Machine learning integration of routine inflammatory biomarkers for predicting remote punctate ischemic lesions following intracerebral hemorrhage: a single-center retrospective study.

Background: Remote Punctate Ischemic Lesions (RPIL) occurring during the acute phase of intracerebral hemorrhage (ICH) represent a paradox of "ischemia within hemorrhage" that is associated with worse functional outcome. While the concept of "thrombo-inflammation" is gaining traction, the predictive utility of novel biomarkers like the Systemic Immune-Inflammation Index (SII) combined with advanced machine learning (ML) remains uncharacterized. Methods: In this retrospective cohort study conducted between January 2019 and December 2025, 12,327 patients with ICH were initially screened, and 6,134 were included after strict exclusion criteria. The cohort was randomly split into training ( Results: LASSO regression identified six key predictors: Age, History of Diabetes, SII, D-Dimer, Glucose, and Fibrinogen. In the ML benchmark, XGBoost achieved the highest discrimination (AUC = 0.799), outperforming the Neural Network (AUC = 0.798) and standard Logistic Regression (AUC = 0.770). The derived nomogram demonstrated excellent calibration (Mean Absolute Error = 0.034) and clinical net benefit in Decision Curve Analysis (DCA). Crucially, inflammatory and coagulation markers (SII, Fibrinogen) were identified as top-tier predictors, corroborating the immuno-thrombotic mechanism. Conclusion: We present a robust ML framework demonstrating that systemic inflammation and hypercoagulability are strongly associated with post-ICH ischemia. The XGBoost model offers precision, while the nomogram provides translational utility for bedside risk stratification. However, as this study relies on single-center data, future multicenter external validation is imperative to confirm the generalizability and clinical applicability of these models before broad implementation.