CardioXAI: A Clinical-Grade Explainable AI System for Heart Disease Prediction

CardioXAI v2.0 is an explainable machine learning system designed specifically for clinical decision support. It combines logistic regression prediction with SHAP interpretability analysis to address the "black box" problem in medical AI, balancing prediction accuracy and explanation stability, and providing medical professionals with a transparent and trustworthy auxiliary diagnostic tool. The project was developed by a team from the Vishwakarma Institute of Technology in Pune, India, with the goal of translating cutting-edge explainable AI technology into clinically usable solutions.

Heart disease is one of the leading causes of death globally, and early risk assessment is crucial. Traditional machine learning models are limited in clinical deployment due to their "black box" nature (doctors cannot understand the reasons behind predictions). CardioXAI v2.0 emerged to address this; its core objectives are to balance prediction accuracy and explanation stability, providing a transparent and reliable AI-assisted tool. The team is from the Vishwakarma Institute of Technology in Pune, India, and is committed to translating explainable AI technology into clinical solutions.

CardioXAI uses a modular full-stack architecture (React frontend, Flask backend, machine learning model layer, MongoDB database) and is deployed via Docker containerization to ensure scalability and ease of deployment. Core features include: binary heart disease risk prediction (low/medium/high risk), probabilistic risk score (0-100%), SHAP waterfall chart visualization, personalized clinical explanations, and professional PDF report generation. User functions include JWT secure authentication (7-day token), MongoDB storage for user/health records, three-step registration process, interactive risk trend dashboard, report history management, and mobile-responsive design.

CardioXAI deeply applies the SHAP method to compute precise linear SHAP decomposition for each patient (equivalent to shap.LinearExplainer). Key implementation details: explanations are computed based on the patient's actual input, individualized feature contribution values (not average), mathematical precision (LinearExplainer ensures faithfulness), and clinical alignment (SHAP values are consistent with medical risk factors). The model is trained on the UCI Cleveland Heart Disease Dataset, which contains 297 preprocessed patient records and 13 clinical attributes (age, gender, chest pain type, etc.).

Test set accuracy is 83.3%, AUC-ROC is 0.9487, and five-fold cross-validation AUC-ROC is 0.9012, indicating good generalization ability. Explanation stability validation: the cosine similarity of similar cases for logistic regression is approximately 0.71 (better than random forest's 0.20); removing top SHAP features leads to an AUC-ROC drop of about 0.07, confirming that explanations reflect the model's actual behavior; SHAP explanations align with known medical risk factors (e.g., age, cholesterol, blood pressure have positive contributions, and exercise-induced angina is a strong positive indicator).

CardioXAI embodies best practices in medical explainable AI, balancing accuracy with explanation understandability/stability, and bridging AI technology with clinical practice. Future plans include: integrating more XAI methods (LIME, Integrated Gradients), comparative analysis of neural networks, EHR system integration, multilingual support, and a mobile application version. The project's user-centric, explanation-focused design philosophy is worthy of reference in the medical AI field.