preprints.org > mathematics & computer science > algebra & number theory > 202010.0113.v1

Working Paper Article Version 1 This version is not peer-reviewed

Version 1 : Received: 5 October 2020 / Approved: 6 October 2020 / Online: 6 October 2020 (09:42:43 CEST)

Sepsis is the most common cause of death in emergency departments (ED) that is associated with suspected infections. Currently, the international heuristic most used by physicians are the SIRS (Systemic Inflammatory Response Syndrome) criteria. However, widespread inaccuracy of this criteria has caused physician mistrust and desensitization to symptoms, thereby increasing mortality rates. This study sourced a data set from a Swedish study of the electronic medical health records of 18,006 patients, and processed the data down to 2, 196 unique patients. 26 separate machine learning models from 7 different families were tested for their performance when given the same inputs as the SIRS criteria’s requirements. It was found that the two-layer feedforward neural network performed best and was eligible for optimization in order to achieve the best classification performance to compare with the SIRS criteria. Training functions were tested for error and execution time, and the best function for use in the neural network was determined to be the Levenberg-Marquardt function. 10 baseline covariates with the most relevance to Sepsis diagnosis were combined with the SIRS criteria to achieve the best and most realistic model performance. The resulting neural network had a sensitivity of 98.1% and a specificity 63.4%, far exceeding other machine learning models and the SIRS criteria.

machine learning; disease prediction; neural networks; sepsis