LiverXAI: An Explainable AI-Driven Liver Disease Prediction System

Artificial intelligence is widely used in medical diagnosis, but the opacity of black-box models hinders their clinical implementation. The LiverXAI project builds an explainable AI (XAI)-driven liver disease prediction system that combines machine learning and XAI technologies to achieve disease prediction, feature importance analysis, and stability testing. It enhances the transparency and credibility of medical AI through an interactive web application, aiming to address the need for explainability in AI decisions from doctors, patients, and regulators.

Black-box models in medical AI face issues such as unclear decision-making basis, difficulty in detecting biases, unreliable handling of edge cases, insufficient regulatory compliance, and challenges in doctor-patient communication. Liver disease is a major global health burden, and early identification is crucial. However, traditional diagnosis has problems like high cost, strong invasiveness, or insufficient sensitivity. Machine learning can assist in early screening but needs to be based on explainability.

LiverXAI is an end-to-end explainable prediction system with core components including:

• Data layer: Uses datasets containing clinical indicators (demographics, liver function, blood tests, etc.), with preprocessing to ensure data quality;
• Model layer: Adopts traditional machine learning models (logistic regression, random forest, etc.) and ensemble methods, uses cross-validation to prevent overfitting, and focuses on sensitivity in evaluation;
• XAI layer: Integrates SHAP (feature contribution), LIME (single-case explanation), feature importance analysis, and decision visualization;
• Stability testing: Adversarial testing, out-of-distribution detection, cross-validation stability;
• Interactive web application: Patient data entry, real-time prediction, explanation visualization, history management, and doctor feedback mechanism. Technical implementation: Backend uses Python + Flask/Django, scikit-learn/XGBoost, etc.; frontend uses React/Vue.js, D3.js, etc.; models are deployed as REST APIs, supporting version management and A/B testing.

The value of LiverXAI lies in building trust through explainability: it assists junior doctors in learning indicator combinations; provides senior doctors with a second opinion for quality control; helps patients understand the basis of risks; feeds back to medical research through feature analysis; and provides explainability documents required for regulatory approval.

Current limitations: Data dependence (quality and representativeness affect generalization), feature limitations (only blood indicators), and explanation depth (statistical rather than causal). Future directions: Multimodal fusion (blood + imaging + genetics), causal inference, personalized treatment recommendations, federated learning (joint data under privacy protection), and continuous learning (adapting to disease changes).

LiverXAI provides references for medical AI development: explainability first (considering XAI from the design stage), end-to-end delivery (complete web application), robustness verification (boundary condition testing), and user-centric design (aligning with doctors' cognitive habits).

LiverXAI represents the direction of medical AI that values explainability and credibility, demonstrating the application of XAI technology in liver disease prediction and the path to transforming it into a clinical tool. It provides an architectural reference for researchers, shows doctors a transparent auxiliary diagnosis method, and means more understandable services for patients. In the future, as XAI matures and regulations improve, such systems will promote the development of precision medicine.