypertrophy (LVH) using a minimal labelled dataset, and to explore the effect of medical therapy on model-derived HCM probability.

Method

A self-supervised foundation model trained in-house on 840,000 unlabelled ECGs was used to generate 512-dimensional embeddings. A total of 147 subjects were included: 49 HCM patients, 49 LVH patients, and 49 controls, each group comprising 33 males and 16 females. LVH patients and controls were matched to HCM patients in a 1:1 ratio by age and sex. Controls were additionally matched by ECG acquisition date. For both HCM and LVH subjects, the first available ECG was selected.

Logistic regression classifiers were trained on frozen embeddings. Performance was estimated using nested cross-validation with a leave-one-out outer, and 4-fold inner scheme. Matched pairs were kept within the same split to prevent data-leakage. Two classification tasks were evaluated: HCM versus controls and HCM versus LVH.

In a subset of 26 HCM patients, unbiased out-of-sample probabilities were generated using the respective leave-one-out models. The last ECG before treatment initiation and the most recent ECG during therapy were compared using a paired Wilcoxon test to assess changes in model-derived HCM probability.

Results

Using only 49 labelled HCM cases, the model demonstrated good discrimination. For HCM versus controls, patient-level AUROC was 0.86 (95% CI 0.78 – 0.93). In the clinically more challenging comparison of HCM versus LVH, AUROC remained robust at 0.77 (95% CI 0.67–0.85).

In the medication subset (N=26), model-derived HCM probability decreased significantly after treatment initiation (mean change -0.13, median change -0.032 [IQR -0.27 to 0.24], p-value = 0.049), suggesting sensitivity of ECG embeddings to therapy-related phenotypic modulation (Figure 1).

Conclusions

Foundation ECG embeddings enabled detection of HCM in this data-limited setting and retained discrimination against LVH despite minimal labelled training data. This approach provides a scalable strategy for AI development in rare cardiovascular diseases, where traditional large-scale supervised learning is often not feasible. Preliminary findings obtained through rigorous out-of-sample validation suggest that embedding-based model outputs may reflect treatment-associated changes in cardiac phenotype. These findings motivate validation in larger, independent cohorts.