Heart Disease Prediction and Interpretability

Abstract

Interpretable machine learning (ML) plays a crucial role in healthcare, where understanding why a model makes a prediction can be as important as the prediction itself. This project explores the use of LIME (Local Interpretable Model-agnostic Explanations) and SHAP (SHapley Additive exPlanations) for explaining heart disease predictions made by several classifiers trained on the UCI Heart Disease dataset. Four models—logistic regression, decision tree, random forest, and XGBoost—were implemented, with XGBoost achieving the highest accuracy of 0.833. LIME provided intuitive, instance-level explanations, while SHAP offered consistent, theory-based global insights. By combining both tools, this study demonstrates how interpretable ML can balance transparency and accuracy in healthcare applications.

---

1. Introduction

1.1 Background and Motivation

Machine learning models have become powerful tools for medical diagnosis, offering high predictive accuracy in detecting diseases. However, complex models—like ensemble or boosted trees—often act as black boxes, providing little insight into their decision-making processes. In high-stakes areas such as healthcare, this opacity limits trust and practical adoption.

To address this, interpretable ML methods have emerged to explain model predictions post-hoc. Two widely adopted approaches are LIME, which approximates local decision boundaries with simple interpretable models, and SHAP, which attributes feature importance using game-theoretic Shapley values. Both methods aim to bridge the gap between accuracy and interpretability, helping medical professionals understand, validate, and trust predictive systems.

1.2 Aim

The goal of this project was to build and interpret ML models for heart disease prediction, focusing on how LIME and SHAP provide complementary insights into model behavior. The analysis emphasizes reproducibility, modular design, and interpretability evaluation.

---

2. Data: UCI Heart Disease Dataset

The dataset contains 303 patient records with 13 clinical features such as chest pain type, cholesterol level, and resting blood pressure. The original multiclass labels were simplified into a binary classification problem:

• 0 – No heart disease
• 1 – Presence of heart disease

This transformation makes the task suitable for interpretable binary classification, where both global feature importance and individual explanations can be effectively analyzed.

---

3. Methods and Experiments

3.1 Model Implementation

Four models were trained and evaluated:

• Logistic Regression (with momentum optimization) – implemented from scratch for stable and fast convergence.
• Decision Tree – using recursive binary splitting with pruning to avoid overfitting.
• Random Forest – combining multiple trees through bootstrapping and feature bagging.
• XGBoost – a gradient-boosted ensemble for benchmark comparison.

All models were evaluated using 5-fold stratified cross-validation with accuracy and mean squared error metrics.

3.2 Interpretability Framework

Two post-hoc explanation methods were applied:

• LIME was implemented from scratch, generating local surrogate models for each prediction by perturbing inputs and fitting a weighted linear model.
• SHAP values were computed using the official library, providing additive, consistent feature attributions grounded in Shapley theory.

A novel extension, Aggregated LIME, was introduced to improve stability by averaging explanations across multiple runs.

---

4. Results

4.1 Model Performance

Model	Accuracy
Logistic Regression	0.817
Decision Tree	0.700
Random Forest	0.817
XGBoost	0.833

XGBoost achieved the highest accuracy, confirming the strength of ensemble methods for medical prediction tasks.

4.2 Interpretability Insights

Both SHAP and LIME consistently identified the same key predictors:

• Thalassemia (thal)
• Chest pain type (cp)
• Fluoroscopy vessel count (ca)
• ST depression (oldpeak)

SHAP produced stable and globally consistent results, aligning with clinical understanding.
LIME provided clear, local explanations tailored to individual cases but required aggregation for stability. The Aggregated LIME approach successfully revealed which features consistently influenced predictions across runs.

---

5. Conclusion

This project demonstrates that interpretable ML techniques can enhance transparency in healthcare prediction systems without sacrificing accuracy. SHAP delivers robust, theoretically sound explanations, while LIME offers accessible, instance-specific insights. Combining both methods forms a comprehensive interpretability pipeline that balances depth, reliability, and usability—crucial qualities for real-world medical applications.