Research map / Probabilistic and scientific modeling / Molecules and proteins
Safe Exploration in Dose Finding Clinical Trials with Heterogeneous Participants
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SAFE-T, a safe adaptive dose-finding procedure using multi-output GP classification and Bayesian optimization over Thall utility, evaluated on 18 synthetic scenarios against 3+3, CRM and C3T.
Problem
Early-phase dose-finding trials need to allocate doses that are safe, benefit heterogeneous participant subgroups, handle non-monotone dose-response, and work with small sample sizes; standard methods (3+3, CRM) ignore efficacy and heterogeneity and rely on parametric assumptions.
Approach
Models toxicity and efficacy with nonparametric multi-output GP classifiers (LMC), enforces a high-probability safety constraint via GP upper confidence bounds on toxicity, explores doses safely upward, then optimizes efficacy with EI under the safe set, and recommends final doses per subgroup by maximizing Thall utility.
Key results
- SAFE-T has lowest or comparable dose error in 16/18 synthetic scenarios and fewest or comparable safety violations in 16/18 scenarios, averaged over 100 trials.
- SAFE-T achieves highest or comparable participant utility in 17/18 scenarios.
- SAFE-T's final dose error decreases steadily with sample size from N=51 to N=537, whereas 3+3 and C3T do not improve.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SAFE-T instead of 3+3, CRM, or C3T | Dose-finding trials across 18 synthetic scenarios · 18 synthetic scenarios | final recommended dose error | Helped |
| Use SAFE-T instead of 3+3, CRM, or C3T | Dose-finding trials across 18 synthetic scenarios · 18 synthetic scenarios | safety constraint violations | Helped |
| Use SAFE-T instead of 3+3, CRM, or C3T | Dose-finding trials across 18 synthetic scenarios · 18 synthetic scenarios | participant utility | Helped |
| Use SAFE-T rather than 3+3 or CRM | Dose-finding trials with varying subgroup ratios · Scenario 11, varying subgroup ratios | safety constraint violations | Helped |
| Use SAFE-T rather than 3+3 or C3T | Dose-finding trials as sample size increases · Scenario 9, sample sizes N = 51 to N = 537 | final dose recommendation error | Helped |
| Use continuous dose allocation | SAFE-T · Continuous scenarios C.1–C.4 | final dose error | Mixed |
| Use UCB or utility optimization instead of EI | SAFE-T · Synthetic scenarios | performance | No clear effect |
| Remove multi-output GP models (ablation) | SAFE-T · Synthetic scenarios | overall performance | Hurt |
| Remove the expansion stage (ablation) | SAFE-T · Synthetic scenarios | overall performance | Hurt |
| Remove the safety constraint (ablation) | SAFE-T · Synthetic scenarios | overall performance | Hurt |
| Replace final Thall-utility dose recommendation (ablation) | SAFE-T · Synthetic scenarios | overall performance | Hurt |
Surprising findings
- 3+3 and C3T do not improve final dose error with increasing sample size, attributed to misspecification and selecting highest safe or most efficacious dose without utility tradeoff.
Limitations
- Restricted to binary toxicity and efficacy outcomes; ignores delayed, continuous, or missing outcomes.
- Requires heterogeneity to be known and pre-specified; cannot handle latent subgroups.
- GP classification lacks theoretical regret/confidence guarantees; GPs are more complex than standard methods for practitioner adoption.
Experimental setup
Benchmarks: 18 synthetic dose-finding scenarios 4 continuous dose scenarios
Metrics: final recommended dose error safety constraint violations participant utility (Thall) proportion of safety violations across subgroup ratios
Baselines: 3+3 CRM C3T
Builds on: Thall utility linear model of co-regionalization STAGEOPT continual reassessment method C3T
Closest papers by meaning
- Data-driven Design of Randomized Control Trials with Guaranteed Treatment Effects (ICML 2025 · similarity 0.85)
- Dynamic Subgroup Identification in Covariate-adjusted Response-adaptive Randomization Experiments (NeurIPS 2024 · similarity 0.84)
- Using Surrogates in Covariate-adjusted Response-adaptive Randomization Experiments with Delayed Outcomes (NeurIPS 2024 · similarity 0.84)
- Contrastive Balancing Representation Learning for Heterogeneous Dose-Response Curves Estimation (AAAI 2024 · similarity 0.83)
- Efficient Adaptive Experimentation with Noncompliance (NeurIPS 2025 · similarity 0.83)
- Enhancing Statistical Validity and Power in Hybrid Controlled Trials: A Randomization Inference Approach with Conformal Selective Borrowing (ICML 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.