A Hypotensive Episode Predictor for Intensive Care based on Heart Rate and Blood Pressure Time Series

Area of Science:

• Critical Care Medicine
• Biomedical Engineering
• Artificial Intelligence in Healthcare

Background:

• Prompt intervention for hypotensive episodes (HEs) is crucial in intensive care units (ICUs).
• Predictive tools for identifying patients at risk of HEs hours in advance are needed.
• Existing methods may lack the automation and real-time predictive capabilities required.

Purpose of the Study:

• To develop and evaluate an automated artificial neural network (ANN) predictor for hypotensive episodes (HEs).
• To assess the impact of prediction-to-onset time gaps on HE prediction performance.
• To explore the utility of weighted posterior probabilities for improving prediction accuracy.

Main Methods:

• An artificial neural network (ANN) model was developed using heart rate and blood pressure time series data from the MIMIC II database.
• The predictor utilized a 30-minute observation window preceding the prediction time.
• The gap between prediction time and the 30-minute HE target window was varied from 1 to 4 hours.

Main Results:

• Prediction performance for HEs degraded as the gap size between prediction and onset increased.
• A weighted posterior probability scheme showed negligible improvement in performance.
• The best mean area under the ROC curve achieved was 0.934, despite low positive predictive values.

Conclusions:

• The developed ANN predictor demonstrates potential for early detection of hypotensive episodes in ICUs.
• Prediction accuracy is sensitive to the lead time of the prediction.
• Further refinement may be needed to improve positive predictive values for clinical utility.