ABSTRACT NUMBER: ESOC2026A1938 EXPLAINABLE MACHINE LEARNING IDENTIFIES PATIENTS' SUBGROUPS BENEFITING FROM INTENSIVE BLOOD PRESSURE REDUCTION AFTER ACUTE INTRACEREBRAL HEMORRHAGEWOULD NOY DEFINE AN ACRONYM IN THE TITLE

Abstract Background and aims Many prior randomized clinical trials of intensive blood pressure (BP) reduction in acute intracerebral hemorrhage (ICH) had neutral results, suggesting that uniform treatment strategies may obscure benefit. We applied explainable machine learning to identify acute ICH patients who benefited from intensive BP reduction in the ATACH-2 trial, which reported neutral overall Results Methods We conducted a post-hoc analysis of ATACH-2 trial, which evaluated intensive BP reduction within 4.5 hours of onset in n=1000 patients with acute supratentorial ICH. An explainable machine learning model was trained using data from the standard BP management (control) arm to predict 3-month poor outcome (modified Rankin Scale score of 4–6). Shapley Additive Explanations (SHAP) values quantified the contribution of admission BP to outcome prediction at an individual-level. Counterfactual BP reduction was simulated to identify patients predicted to benefit from intensive BP lowering. Outcomes among treatment-arm subgroup predicted to benefit were compared with propensity-score-matched controls with similar admission NIHSS, age, BP, and hematoma volume. Results The model identified 130 patients in ATACH-2 treatment arm predicted to benefit from BP reduction. Compared with 94 propensity-score-matched controls, intensive BP lowering was associated with lower odds of poor outcome (odds ratio 0.56, 95% CI 0.33–0.96; p=0.034). Treatment assignment (p=0.009), admission NIHSS score (p<0.001), and age (p<0.001) were independent predictors of outcome. Conclusions Explainable machine learning identified a subgroup of acute ICH patients who benefited from intensive BP reduction despite neutral overall trial results, supporting a precision, risk-based approach to BP management and informing future trial design. Conflict of interest Anh T. Tran, PhD:nothing to disclose; Adnan I. Qureshi, MD:nothing to disclose; Joshua Z. Willey, MD:nothing to disclose; Santosh B. Murthy, MD:nothing to disclose; Guido J. Falcone, MD:nothing to disclose; Lee H. Schwamm, MD:nothing to disclose ; Kevin N. Sheth, MD:nothing to disclose; Seyedmehdi Payabvash, MD:nothing to disclose