Risk-inclusive Contextual Bandits for Early Phase Clinical Trials

Early-phase clinical trials face the challenge of selecting optimal drug doses that balance safety and efficacy due to uncertain dose-response relationships and varied participant characteristics. Traditional randomized dose allocation often exposes participants to sub-optimal doses by not considering individual covariates, necessitating larger sample sizes and prolonging drug development. This paper introduces a risk-inclusive contextual bandit algorithm that utilizes multi-arm bandit (MAB) strategies to optimize dosing through participant-specific data integration. By combining two separate Thompson samplers, one for efficacy and one for safety, the algorithm enhances the balance between efficacy and safety in dose allocation. The effect sizes are estimated with a generalized version of asymptotic confidence sequences (AsympCS), offering a uniform coverage guarantee for sequential causal inference over time. The validity of AsympCS is also established in the MAB setup with a possibly mis-specified model. The empirical results demonstrate the strengths of this method in optimizing dose allocation compared to randomized allocations and traditional contextual bandits focused solely on efficacy. Moreover, an application on real data generated from a recent Phase IIb study aligns with actual findings.

💡 Research Summary

This paper tackles a central challenge in early‑phase dose‑ranging clinical trials: how to allocate drug doses so that both efficacy and safety are optimized while accounting for patient‑specific covariates. Traditional designs such as equal randomization (ER) ignore covariates and treat safety and efficacy separately, often leading to sub‑optimal exposure, larger sample sizes, and longer development timelines. The authors propose a novel “Risk‑inclusive Thompson Sampling” (RiTS) framework that casts the trial as a contextual multi‑armed bandit (CMAB) problem and integrates two parallel Thompson samplers—one for efficacy and one for safety—combined through a user‑specified weight parameter w.

The methodological pipeline consists of two stages. In the trial‑optimization stage, a parametric working model assumes linear relationships between covariates X and the potential outcomes: Rₙ(a)=βₐ,0+Xₙᵀβₐ+εₙ for efficacy and Sₙ(a)=γₐ,0+Xₙᵀγₐ+δₙ for safety, with normal errors. Bayesian priors on the regression coefficients are updated after each patient, yielding posterior distributions. For each incoming participant, M draws are taken from the efficacy and safety posteriors, producing predicted efficacy (b₀+Xᵀb) and safety (g₀+Xᵀg) scores. The composite score ω = w·(efficacy prediction) + (1‑w)·(safety prediction) is computed for each arm; the arm that maximizes ω in a given draw receives a vote, and the proportion of votes across M draws defines the allocation probability qₙ(a,Xₙ). A clipping rule enforces a minimum probability m for every arm, guaranteeing exploration of all dose levels. This design allows clinicians to encode prior safety thresholds or therapeutic goals directly into w, making the algorithm adaptable across therapeutic areas (e.g., oncology vs. dermatology).

The second stage addresses statistical inference after the trial. The authors extend Asymptotic Confidence Sequences (AsympCS) to the CMAB setting, providing time‑uniform confidence intervals for treatment effect sizes Δ(a)=E