Comparison of Machine Learning Methods to Predict Early Mortality After Evacuation of Chronic Subdural Hematoma

BACKGROUND AND OBJECTIVES:We developed a series of machine learning models to predict early mortality after chronic subdural hematoma (cSDH) evacuation.METHODS:We retrospectively collected patients treated surgically for cSDH at 4 level 1 trauma centers (2009-2021). Previously, we developed a deep learning segmentation tool to automatically calculate preoperative and postoperative cSDH volumes. Using cSDH volumes and clinical information, we developed 6 machine learning models including logistic regression (LR), support vector machine, neural network (NN), decision tree (DT), Nave Bayes, and XGBoost to predict 30-day mortality after surgery. We applied least absolute shrinkage and selection operator regression to select a subset of predictors for consistent model input. To account for class imbalance, we used synthetic minority oversampling technique. We used 10-fold cross validation to evaluate model performance.RESULTS:We included 731 patients. Our final models included age, admission Glasgow Coma Scale, unilateral/bilateral hematoma, antiplatelet status, platelet count, preoperative volume, and method of surgical evacuation. The 30-day mortality rate was 7.5%. Overall, our models demonstrated moderate discriminative ability with area under the receiver operating characteristics curves (AUCs) ranging from 0.64 for DT (95% CI: 0.56-0.72) to 0.75 for LR (95% CI: 0.69-0.81). AUC for DT was significantly lower than LR (P <.03). AUCs for support vector machine (AUC = 0.73; 95% CI: 0.67-0.79), NN (0.69; 95% CI: 0.62-0.76), Nave Bayes (0.70; 95% CI: 0.63-0.78), and XGBoost (0.73; 95% CI: 0.66-0.80) were not significantly different from LR. LR achieved the highest balanced accuracy (0.69) whereas DT and NN had the lowest (0.61). Age, craniotomy, Glasgow Coma Scale, larger preoperative volumes, unilateral cSDH, and lower platelet count were associated with increased risk of mortality on multivariate analysis.CONCLUSION:The LR model demonstrated the best performance of discriminative ability, balanced accuracy, and recall, whereas DT modeling performed worst. Using an automated segmentation software, our models demonstrate an ability to identify patients at high risk of mortality after treatment for cSDH.