Time-Padded Random Forest Ensemble to Capture Changes in Physiology Leading to Sepsis Development
Benjamin Sweely, Austin Park, Lia Winter, Longjian Liu, Xiaopeng Zhao · Computing in cardiology · 2019
Background: Sepsis is the leading cause of in-hospital deaths, and it is one of the costliest complications to treat.Detection of sepsis is complicated and not yet efficient.Each hour of delay in treatment for a septic patient results in a 4-8% increase in chance of mortality.Method: The dataset provided consists of files that contain hourly parameter measurements for over 40,000 unique patients.Due to the complex nature of this challenge problem, a model of similar complexity was necessary.A boosted random forest ensemble was chosen and developed in MATLAB in hopes of producing the best results for this challenge.The provided data was time padded for 8 additional hours' worth of data, 10-fold cross-validated, and imputed with previous data.Many ensemble methods were tested with Random Under-Sampling Boosting performing the best.For this model, the hyper-parameters were optimized via a grid search to find an optimal model.Results: Using the optimized hyper-parameters along with the correct pre-processing techniques, a 10-fold average utility score of 0.421 was achieved on the training sets A and B combined.We participated in Physionet Challenge under the name SOS: Searching of Sepsis and the utility score on full test set is 0.314.Our official rank is #14.