Survival–Toxicity Trade-Offs Between Trifluridine/Tipiracil and Regorafenib in Refractory Metastatic Colorectal Cancer: A Simulation-Based Reinforcement Learning Study
Perihan Perkin11Department of Medical Oncology, Ankara Bilkent City Hospital, Ankara, Türkiye
Objectives: We examined whether survival and grade 3–4 toxicity could be jointly modeled for trifluridine/tipiracil (FTD-TPI) and regorafenib in refractory metastatic colorectal cancer using reinforcement learning (RL). Methods: A synthetic cohort of 1,039 virtual patients was generated from later-line trials and allocated to FTD-TPI (n=534) or regorafenib (n=505). Survival was simulated with Weibull distributions incorporating age, ECOG performance status, and KRAS status. Grade 3–4 adverse events were modeled using published probabilities. A one-step Q-learning model integrated survival and toxicity through a prespecified reward function. Cox models descriptively compared survival distributions. Results: Median simulated survival was 6.91 months with FTD-TPI and 6.59 months with regorafenib. The unadjusted hazard ratio was 0.893 (95% CI, 0.790–1.008; p=0.068), and the adjusted hazard ratio was 0.905 (95% CI, 0.801–1.023; p=0.110). Mean modeled rewards were ?18.38 and ?25.24, respectively. The learned policy mapped 68.5% of profiles to FTD-TPI. Conclusion: FTD-TPI showed a more favorable modeled survival–toxicity profile, but the survival difference was not statistically significant and clinical superiority was not established. The framework is hypothesis-generating rather than a validated decision-support tool. Keywords: Colorectal neoplasms, computer simulation, regorafenib, reinforcement machine learning, trifluridine tipiracil drug combination
Cite This Article
Perkin P. Survival–Toxicity Trade-Offs Between Trifluridine/Tipiracil and Regorafenib in Refractory Metastatic Colorectal Cancer: A Simulation-Based Reinforcement Learning Study. EJMI. 2026; 10(4): 601-609
Corresponding Author: