MONITOR: a multi-domain machine learning approach to predicting in-hospital mortality

Abstract

Background: Current machine-learning (ML) models have been developed to predict mortality for specific diseases, procedures, and setting at a given time; however, the risk for in-hospital mortality changes throughout a patient’s hospital stay. A model that could predict in-hospital mortality throughout a patient’s stay regardless of disease or procedure could improve clinical outcomes. Methods: We conducted a prognostic study where cohorts were created from electronic health records (EHR) with encounters between January 1, 2014 and January 30, 2020 at tertiary academic hospital and community hospital. The initial dataset contained 228,405 patients. EHR of 176,526 patients remained in the study after adjusting for age (18 or older), length of stay (LOS) (between 0 and 365 days), and encounter dates within study period. Training and testing cohorts, stratified by length-of-stay and in-hospital mortality, were created with an 80/20 split. Results: The study included 176,526 patients {mean [interquartile range (IQR)] age of 52.2 [34–68] years; 55.3% female, 63.7% white, 92.7% non-Hispanic} who were admitted for 5.6 [2–6] days. The in-hospital mortality rate for the training and testing cohorts was 3.0%. The CatBoost classifier model, trained with a combination of undersampling and oversampling, demonstrated a F2 score of 0.510 [95% confidence intervals (CI): 0.496–0.516]. The F2 score is highest for patients with a one-day LOS (0.811; 95% CI: 0.776– 0.843). Even though the F2 score is lower for patients who stayed more than a day, the F2 score generally increases each day until the day of discharge or mortality. Conclusions: This study investigated an ML model that predicted risk of in-hospital mortality regardless of patient demographics and level of care setting. The model accounted for changes in patient condition throughout the LOS. An implementation study should be conducted to determine how this model can be integrated into clinical workflow to support decision making.