ATSDH is a critical neurosurgical emergency: while rare, ATSDH can lead to severe neurological complications and substantial morbidity and mortality rates.

This study attempts to identify modifiable risk factors, such as optimal surgical techniques and post-operative care protocols, to better characterize ATSDH sequelae and thereby improve survival and quality of life for ATSDH patients.

Data for 183 patients presenting with ATSDH were collected, including patient demographics, intervention, and pre- and post- operative evaluation along with individual outcomes. 

Patient information was analyzed through XGBoost to predict mortality. After training, the weights of the model were evaluated to determine which features were most significant in making its prediction. 

5-fold cross validation for mortality had an accuracy of 90.13% (F1 score: 90.145, MSE: 0.099). The highest weights were basal cistern compression/effacement, use of mannitol/hypertonic saline, and Rotterdam score. Lowest weights included whether a lesion was present in epidural, subdural, or intraparenchymal space.

When predicting for TBI outcomes ranging from "good recovery" to "death," accuracy for these six levels was 43.84% (MSE: 1.69). Heaviest weights included whether the patient was on blood thinner(s) prior to arrival while lowest weights included the total craniotomies performed.

The machine learning technique employed in this study is an ensemble of weak learners (simple decision trees). Despite such an architecture and with minimum adjustment of default hyperparameters, the model was able to achieve high accuracy with low MSE. While the accuracy for classifying the TBI outcome was lower, it is significantly higher than random. Evaluating the weights revealed specific factors that would be important to focus on for developing targeted treatment strategies and improving neurological outcomes in patients with traumatic SDH.