Evaluation of AI-Derived Survival Estimates Against Real-World Outcomes in Extensive-Stage Small Cell Lung Cancer
TL;DR
AI-derived survival estimates based on baseline data tend to overestimate survival and show limited agreement with real-world outcomes in extensive-stage SCLC, suggesting these tools may provide general prognostic insight but are not yet suitable for independent clinical decision-making.
Abstract
Objectives: Artificial intelligence (AI) is increasingly explored as a prognostic tool in oncology; however, its reliability in real-world settings remains uncertain. We aimed to compare AI-derived survival estimates with observed outcomes in patients with extensive-stage small cell lung cancer (SCLC). Methods: This retrospective cohort study included 146 patients treated between 2016 and 2025. Baseline demographic, clinical, metastatic, treatment, and laboratory data were entered into a large language model–based system to generate predicted overall survival (OS) and progression-free survival (PFS). Predicted values were compared with observed outcomes. Agreement was assessed using Bland–Altman analysis. Results: The median observed OS was 10.97 months, whereas the AI-predicted median OS was 13.2 months, indicating an overestimation of 2.2 months. For PFS, the observed and predicted medians were 6.53 and 7.2 months, respectively, corresponding to an overestimation of 0.7 months. Bland–Altman analyses demonstrated systematic positive bias and wide limits of agreement, indicating substantial variability at the individual level, more pronounced for OS. Conclusion: AI-derived survival estimates based on baseline data tend to overestimate survival and show limited agreement with real-world outcomes in extensive-stage SCLC. These tools may provide general prognostic insight but are not yet suitable for independent clinical decision-making.