Adjuvant nivolumab vs pembrolizumab in stage IIB/IIC melanoma: a reinforcement learning-based simulation study

Adjuvant programmed cell death protein 1 (PD-1) inhibitors (nivolumab, pembrolizumab) improve recurrence-free survival (RFS) in stage IIB-IIC melanoma, yet no head-to-head trial directly compares them. Traditional indirect methods estimate relative efficacy but often fail to integrate toxicity and patient-level trade-offs. Reinforcement learning (RL) provides a framework to simulate decision-making under uncertainty and competing clinical priorities.

We developed an RL model treating each simulated patient as the environment, with state variables including age, ECOG status, stage, time-to-recurrence, and adverse event (AE) outcomes. Actions were treatment choices between nivolumab and pembrolizumab. Rewards combined gains in RFS (+ 1 per 2 months) with penalties for grade 3-4 AEs and discontinuations, incorporating both raw and placebo-adjusted AE rates. Q-learning was iterated across 1000 virtual trial episodes until policy convergence.

The RL-derived policies reflected conditional treatment preferences rather than a single optimal agent. In scenarios weighted toward tolerability, nivolumab was favored due to lower grade 3-4 AE and discontinuation rates. When incremental RFS gains were prioritized, pembrolizumab emerged as the preferred option. Placebo-adjusted versus raw AE modeling materially influenced the balance of preferences, underscoring the importance of attribution in comparative safety assessment.

Our RL framework complements existing comparative methods by making treatment trade-offs explicit and scenario-dependent. Rather than declaring a universal "best" PD-1 inhibitor, the model contextualizes efficacy-toxicity balances, supporting transparent decision-making in settings where small absolute differences may meaningfully influence patient and clinician preferences.