Professor Yeonhee Park of the Department of Statistics at Sungkyunkwan University has developed a novel statistical framework — MARGO (Machine Learning-Assisted Adaptive Randomization for Group Sequential Trials Based on Overlap Weights) — that makes machine learning practically applicable in clinical trial design. This work provides the first rigorous solution to the fundamental statistical challenges that arise when integrating ML/AI-driven decision-making into the scientifically demanding environment of clinical trials.
The Promise and the Barrier: Why ML/AI Alone Is Not Enough
Machine learning and artificial intelligence have garnered widespread attention as transformative tools for personalized treatment assignment in clinical trials. In particular, adaptive randomization — which dynamically adjusts treatment allocation based on accumulating trial data — is a promising approach for improving patient outcomes by steering more participants toward more effective treatments. However, applying this approach in practice can introduce a critical statistical problem. When patient characteristics (e.g., biomarkers) are used to guide treatment assignment, systematic imbalances can emerge between treatment groups. This covariate imbalance leads to biased treatment effect estimates and an inflated type I error rate, risking false conclusions. The problem is further compounded in group sequential designs, which include planned interim analyses for early stopping decisions.
Machine Learning Meets Causal Inference: A Two-in-One Solution
To address this fundamental challenge, MARGO integrates machine learning-based predictive models with overlap weights (OW), a propensity score–based approach widely used in causal inference to adjust for covariate imbalance. MARGO uses patient covariate information to predict the probability of treatment success via machine learning, then uses these predictions to preferentially assign patients to the more effective treatment. Simultaneously, OW corrects covariate imbalance across treatment groups, effectively controlling the bias and type I error inflation induced by adaptive randomization. The framework was evaluated using four machine learning algorithms: Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Random Forest (RF), and Multi-Layer Perceptron (MLP).
Rigorously Validated Performance
Through extensive simulation studies, MARGO demonstrated superior performance over conventional fixed randomization and existing adaptive randomization methods across three key dimensions. First, MARGO allocated a greater proportion of patients to the more effective treatment. Second, it maintained the overall type I error rate below the target threshold of 0.05 — even in scenarios where conventional methods inflated the error rate to as high as 0.08–0.18. Third, it preserved high statistical power under alternative scenarios while reducing the number of treatment failures. Together, these results demonstrate that MARGO can simultaneously improve the ethical standards and scientific integrity of clinical trials.
Beyond “Using AI” Toward “Trusting AI in Clinical Trials”
The most important contribution of this research goes beyond simply applying machine learning to clinical trials — it rigorously resolves the fundamental statistical problems that emerge in that process. MARGO is designed to accommodate a wide range of AI models and holds broad potential for extension to precision medicine and data-driven decision-making across diverse fields.
This study was published in Statistics in Medicine.
