Risk Estimation for ICU Patients with Personalized Anomaly-Encoded Bedside Patient Data

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Critical Care Medicine

Background:

• Conventional threshold-based monitoring in intensive care units (ICUs) has limitations in capturing dynamic patient health changes.
• Continuous patient data analysis is crucial for timely and accurate health risk assessment in critical care settings.

Purpose of the Study:

• To develop a novel framework for continuous health risk estimation in ICU patients.
• To leverage anomaly-encoded patient data for improved risk prediction models.
• To enhance bedside patient monitoring beyond traditional threshold-based systems.

Main Methods:

• Utilized Gaussian process models to learn temporal trends and diurnal patterns in patient data.
• Developed a two-module framework: one for anomaly detection and trend learning, another for risk prediction.
• Encoded detected anomalies and learned Gaussian models into patient data for subsequent risk estimation.

Main Results:

• The proposed framework enables dynamically adaptable bedside patient monitoring.
• Prediction models incorporating anomaly-encoded data demonstrated superior performance over models using raw patient measurements.
• The framework's predictive accuracy for in-hospital mortality and length of stay is comparable to state-of-the-art models.

Conclusions:

• Anomaly-encoded data enhances the accuracy of patient health risk estimation in ICUs.
• The novel framework offers a more adaptive and effective approach to critical care patient monitoring.
• This method holds promise for improving patient outcomes through advanced predictive analytics.