Prediction of postoperative intracerebral hemorrhage following brain biopsy using integrated statistical and machine learning analyses

Postoperative intracerebral hemorrhage (P-ICH) is the most common complication of brain biopsy, yet predictors of clinically significant symptomatic events remain poorly defined. We performed a retrospective dual-center cohort study of 464 consecutive adult patients undergoing brain biopsy between 2014 and 2024. P-ICH was classified by clinical severity (none, asymptomatic, symptomatic) and timing (immediate or delayed). Multivariable ordinal logistic regression and conditional inference tree (CIT) analysis were used for interpretable risk stratification, and machine learning (ML) models based on perioperative variables were developed to predict symptomatic P-ICH using nested cross-validation. Immediate P-ICH occurred in 54.7% of cases, including 49.4% asymptomatic and 5.4% symptomatic hemorrhage. Increasing age (adjusted odds ratio 1.015 per year, p = 0.015) and basal ganglia (BG) location (adjusted odds ratio 8.305, p < 0.001) were independently associated with greater hemorrhage severity. Lesion location was the primary determinant in CIT analysis, with age providing secondary stratification among non-BG lesions. ML models demonstrated moderate discrimination, with Random Forest achieving an area under the curve (AUC) of 0.711, and consistently identified lesion location as the most important predictor. These findings support risk-adapted trajectory planning, patient counseling, and postoperative monitoring in high-risk populations.