Predictors of Six-Month Functional Outcome After Chronic Subdural Hematoma Surgery: Logistic Regression Versus Machine Learning

Background/Objectives: Six-month functional outcome after chronic subdural hematoma (CSDH) surgery has not been reliably predicted in Southeastern European cohorts. We asked whether machine learning (ML) classifiers outperform logistic regression. Methods: This study included a single-center retrospective cohort of 78 surgically treated CSDH patients (Novi Sad, Serbia). The primary outcome was favorable six-month outcome (Glasgow Outcome Scale 4–5). Missing covariates were multiply imputed (multiple imputation by chained equations, MICE, m = 20), and estimates were pooled using Rubin’s rules. A three-predictor model was pre-specified; six ML classifiers were trained on five preoperative predictors and evaluated using stratified 10-fold repeated cross-validation, with metrics derived from held-out predictions. Results: Favorable outcome occurred in 50/78 (64.1%). Preoperative Karnofsky Performance Status (KPS) was the only independent predictor (odds ratio, OR = 1.14 per point, 95% confidence interval, CI 1.07–1.20, p < 0.001). Cortical atrophy grade was associated univariately (OR = 0.57, 95% CI 0.34–0.96, p = 0.033) but not after adjustment for KPS (OR = 1.08, 95% CI 0.53–2.22, p = 0.833). Discrimination was indistinguishable across the six classifiers (AUC 0.870–0.885; all pairwise DeLong p ≥ 0.123); the three-predictor model reached an AUC of 0.903 (0.833–0.972). Conclusions: Preoperative KPS was the only independent predictor. No ML classifier outperformed logistic regression; a three-predictor model performed at least as well. External validation is required.