Bayesian / Adaptive Trial Design
一行要約
Bayesian adaptive trial design は事前分布 + 蓄積データから posterior を逐次更新し、dose escalation, randomization ratio, sample size, futility/efficacy stopping を data-driven に変更する試験設計手法。Phase I では BOIN / mTPI / CRM が 3+3 を超えて MTD identification の精度・効率を向上、phase II/III では response-adaptive randomization (BATTLE-2, I-SPY 2), master protocol (basket / umbrella / platform) で biomarker-defined subpopulation に高速対応。Oncology drug development の主流設計の一つ。
原理
Bayesian foundation
posterior ∝ likelihood × prior。phase I では DLT 確率を Bayesian model (BLRM, CRM) で更新し、各 dose の posterior probability で escalation 判断。BOIN は non-parametric 簡略版で実装容易、phase I の世界標準化が進行。
主要 design
- BOIN (Bayesian Optimal Interval): dose escalation を simple decision rule で実装、3+3 を上回る MTD identification、Yuan/Liu MD Anderson 開発
- CRM (Continual Reassessment Method): dose-toxicity model を Bayesian に逐次更新
- Response-adaptive randomization: 効果群への randomization ratio を高める (I-SPY 2: biomarker × treatment 各 arm graduate / drop)
- Group sequential: O’Brien-Fleming, Pocock 等で interim 検定
- Sample size re-estimation: blinded / unblinded で interim variance / treatment effect から N 再計算
- Master protocol: basket (single drug, multi-tumor with shared biomarker), umbrella (single tumor, multi-biomarker, multi-drug), platform (multi-drug, perpetual, e.g. STAMPEDE / I-SPY 2)
主要エビデンス / 適用領域
- NSCLC: LIBRETTO-001 (selpercatinib RET basket), TRIDENT-1 (repotrectinib ROS1/NTRK basket), B-FAST / BLOOM (umbrella)
- Breast cancer: I-SPY 2 (response-adaptive, neoadjuvant), graduated drugs include pembrolizumab, durvalumab+olaparib
- Phase I oncology: BOIN が 3+3 を凌駕する efficiency、FDA も recognize
- STAMPEDE (prostate): platform trial で複数 arm を同時試験、control 共有
- COVID RECOVERY (oncology 外): platform trial paradigm の汎用化例
適用分野と限界
- 強み: 効率向上 (smaller N, faster decision), biomarker-driven subpopulation 高速 evaluation, drug attrition rate 改善, ethical (患者を有効 arm に多く)
- 限界: regulatory acceptance は design 依存 (basket は OK、response-adaptive RCT は limited use)、複雑性で IRB / DSMB review burden 増加、interim simulation で type-1 error 厳密制御要、small subgroup での precision 制限、stat 専門人材依存
Open Questions
- AI / RL-based adaptive design: reinforcement learning で adaptive policy 最適化
- External control arm + Bayesian borrowing: RWD を informative prior に
- Decentralized / digital trial との統合
- Biomarker uncertainty を組込んだ design: imperfect biomarker assay の影響を design 段階でモデル化