Near Perfect Prediction of Mortality 24 Hours in Advance in a Cardiovascular Intensive Care Unit Using Artificial Intelligence

J. Del Gaizo¹, Z. W. Sollie¹, A. Wright¹, A. Kilic^2
1Medical University of South Carolina, Charleston, South Carolina ²MUSC, Charleston, South Carolina

Purpose: Artificial intelligence (AI) may be able to detect high probability of mortality in a cardiovascular intensive care unit (CVICU) hours in advance. This study evaluated the prognostic utility of AI leveraging only continuous vital sign and laboratory data in a CVICU to predict mortality up to 24 hours before death.

Methods: Continuous physiological data including vital signs and serial laboratory parameters (393 features) were collected from patients admitted to a single-center CVICU between January 28, 2021 and November 30, 2023. The primary outcome was CVICU all-cause mortality. An AI model (customized neural network) was trained to predict mortality from 48 hours of sensor data and 24 hours of laboratory data from all patients; data within 15 minutes of death was excluded. Summary statistics of each sensor’s data for each patient were fed into a pre-processing pipeline, followed by the neural network. After the network was trained, the same model was used to predict if a subject had mortality at pre-death buffers of 15-minutes, 6-hours, 12-hours, and 24-hours. The models were assessed using area under the receiver operating curve (AUC) on stratified holdout sets across 50 iterations.

Results: A total of 2,845 CVICU patients were identified, with 214 mortalities and 2,631 survivors.
The model obtained exceptional performance in the holdout sets at each of the pre-mortality time buffers: 0.974±.017 AUC (15 minutes); 0.961±.018 AUC (6 hour); 0.950±.023 AUC (12 hour); and 0.939±.022 AUC (24 hour). The most predictive features for mortality included 1) lactate values; 2) Oxygen saturation; 3) systolic arterial blood pressure; 4) central venous pressure; 5) ventilation rate; 6) mean arterial blood pressure; and 7) arterial blood gas values.

Conclusion: A low parameter neural network can achieve state-of-the-art AUC for detection of CVICU mortality as early as 24-hours pre-mortality. These AI models may have important implications in early warning signs for clinicians and for prognostic discussions. Furthermore, it is readily deployable as only vital sign and laboratory data are required.

Identify the source of the funding for this research project: There is no dedicated funding to disclose for this project.