Detecting Organ Failure in Motor Vehicle Trauma Patients: A Machine Learning Approach

Motor vehicle accidents are prevalent throughout the U.S.; over five million crashes occur each year. Typically, patients are appropriately diagnosed only after being transported to nearby medical centers, which currently takes about 7–15 minutes on average. This project aims to expedite this diagnostic process by training neural networks on Intensive Care Unit (ICU) data to automate the identification of injuries caused by motor vehicle accidents. Data was aggregated from the Beth Israel Deaconess Medical Center in Boston, Massachusetts via the Medical Information Mart for Intensive Care III (MIMIC-III) database, for patients' electrocardiogram (ECG) and respiratory rate reports. Natural Language Processing was used to isolate the patients of interest and classify them based on the area of injury, i.e., any of eight combinations of the three vital organs: brain, heart, and lungs. Upon isolation of the waveform data, noise was filtered via normalization, Butterworth and forward-backward filter application, and Fast Fourier transformation. The trained Artificial Neural Network (ANN) contained 23 dense layers with the number of neurons decreasing per two layers with ReLU activation and 10% dropout. Multiclass activation was used for the final layer, in order to allow training for detection of multiple areas of trauma. The trained Convolutional Neural Network (CNN) had 12 convolutions, utilizing batch normalization, six pooling, one flatten, and two dense layers. Both models used AdamOptimizer. The CNN and ANN produced F1 scores of 0.82 and 0.56 respectively, suggesting the models could potentially perform accurate early detection of injury area and expedite hospital treatment of trauma patients.