Prediction of Neurological Deterioration of Patients with Mild Traumatic Brain Injury Using Machine Learning

Abstract

Possible Neurological Deterioration (ND) of patients with Traumatic Brain Injury (TBI) is difficult to identify especially the mild and moderate injuries. When ND happens, death or lifelong disability is prevalent. Early prediction of possible ND would allow medical and healthcare institutions to provide the needed medical treatment. This paper presents the results that show Machine Learning (ML) can be used to create predicative models with high prediction rates even with a small set of patient records (219 patient records with 54 variables). From the patient records, 20 randomized data sets with preconditions on the testing and training data were created and fed to selected Artificial Neural Network (ANN) and Classification Algorithms. Preconditions on testing and training data can affect the prediction models created by the different algorithms. The best prediction models created by the ANN algorithms (multilayer perceptron (MLP), recurrent neural network (RNN), and long short-term memory (LSTM)) and two classification algorithms (linear regression and logistic regression algorithms) are considered acceptable and could be applied as medical decision support to identify patients that may potentially have ND. Early prediction of a possible ND of a patient can now be easily carried out as soon as his or her records and medical test results are ready and match the 54 variables needed for prediction.