Early Prediction of Sepsis Using a Sliding Window-based AdaBoost Learning and Bayesian Regression

Meicheng Yang, Hongxiang Gao, Xingyao Wang, Fan Zhou, Jianan Di, Tiantian Wang, Jianqing Li, Chengyu Liu
Southeast University


Sepsis is a life-threatening condition and has caused extensive con-cern. The PhysioNet/Computing in Cardiology Challenge 2019 focuses on the early prediction of sepsis from multi-model clinical data, aiming to reduce the sepsis severe morbidity, mortality and medical costs. To address this challenge, we proposed a sliding window-based Ada-Boost learning and Bayesian regression method to predict the onset of sepsis up to six hours in advance. First, balanced records from non-sepsis group were randomly selected from the training data, generating a total of 4,600 records with equal number records of sepsis and non-sepsis. Then, a six-hour sliding window was applied to segment each record with a step of one hour, resulting in a total of 220,647 six-hour segments. If sepsis event fell within the window, this segment was labelled as ‘sepsis’, oth-erwise no. Features were extracted from all available patient co-variates, as well as from their differential time series. Gentle AdaBoost learning model was trained for determining a six-hour segment as ‘sepsis’ or not. The 0/1 results for the six-hour segments were merged using a Bayesian regression method to form the final onset prediction of sepsis event. Evaluations were performed using a standard five-fold cross-validation. Results reported an area under receiver operating characteristic (AUROC) curve of 0.787 for sepsis 0/1 classification on the six-hour segments. The overall score of Utility defined by the challenge organizer was 0.416 from the five-fold cross-validation. In addition, the maximum running time was only 19% of quota, indicating the computing efficiency of our proposed method.