Artificial neural networks in traumatic brain injury: predicting outcomes after surgical removal of acute subdural hematoma

Abstract

Background and Aim: Prognostication of acute traumatic brain injury (TBI) remains challenging. We evaluated prognostic ability of artificial neural network (ANN) for predicting outcomes after surgical removal of acute subdural hematoma. Methods: Consecutive TBI patients who underwent surgical evacuation of acute subdural hematoma in a period from January 2009 until January 2016 were prospectively evaluated for age, gender, admission Glasgow Coma Scale (GCS) score, hematoma thickness, midline shift and surgical management. Discharge outcomes were assessed using the Glasgow Outcome Scale (GOS). ANN and Radial basis function neural network (RBFNN) were designed to predict in-hospital mortality rate and unfavorable outcome at hospital discharge. Training set consisted of 70% of the data, and the remaining 30% set was used to test the models. Results: Six-hundred and ninety-five patients (73.4% male; median age 59 years) were studied. Decompressive craniectomy was performed in 266 (38.3%) patients. In-hospital mortality rate was 32.1%, and poor discharge outcome (death, persistent vegetative state or severe disability) rate was 60.4%. For mortality prediction, the best ANN model with 10 hidden neurons showed 75.6% accuracy in training dataset and 74% accuracy in test dataset (51.5% specificity, 84.5% sensitivity). The RBFNN with 40 neurons in a hidden layer had higher accuracy at level of 79.7% in training data set and 77.4% in test dataset (62.1% specificity, 84.5% sensitivity). The best ANN model with 10 hidden neurons predicted poor outcome with 79.9% accuracy in training dataset and 79.3% accuracy in test dataset (84.9% specificity, 70.7% sensitivity). The RBFNN performed with better accuracy of 80.9% in training dataset and 80.8% in test dataset (84.9% specificity, 74.4% sensitivity). The ANN and RBFNN outperformed accuracy of binary logistic regression. Conclusions: ANN can be useful to support prediction of o[...].