Differentiating HFmr/rEF from HFpEF using standard 12-lead ECG measurements: an interpretable machine learning study

Differentiating heart failure (HF) with mildly reduced/reduced ejection fraction (HFmr/rEF) from HF with preserved ejection fraction (HFpEF) guides therapy but echocardiography may be delayed or unavailable. We developed and validated machine learning models using routine 12-lead ECG data to classify HF phenotypes.

In this retrospective cohort of hospitalised patients with HF, predictors available at or before the index ECG were used. HFmrEF was pooled with HFrEF (left ventricular ejection fraction <50%) for model development. Data were split 70/30 into training and held-out test sets. Random forest (RF), Extreme Gradient Boosting and support vector machine models were trained and tuned using fivefold cross-validation in the training set. Boruta was used to select key ECG features. Test-set performance was evaluated by area under the curve (AUC) and accuracy; AUCs were compared using DeLong's test.

Overall, 495 patients were included (254 HFmr/rEF; 241 HFpEF). RF consistently performed best. Using ECG features alone, RF achieved an AUC of 0.821 (95% CI 0.752 to 0.890) and accuracy of 75.7%. A parsimonious RF model using 12 Boruta-selected ECG variables achieved an AUC of 0.832 (95% CI 0.766 to 0.899) and accuracy of 76.4%, with no significant AUC difference versus a comprehensive RF model using clinical/laboratory/ECG predictors (AUC 0.804, 95% CI 0.734 to 0.875; p=0.288) or the model using all ECG features (p=0.092). Adding X-ray cardiomegaly did not improve performance.

A parsimonious RF model based on a small set of standard ECG measurements differentiates HFmr/rEF from HFpEF with good discrimination, supporting ECG as an adjunct for phenotyping when echocardiography is not immediately available.