Automatic localization of myocardial infarction using vectorcardiography

For patients experiencing myocardial infarction (MI), localizing the affected cardiac region using electrocardiography (ECG) can reduce the time to reperfusion therapy, reducing morbidity and mortality. Extracting relevant information from ECG signals is not trivial, and computational methods have been developed aiming to assist physicians in making faster and better decisions in emergency situations. However, their clinical adoption remains limited due to the high false alarm rates consequence of the low generalizability of these methods. This research compares the performance of three machine learning techniques - Lasso, Support Vector Machine, and Gradient Boosting Machine - with varying degrees of complexity in localizing MI. Vectorcardiography-derived features were used as input to the models due to their ability to capture spatial and temporal information regarding the heart's electrical activity. An autoencoder was employed to smooth the feature space, facilitating more efficient model training and improving generalization. To further address generalizability challenges, an inter-patient validation approach was employed. Models were trained on the PTB-XL dataset and externally validated on the PTB Diagnostic dataset. Results demonstrate that Lasso, a simpler model, achieved the highest AUC of 0.74 on the external dataset, outperforming more complex models such as SVM (0.72) and GBM (0.68). Also, the combination of Lasso with the autoencoder provided superior generalization compared to other state-of-the-art methods reported in the MI localization literature. This highlights the proposed method's suitability for clinical settings, where model generalizability and reliability are critical. Furthermore, our method offers the advantage of explainability, allowing the extraction of clinical and physiological insights from the data and bridging the gap between computational methods and clinicians.