Explainable AI-Based Heart Disease Classification: A Deep Learning Framework with Limited Features and Clinical Interpretability for Resource-Constrained Healthcare Settings

Abstract

Background: Cardiovascular disease remains the leading cause of death globally, necessitating accurate early detection systems, particularly in resource-limited areas where comprehensive diagnostic tools are unavailable. Current machine learning approaches for heart disease prediction are often clinically non-interpretative and require numerous features, limiting their application in resource-limited settings.

Objective: This study aims to develop and validate an explainable artificial intelligence (AI) framework for heart disease classification using a minimal feature set while maintaining high prediction accuracy and providing clinically understandable explanations.

Method: We implemented and compared three machine learning models (Logistic Regression, Random Forest, and Deep Neural Network) on the UCI Heart Disease dataset (n=297) with feature selection that reduced variables from 20 to 8 relevant clinical parameters. Model interpretability was enhanced using SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations) techniques. Performance evaluation included AUC (Area Under the Curve), precision, recall, and F1-score metrics.

Results: Logistic regression achieved the highest performance with AUC=0.9699, precision=1.000, and recall=0.821. Random Forest showed AUC=0.9431 with balanced precision (0.917) and recall (0.786), while Deep Neural Network achieved AUC=0.9531. Clinical interpretation revealed maximum heart rate (thalach), type of chest pain (cp), and age as the most significant predictive features. The approach with limited features maintained high accuracy while reducing computational complexity by 60%.

Conclusion: This explainable AI framework demonstrates superior performance with clinical interpretability, making it suitable for implementation in healthcare facilities with limited resources. Minimal feature requirements and transparent decision-making processes enhance its practical application for early detection of heart disease and clinical decision support.