Patient-Level Modeling and Classification of ST-segment in 12-Lead Electrocardiogram for Automated Myocardial Ischemia Detection

The earliest manifestations of myocardial ischemia have been reported in the ST-segment. Hence, accurate detection of ST-segment changes (ST elevation (STE) and ST depression (STD)) in 12-lead ECG recordings is essential for diagnosing myocardial ischemia. This study proposes a clinically relevant patient-level ST-segment classification approach using 12-lead ECG recordings, suitable for varied-length recordings. The ST-segment is modeled using least-squares polynomial regression, resulting in an STp segment. Handcrafted features are then extracted from this segment and its amplitude envelope. Seven ensemble machine learning classifiers-balanced random forest (BRF), balanced bagging classifier with a support vector machine (BBC-SVM) using a radial basis function kernel, and BBC with a multi-layer perceptron (BBC-MLP), gradient boosting (GB), adaptive boosting (AB), extreme gradient boosting (XGB), and light gradient boosting (LGB))-effective at handling class imbalance, are evaluated and compared. The LGB classifier outperformed the others, achieving F1 scores of 0.94 (Normal), 0.86 (STD), and 0.72 (STE), resulting in an overall average of 0.84 in the 12-lead case. A specific six-lead combination (I, II, aVL, V1, V4, and V6) achieved an average F1 score of 0.83. Additionally, the proposed approach was validated on the European ST-T database and yielded intra-patient average F1 scores ranging from 0.87 to 1. The proposed approach demonstrated state-of-the-art performance while ensuring clinical understanding and model transparency.Clinical relevance- This study presents a clinically interpretable and effective ST-segment classification framework at the patient-level, using handcrafted features derived from 12-lead ECGs. While achieving good performance in distinguishing Normal, STD, and STE, the approach holds potential for clinical application, assisting cardiologists in the detection and decision-making of ST-segment changes.