AI-Enhanced ECG for pulmonary hypertension detection: integrating machine learning and deep learning approaches

Abstract Background Pulmonary hypertension (PH) is commonly associated with heart disease and carries significantly prognostic implication. Right heart catheterization (RHC) is the gold standard for diagnosis, but non-invasive detection methods are needed for early identification and risk assessment. This study aimed to develop an electrocardiography (ECG)-based machine learning algorithm for PH detection. Methods A total of 53,069 patients who underwent ECG and echocardiography were enrolled. PH was defined as right ventricular systolic pressure of 40 mmHg on echocardiography, with 1,612 patients also undergoing RHC validation. Machine learning models, including CatBoost, XGBoost, LightGBM, and a deep learning model (ResNet1D18), were developed using an 80/20 training-testing split with 5-fold cross-validation. Results The machine learning models demonstrated strong predictive performance for PH detection. The deep learning model, ResNet1D18 achieved the highest accuracy (96.7%) and area under curve (AUC: 0.953), outperforming other models, including CatBoost (AUC: 0.850), XGBoost (AUC: 0.842), and LightGBM (AUC: 0.851). Feature selection analysis suggests the R wave amplitude in lead V1 and aVF may aid in PH identification. Conclusion Machine learning algorithms effectively detect PH from ECG, with deep learning models showing superior performance. Further feature selection may help refine AI-driven ECG models, enhancing early PH detection and risk stratification.