Development of machine learning models for predicting unfavorable functional outcomes from preoperative data in patients with chronic subdural hematomas

Abstract

Chronic subdural hematoma (CSDH) often causes neurological deterioration and is treated with hematoma evacuation. This study aimed to assess the feasibility of various machine learning models to preoperatively predict the functional outcome of patients with CSDH. Data were retrospectively collected from patients who underwent CSDH surgery at two institutions: one for internal validation and the other for external validation. The poor functional outcome was defined as a modified Rankin scale score of 3–6 upon hospital discharge. The unfavorable outcome was predicted using four machine learning algorithms on an internal held-out cohort (n = 188): logistic regression, support vector machine (SVM), random forest, and light gradient boosting machine. The prediction performance of these models was also validated in an external cohort (n = 99). The area under the curve of the receiver operating characteristic curve (ROC-AUC) of each machine learning-based model was found to be high in both validations (internal: 0.906–0.925, external: 0.833–0.860). In external validation, the SVM model demonstrated the highest ROC-AUC of 0.860 and accuracy of 0.919. This study revealed the potential of machine learning algorithms in predicting unfavorable outcomes at discharge among patients with CSDH undergoing burr hole surgery.