Predicting Imminent Mortality in a Cardiovascular Intensive Care Unit Using Artificial Intelligence

Z. Whitten. Sollie¹, J. Del Gaizo², R. Mathi³, B. Welch⁴, S. M. Zeigler², A. Kilic^3
1Medical University of South Carolina, Wando, South Carolina ²Medical University of South Carolina, Charleston, South Carolina ³MUSC, Charleston, South Carolina ⁴Medical University of South Carolina (MUSC), Mount Pleasant, South Carolina

Purpose: Automated systems that combine physiological data with recent laboratory parameters have the potential to warn of imminent mortality in critically ill patients. The aim of this study was to develop and evaluate an artificial intelligence (AI)-based model for predicting imminent mortality in a cardiovascular intensive care unit (CVICU).

Methods: Continuous data were collected from patients admitted to a single-center CVICU between 2021-2022. The primary outcome was imminent mortality occurring in 15 minutes using data from the prior 6 hours. There were 2 AI-based approaches: (1) a limited approach using only arterial blood pressure measurements that is likely robust across institutions, and (2) an encompassing approach that uses more extensive laboratory (393 features) and physiological data (42 sensors) but may be susceptible to institution-based overfitting. The pipeline concatenates the sensor summaries and laboratory data, removes 0-variance train set variables, median imputes, normalizes, and feeds the data to a linear support vector classifier to predict imminent mortality. Models were assessed using area under the receiver operating curve (AUC) on stratified holdout sets. In each iteration, training (72.25%), validation (12.75%) and true holdout sets (15%) were created.

Results: A total of 324 (97 with CVICU mortality) and 1,472 (144 with CVICU mortality) patients were used in the limited and encompassing approaches, respectively. For the limited approach, the holdout AUCs across 10 iterations were: (mean AUC, (IQR)): Systolic arterial blood press (ABPs), 0.869 (IQR: 0.851, 0.900), Diastolic arterial blood pressure (ABPd), 0.788 (IQR: 0.750, 0.825), Combination arterial blood pressure (ABP-both), 0.844 (IQR: 0.803, 0.860). For the encompassing approach, the holdout AUCs across 50 iterations included: physiologic sensors only, 0.952 (IQR: 0.936, 0.967); laboratory parameters only, 0.867 (IQR: 0.842, 0.896); Both, 0.977 (IQR: 0.970, 0.986). (Figure)

Conclusion: AI can be employed for automated detection of imminent mortality using 6 hours of CVICU sensor data. Excellent performing models were attained in our limited approach, which improved even further to state-of-the-art performance with an AUC of 0.98 when expanded in our encompassing approach.

Identify the source of the funding for this research project: No specific funding was used for this research project