Hallucination-Aware Clinical Decision Systems: Detecting and Mitigating AI False Outputs in Real-Time Diagnosis

This paper presents a framework for building hallucination-aware Clinical Decision Support Systems (CDSS) that detect and reduce false or unverified AI outputs during real-time diagnosis. It combines three layers of defence. Retrieval-Augmented Generation grounds model outputs in medical guidelines, drug databases and patient Electronic Health Records. Uncertainty Quantification, using Bayesian Neural Networks, MC Dropout and Deep Ensembles, signals when the model is unsure. Granular fact-checking splits each output into atomic claims and verifies them against the patient record. The paper also discusses evaluation benchmarks (Med-HALT, FactScore, Med-NLI), human-in-the-loop review, trust calibration and guided deferral to clinicians. It concludes with future directions in regulation, agentic verification and longitudinal evaluation. Keywords: clinical decision support, AI hallucination, retrieval-augmented generation, uncertainty quantification, fact-checking, human-in-the-loop, patient safety