Sepsis prediction model based on vital signs related features

Abstract

In this paper, we present our solution for early detection of sepsis by joining the PhysioNet/Computing in Cardiology Challenge 2019. Our proposed algorithm uses three different models for sepsis prediction. The model takes into account the amount of time the patients have already spent in the intensive care unit. The first model uses 64 features and is applied if the patient stays in ICU for the first 9 hours. Second and third models use 111 more advanced features. The second prediction model is activated if the patient stays for more than 9 hours. The third one is activated for more extended stays if the patient stays for more than 60 hours. The time patients spent in the ICU or were hospitalized is an essential indicator for the risk of developing sepsis. During the longer stays in hospital number of intravenous measurements and other procedures increases, increasing the risk of blood infection. Therefore, feature extraction in our algorithm was based on these metrics. The best-received score with the models, trained using Gentle Boosting on a training set with ADASYN balancing was lower than the best score with the models, trained on a dataset with randomly removed samples. The official our team VGTU utility score on full test set was 0.014 and ranked at 66 place.