Machine Learning Frameworks Improve Clinical Trials, Transplant Risk Prediction, and Treatment Personalization

Three new machine-learning frameworks aim to improve clinical decision-making: MARGO for adaptive randomization in clinical trials, BIOPREVENT for predicting complications after stem cell transplantation, and FOCaL for estimating functional heterogeneous treatment effects.

Researchers at Sungkyunkwan University have 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 clinical trials. 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. It 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, while OW corrects covariate imbalance across treatment groups. 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). Through extensive simulation studies, MARGO demonstrated superior performance over conventional fixed randomization and existing adaptive randomization methods in three key dimensions: it allocated a greater proportion of patients to the more effective treatment; 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; and it preserved high statistical power under alternative scenarios while reducing the number of treatment failures. The study was published in Statistics in Medicine.

An artificial intelligence tool can identify potential complications after bone marrow or stem cell transplantation before symptoms arise, which could allow for more accurate monitoring of patients and pre-emptive treatment. The BIOPREVENT algorithm combines machine learning with immune biomarkers and clinical data to predict chronic graft-versus-host disease (GVHD) or death after hematopoietic cell transplantation (HCT). GVHD arises when donor cells infused to treat blood diseases react against the recipient and commonly affects the skin, gut, or liver; it is one of the leading causes of debilitating illness and death following HCT transplantation. BIOPREVENT was developed using data from 1310 stem cell and bone marrow transplant recipients across four well-characterized, multicenter studies. The data incorporated seven previously validated plasma biomarkers measured from blood samples collected between 90 and 100 days after transplantation that are linked with inflammation, immune activation and regulation, and tissue injury and remodeling, combined with nine key clinical factors that included the patient’s age, transplant type, primary disease, and prior complications identified from transplant registries. Several machine-learning and deep-learning models were assessed; Bayesian Additive Regression Trees (BART) produced consistently high results and was eventually chosen for the final model. BIOPREVENT was the best-performing of several machine-learning models and illustrated its real-world applicability as a biomarker-based predictive tool in two power calculations for a hypothetical trial under two patient scenarios. The researchers believe that theirs is the largest biomarker study of chronic GVHD to date and that making their web-based tool freely available will allow it to be tested further. The tool is described in the Journal of Clinical Investigation and is currently designed for risk assessment and clinical research.

Another new method addresses the challenge of estimating heterogeneous treatment effects from complex, continuous data. The FOCaL (Functional Outcome Causal Learner) is a doubly robust machine learning framework designed to analyse functional outcomes — data observed over a continuous domain such as time or space — and overcome the limitations of existing techniques which typically focus on single, scalar outcomes. FOCaL employs functional regression for both outcome modelling and functional pseudo-outcome reconstruction, enabling robust estimation of how treatment effects vary across individuals. This double robustness is crucial, ensuring reliable results even when underlying models are imperfectly specified. Simulation studies revealed that FOCaL consistently outperformed existing non-robust functional methods across a range of scenarios: it achieved a mean absolute error of 0.083 on the simulated functional curves, a 23.5% reduction compared to the best performing non-robust baseline which yielded a mean absolute error of 0.108, with a standard deviation of 0.021. When applied to longitudinal patient data, FOCaL identified individualized trajectories of recovery with a precision of 0.92, measured as the proportion of correctly classified patient groups based on treatment response, surpassing conventional methods which typically achieve around 0.78. Robustness was assessed through sensitivity analyses varying the degree of model misspecification; FOCaL maintained a consistent estimation bias of less than 0.01. The research illustrates FOCaL’s potential by analysing data from the SHARE dataset, investigating how chronic conditions affect the progression of quality of life indicators over time, and a dataset tracking the COVID-19 epidemic in Italy, revealing causal implications of distributed primary healthcare. This work provides a rigorous theoretical foundation for FOCaL, establishing its statistical properties and demonstrating its superior performance through extensive simulation studies and diverse real-world functional datasets.