ABSTRACT
The increasing availability of electronic health records and administrative data and the adoption of computer-based technologies in healthcare have significantly focused on medical informatics. Sepsis is a time-critical condition with high mortality, yet it is often not identified in a timely fashion. The early detection and diagnosis of sepsis can increase the likelihood of survival and improve long-term outcomes for patients. In this paper, we use SHapley Additive exPlanations (SHAP) analysis to explore the variables most highly associated with developing sepsis in patients and evaluating different supervised learning models for classification. To develop our predictive models, we used the data collected after the first and the fifth hour of admission and evaluated the contribution of different features to the prediction results for both time intervals. The results of our study show that, while there is a high level of missing data during the early stages of admission, this data can be effectively utilized for the early prediction of sepsis. We also found a high level of inconsistency between the contributing features at different stages of admission, which should be considered when developing machine learning models.
