Explainable machine learning identifies subgroup that may benefit from intensive blood pressure reduction in intracerebral hemorrhage

Post hoc analysis of the ATACH-2 trial used explainable ML to find a subgroup of 56 ICH patients in whom intensive BP lowering reduced poor outcomes (OR 0.21, 95% CI 0.08–0.54). Model AUC was 0.83 in independent validation.

• An XGBoost model trained on control-arm data predicted 3-month poor outcome (mRS >3) with an independent validation AUC of 0.83 (95% CI 0.77–0.89), using SHAP and counterfactual perturbation to simulate BP reduction.
• The candidate treatment-responsive subgroup had higher baseline systolic BP, more severe neurological deficits, lower blood glucose, and more frequent basal ganglia involvement than the remainder of their respective groups.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.