Identifying treatment-responsive patient subgroups in a neutral clinical trial of Intensive blood pressure reduction in acute intracerebral hemorrhage: A post hoc explainable machine learning analysis.

The Antihypertensive Treatment of Acute Cerebral Hemorrhage (ATACH-2) trial reported no overall benefit from intensive blood pressure (BP) reduction in intracerebral hemorrhage (ICH), potentially masking benefit in select patient subgroups. We evaluated whether explainable machine learning could identify treatment-arm subgroups in whom BP reduction is associated with improved outcomes. Using the ATACH-2 dataset, we trained an XGBoost model on two-thirds of the control arm (n = 326) to predict 3-month poor outcome (modified Rankin Scale >3). The model was then applied to treatment-arm patients (n = 499) and held-out controls (n = 163). SHapley Additive exPlanations (SHAP) quantified individual BP risk contributions, and counterfactual perturbation simulated BP reductions. We applied an exploratory, hypothesis-generating grid search to identify selection strategies yielding the lowest odds ratio (OR) for poor outcome in treatment subgroups versus controls. The model achieved an area under the curve of 0.85 (95% confidence interval [CI], 0.81-0.89) in cross-validation and 0.83 (95% CI, 0.77-0.89) in independent validation. Using a combined SHAP and counterfactual analysis, we identified a candidate treatment-responsive subgroup (n = 56) in whom intensive BP reduction was independently associated with lower odds of poor outcome compared with held-out controls (OR = 0.21, 95% CI, 0.08-0.54; p = 0.002). This association remained significant when compared with the subset of held-out controls meeting the same selection criteria. Compared with the remainder of their respective groups, these patients had higher baseline systolic BP, more severe neurological deficits, lower blood glucose levels, and more frequent basal ganglia involvement. These findings highlight the potential of explainable machine learning for precision BP management in acute ICH.