Predicting Unplanned Trauma ICU Admissions for Initial Nonoperative, Non-ICU Patients.

Abstract

Introduction: Unplanned intensive care unit (ICU) admissions are associated with increased morbidity and mortality. This study uses interpretable machine learning to predict unplanned ICU admissions for initial nonoperative trauma patients admitted to non-ICU locations.

Methods: TQIP (2020-2021) was queried for initial nonoperative adult patients admitted to non-ICU locations. Univariable analysis compared patients who required an unplanned ICU admission to those who did not. Using variables that could be known at hospital admission, gradient boosting machines (CatBoost, LightGBM, XGBoost) were trained on 2021 data and tested on 2020 data. SHapley Additive exPlanations (SHAP) were used for interpretation.

Results: The cohort had 1,107,822 patients; 1.6% had an unplanned ICU admission. Unplanned ICU admissions were older (71 [58-80] vs. 61 [39-76] years, p < 0.01), had a higher Injury Severity Score (ISS) (9 [8-13] vs. 9 [4-10], p < 0.01), longer length of stay (11 [7-17] vs. 4 [3-6] days, p < 0.01), higher rates of all complications and most comorbidities and injuries (p < 0.05). All models had an AUC of 0.78 and an F1 score of 0.12, indicating poor performance in predicting the minority class. Mean absolute SHAP values revealed ISS (0.46), age (0.29), and absence of comorbidities (0.16) as most influential in predictions. Dependency plots showed greater SHAP values for greater ISS, age, and presence of comorbidities.

Conclusions: Machine learning may outperform prior attempts at predicting the risk of unplanned ICU admissions in trauma patients while identifying unique predictors. Despite this progress, further research is needed to improve predictive performance by addressing class imbalance limitations.