A causal machine learning framework for treatment personalization in clinical trials: application to ulcerative colitis

Abstract

Randomized controlled trials estimate average treatment effects, but treatment-response heterogeneity motivates personalized approaches that identify which patients benefit most from specific interventions. Causal machine learning methods offer a principled framework for estimating heterogeneous treatment effects and evaluating individualized treatment policies, yet their application to clinical trial data requires careful attention to the distinction between statistically detectable heterogeneity and improvements in decision performance. We present a causal machine learning framework that combines X-learner estimation, permutation-based feature importance, best linear predictor testing, and doubly robust policy-value evaluation with out-of-fold validation, applicable to randomized controlled trials across therapeutic areas. As a case study, we demonstrate its application using patient-level data from the UNIFI maintenance trial of ustekinumab in ulcerative colitis. We analyzed the randomized maintenance cohort comparing placebo (n=144), standard-dose ustekinumab every 12 weeks (n=143), and dose-intensified ustekinumab every 8 weeks (n=274). Using cross-fitted gradient-boosting models, we estimated conditional average treatment effects from baseline demographics, medication history, week-8 clinical scores, laboratory biomarkers, and video-derived endoscopic features, and assessed whether endoscopic features add value beyond routinely available clinical markers; confidence intervals were obtained by bootstrap resampling. Clinical and demographic features showed higher aggregate importance than endoscopic features for predicting treatment-effect heterogeneity (non-endoscopic 0.74 ± 0.02 versus endoscopic 0.56 ± 0.01 for ustekinumab versus placebo; ratio 1.33 \(\times \) ). Best linear predictor testing indicated an endoscopic contribution for ustekinumab versus placebo (z=7.89, p<0.001) but not for dose intensification (z=1.18, p=0.50). Doubly robust policy evaluation showed no meaningful improvement in expected remission from incorporating endoscopic features (ustekinumab versus placebo: \(\Delta \) = +4.2 percentage points, 95% CI \(-0.2\) to +8.6). Out-of-fold three-arm policy evaluation showed lower treatment-assignment performance when endoscopic features were included (30.9% versus 37.7% remission). Statistically detectable heterogeneity did not translate into improved treatment assignment, and week-8 endoscopic features added no decision value beyond routinely available clinical markers in this cohort. This framework provides a template for rigorous causal analysis of treatment personalization in randomized controlled trials, distinguishing statistical heterogeneity detection from practical improvements in individualized treatment assignment.