Explainable Machine Learning for Early Prediction of In-Hospital Mortality Among Patients with Sepsis: A Clinical Decision-Support Study

Abstract

Sepsis is a serious clinical problem which needs prompt recognition of patients at high risk of in-hospital death. This research proposed and evaluated explainable machine-learning models to predict early mortality based on demographic and hospitalization data. A retrospective predictive-modelling design was used with 1,511 individual patient records for age, sex, hospital episode number and in-hospital outcome. Data were split into a training set and a testing set in stratified manner, and the performance of logistic regression, decision tree, random forest, support vector machine and gradient boosting models were assessed. Model performance was evaluated based on the metrics of ROC-AUC, precision–recall AUC, accuracy, sensitivity, specificity, F1 score, balanced accuracy and Brier score. Logistic-regression estimates and permutation-based feature importance were used to examine explainability. Overall mortality was 35.5%. The non-survivors were older than the survivors and there was a sharp rise in mortality at higher ages. Logistic regression had the best discrimination (ROC-AUC of 0.720), sensitivity of 67.3%, specificity of 61.2% and accuracy of 63.4%. The strongest predictor of mortality was age with an inverse association observed for hospital episode number. Sex was not independently significant. The results suggest that straightforward, understandable variables are useful for initial mortality-risk stratification, but that the moderate performance restricts the ability of these variables to be used independently at the bedside. Before clinical implementation, there is a need of external validation and integration of physiological, laboratory, organ-function and treatment-related variables.