Prediction of Mortality in Trauma Patients with Insufficient Training Data Using Deep Learning

Abstract

The trauma victim is in an emergency phase and needs immediate help. This phase is very sensitive to time, considering that the most common trauma is bleeding. The patient will die if the help is not given immediately. Trauma is comprehended as a degree of severity. Medical treatment is given based on this degree. However, methods to measure the degree currently used require quite some time, which risks the patient’s life. This study proposes a standardized measurement of the degree of trauma and predicts patient mortality based on selected features. We select and discard medical data that does not affect the degree of trauma. The data becomes simple. The primary trauma dataset was collected from the emergency room at Hasan Sadikin Hospital in Bandung, Indonesia. The data has 19 features and two prediction classes for mortality, namely death and life. Data mining techniques were carried out to provide a sufficient dataset so that the algorithm could work optimally. This study performs several steps of data pre-processing as follows: cleaning, correcting, and augmenting the dataset in order to obtain significant results. Then reduce some features in light of the fact that some have overlapped. We implement a long-short-term memory deep learning algorithm to build a mortality prediction model. Although there were inconsistencies in the data and an insufficient amount, the result was promising. Features become concise and comprehensive. It was accurate up to 92% of the time and took less time than the other methods.