Integrating Manual ECG Feature Extraction with Ensemble Learning for Myocardial Infarction Diagnosis

Abstract Background: Myocardial infarction (MI) is a leading cause of mortality worldwide. Electrocardiograms (ECGs) are primary diagnostic tools but face limitations with high-dimensional, imbalanced data. Combining manual ECG waveform feature extraction with machine learning may improve diagnostic accuracy. Aim: The aim of this study is to improve the accuracy and efficiency of MI diagnosis by integrating manually extracted ECG waveform features with ensemble machine learning models and evaluating their performance on a comprehensive, large-scale ECG dataset. Methods: We utilized the PTB-XL dataset, which consists of 15,014 12-lead ECG recordings. Manual feature extraction was conducted on 94 waveform-specific variables, including the durations and areas of P-waves, QRS complexes, and T-waves. The preprocessing steps included signal filtering, detrending, and data augmentation to mitigate noise and address class imbalance issues. Three ensemble learning algorithms – XGBoost, Random Forest, and AdaBoost – were trained and validated using a 70%, 20%, and 10% split for training, validation, and testing, respectively. Results: Compared with the other models, the XGBoost model demonstrated superior performance, with an accuracy of 86.12%, a sensitivity of 84.81%, a specificity of 87.43%, and an area under the receiver operating characteristic curve (AUROC) of 93.59%. Subcategory analyses indicated strong diagnostic performance in detecting anterior MI (AUROC: 94.09%), inferior MI (AUROC: 90.14%), and lateral MI (AUROC: 89.76%). The ST interval emerged as the most influential feature for accurate diagnosis. Conclusion: Manual ECG feature extraction combined with ensemble learning improves MI diagnosis. Future work should explore automated deep learning and additional physiological features.