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

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence in Medicine

Background:

• Differentiating heart failure phenotypes is crucial for guiding therapy.
• Echocardiography, the standard for differentiating heart failure with preserved ejection fraction (HFpEF) from heart failure with mildly reduced/reduced ejection fraction (HFmr/rEF), may be delayed or unavailable.
• Routine 12-lead electrocardiogram (ECG) data offers a potential alternative for rapid phenotyping.

Purpose of the Study:

• To develop and validate machine learning models for classifying HF phenotypes using routine 12-lead ECG data.
• To assess the performance of different machine learning models in distinguishing HFmr/rEF from HFpEF.
• To identify key ECG features for accurate heart failure phenotyping.

Main Methods:

• Retrospective cohort study of 495 hospitalized patients with heart failure.
• Development of Random Forest (RF), Extreme Gradient Boosting, and Support Vector Machine models using predictors available at or before the index ECG.
• Feature selection using Boruta algorithm to identify key ECG variables.
• Model performance evaluated using Area Under the Curve (AUC) and accuracy via cross-validation and a held-out test set.

Main Results:

• The Random Forest (RF) model demonstrated consistent superior performance.
• A parsimonious RF model using 12 selected ECG variables achieved an AUC of 0.832 and accuracy of 76.4%.
• This model showed no significant performance difference compared to comprehensive models using clinical/laboratory/ECG predictors or all ECG features. Adding chest X-ray cardiomegaly did not improve results.

Conclusions:

• A parsimonious Random Forest model utilizing standard ECG measurements effectively differentiates HFmr/rEF from HFpEF.
• The model provides good discrimination, supporting the use of ECG as an adjunct tool for heart failure phenotyping.
• This approach is particularly valuable when echocardiography is not immediately accessible.