CertiHealth: Towards Certified, Uncertainty-Aware, and Explainable AI for Medical Decision-Making

AAAI 2026 Workshop TrustAgent Submission55 Authors

Keywords: Certified robustness, Uncertainty awareness, Explainable AI, Medical decision-making, Trustworthy AI, Distributional robustness

TL;DR: CertiHealth builds safe, explainable, and certifiable medical AI that stays reliable under uncertainty and data shifts.

Abstract: Artificial intelligence has shown promise in automating clinical diagnosis and decision support, yet most medical AI systems remain unreliable, opaque, and unverified [1, 4, 5]. Existing approaches typically address either model interpretability [6, 7], uncertainty quantification [3], or robustness certification [1, 2] in isolation, leaving critical gaps in safety and trust [11, 12]. This paper introduces CertiHealth, a unified framework for developing certified, uncertainty-aware, and explainable AI models for medical decision-making. CertiHealth constrains neural architectures through Lipschitz continuity [1], enabling formal robustness guarantees against bounded input perturbations. It integrates probabilistic uncertainty estimation [3] to assess predictive confidence and employs interpretable attribution mechanisms [6, 7] to provide transparent, clinically meaningful explanations. Evaluated on diagnostic tasks using the MIMIC-IV clinical dataset, CertiHealth demonstrates improved calibration, verifiable robustness, and alignment between model explanations and known medical risk factors. By combining mathematical certification, quantified uncertainty, and human-centered interpretability [8, 9, 13], CertiHealth advances the development of verifiably trustworthy medical AI suitable for safety-critical healthcare environments [10, 12, 14].

Submission Number: 55