An Explainable AI Diagnostic Framework with SHAP-Driven Decision Interpretability

Healthcare is being revolutionized by AI with faster, data supported decision support and timely diagnosis. In this paper, we introduce Doctor AI, a symptom-based disease prediction system which predicts likely diseases based on a limited set of symptoms provided by the user. The proposed approach is based on ensemble learning for enhancing various types of performance. We perform experiments over a curated dataset with 41 diseases and 4,290 samples and observe that Doctor AI achieves full test accuracy in the considered setting. In order to improve clinical confidence and transparency, authors employed SHAP to measure feature-wise contribution and explain away (every) prediction. To further support the healthcare professionals, an explainability module is provided to bring forward highly influential symptoms beneath model outputs that accompany validation. In conclusion, Doctor AI has shown that the synergy between high performing ensembles with a model agnostic explainability can boost user confidence levels and lead to successful practical deployments of AI assisted digital health care systems.