Augmenting Insufficiently Accruing Oncology Clinical Trials Using Generative Models: Validation Study.

Area of Science:

• Clinical Trials
• Artificial Intelligence
• Oncology Research

Background:

• Insufficient patient accrual is a significant challenge in clinical trials, leading to underpowered studies and increased costs.
• Existing methods using generative models to simulate patients have limitations in scope and evaluation.
• Real-world data can serve as external controls, but accrual issues impact all study arms.

Purpose of the Study:

• To comprehensively evaluate the utility of generative models in simulating additional patients to address insufficient clinical trial accrual.
• To assess the effectiveness of different generative models in augmenting trial data.

Main Methods:

• Retrospective analysis of 10 datasets from 9 completed cancer trials.
• Simulated insufficient accrual by removing 10-50% of patients and used generative models to replace them.
• Evaluated four generative models (sequential synthesis, Bayesian network, GAN, VAE) and bootstrap sampling.
• Replicated published analyses using decision agreement, estimate agreement, standardized difference, and CI overlap metrics.

Main Results:

• Sequential synthesis demonstrated high performance (88-100% decision agreement) for up to 40% patient removal.
• Bootstrap sampling showed moderate effectiveness (78-89% decision agreement).
• No systematic difference found between early and later recruited patients, supporting generative model efficacy.
• High fidelity of generated data to training data was observed (Hellinger distance).

Conclusions:

• Sequential synthesis can simulate a full dataset for oncology trials with as little as 60% target recruitment, offering an alternative to underpowered studies.
• Generative models show potential to rescue poorly accruing clinical trials.
• Further research is needed to confirm findings and generalize to other diseases.