Comparative Study of Explainable Machine Learning Techniques for Interpretable Healthcare Risk Prediction Systems

Abstract

This paper provides a comparative approach to machine learning (ML) models and explainable artificial intelligence (XAI) methods of developing interpretable healthcare risk prediction systems. Based on a quantitative experimental design, three ML models, such as decision tree, random forest and neural network, have been tested, together with explainability techniques, including SHAP, LIME and feature importance. To evaluate the performance of models with respect to accuracy, precision, recall and interpretability factors, a structured healthcare dataset containing patient records was utilized. The findings show that the neural networks were the most accurate in predicting, and the decision trees were the most interpretable. SHAP was the most consistent and effective XAI method to explain model predictions, compared to LIME and feature importance methods. There was a strong negative correlation between accuracy and interpretability, which emphasized the nature of a trade-off between accuracy and interpretability in model selection. The results highlight that explainability methods combine with high-performing models can ensure more transparency, trust, and usability in clinical decision-making. The research finds that hybrid solutions that integrate predictive accuracy with strong interpretability provide the most feasible solution to healthcare AI systems. Such reflections can help develop ethical, reliable, and clinically applicable machine learning models.