Interpretability in Cardiovascular Diagnostics: Classical Machine Learning vs. Deep Learning

Le résumé fourni par la source

An overview of the integration of Al-based ECG applications in different cardiovascular disorders is given in this study. Globally, one of the major causes of death is cardiovascular diseases (CVD), hence prompt and precise diagnosis is essential. This study’s primary objective is to compare DL with conventional ML models for ECG based CVD diagnosis and to investigate the conflicts between model performance and interpretability. We trained our proposed framework on a publicly available 12-lead ECG dataset using models namely- CNN, LSTM, MLP, SVM, LR and RF are some of the prediction techniques.Recursive Feature Elimination technique is utilized to extract and select important features, while SHAP values were used to ensure interpretability of the model. The performance of DL models applied to structured data did not match that of the classical ML Random Forest and Logistic Regression. Even though DL models have strong predictive capabilities, their opaque nature (black box nature) impedes their applicability in the real clinical practice. The findings emphasize the importance of explainable AI in health care and support the use of hybrid approaches to maintain diagnostic accuracy while enhancing usability and trust in a health care environment.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.