Aplicação de redes neurais para diagnóstico cardíaco: uma abordagem comparativa com machine learning e interpretação via SHAP

Abstract

The diagnosis of cardiac conditions through electrocardiograms (ECGs) is a complex process that demands time and expertise from healthcare professionals. With the advancement of computational techniques, there is an opportunity to develop more accurate and interpretable diagnostic support tools. This work aimed to compare machine learning (ML) and deep learning (DL) approaches in the classification of cardiac conditions using the CPSC2018 database, evaluating both 12-lead and single-lead configurations. The methodology employed a ResNet34 architecture for DL and XGBoost for ML, along with SHAP interpretability techniques. The results demonstrated the superiority of ResNet34, which achieved an F1-Score of 0.815 and accuracy of 96.5% with 12 leads, outperforming XGBoost (0.758 and 95.8%, respectively). In the lead I configuration, ResNet34 maintained robust results, with an F1-Score of 0.774 and accuracy of 95.6%, while XGBoost showed more modest results, with 0.655 and 94.2%. The interpretability analysis revealed that while XGBoost provides clear insight into how specific features influence its decisions, ResNet34 provides a more visual and clinically relevant perspective in both configurations. The study concludes that these techniques have potential to serve as valuable auxiliary tools in cardiac diagnosis, both in clinical settings with 12-lead ECGs and in wearable devices using single-lead, where they could act as relevance sensors to identify signal segments that deserve more detailed evaluation by specialists. In all cases, their application should be complementary to medical expertise.