Uncertainty quantification framework for AI-based survival models: Application to lung cancer prognosis in older...

Reliable prediction is critical for prognosis communication and informed decision-making, yet remains challenging for cancer patients-especially older adults with non-small cell lung cancer (NSCLC)-due to the uncertainty inherent in model-based predictions. This study presents a framework that integrates uncertainty quantification (UQ) into individual survival prediction using electronic health records from 4243 older NSCLC patients in Korea. We applied four Cox proportional hazard-based survival models and four artificial intelligence (AI)-based survival models, including DeepSurv, to predict 2-year survival probabilities. We introduce two novel UQ metrics: certainty score, capturing the relative model confidence in predicted mortality risk, and predictive multiplicity, quantifying model disagreement in risk stratification. Although the survival models achieved high mean areas under the receiver operating characteristics curve ranging from 0.840 to 0.851, 26% of patients in the test set were assigned to conflicting risk groups depending on the model used, indicating considerable variability in model-predicted prognosis. DeepSurv demonstrated the highest average certainty. All models showed substantial degrees of predictive ambiguity and discrepancy. We also developed a visual informatics tool that presents personalized best-, worst-, and most likely-case scenarios, risk group stratification, and interpretable feature importance to improve transparency and facilitate shared decision-making. This framework offers a practical approach for integrating uncertainty into AI-based prognosis, addressing the challenge of enhancing confidence in cancer prognosis communication by quantifying and visualizing model uncertainty. It can support clinicians in tailoring prognostic discussions based on the level of model consensus and confidence, helping guide when to communicate prognosis cautiously or emphasize shared decision-making. The proposed framework is model-agnostic and readily applicable to real-world clinical settings.