A transformer-based diffusion probabilistic model for heart rate and blood pressure forecasting in Intensive Care...

Area of Science:

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

Background:

• Intensive Care Unit (ICU) vital sign monitoring is critical for timely patient interventions.
• Accurate predictive systems are essential for improving patient outcomes in critical care settings.
• Existing methods for vital sign forecasting in the ICU require enhancement.

Purpose of the Study:

• To develop a novel deep learning approach for forecasting Heart Rate (HR), Systolic Blood Pressure (SBP), and Diastolic Blood Pressure (DBP) in the ICU.
• To introduce the Transformer-based Diffusion Probabilistic Model for Sparse Time Series Forecasting (TDSTF).
• To evaluate the performance and efficiency of the proposed TDSTF model against existing methods.

Main Methods:

• Utilized 24,886 ICU stays from the MIMIC-III database for model training and testing.
• Developed the TDSTF model by integrating Transformer and diffusion probabilistic models for sparse time series forecasting.
• Compared TDSTF performance against baseline models using metrics such as Standardized Average Continuous Ranked Probability Score (SACRPS) and Mean Squared Error (MSE).

Main Results:

• TDSTF achieved a SACRPS of 0.4438 and an MSE of 0.4168.
• Demonstrated an 18.9% improvement in SACRPS and a 34.3% improvement in MSE over the best baseline model.
• Exhibited a computational efficiency, with an inference speed over 17 times faster than the best baseline model.

Conclusions:

• The TDSTF model is an effective and efficient deep learning solution for forecasting vital signs in the ICU.
• TDSTF significantly outperforms existing models in predicting vital sign distributions and overall accuracy.
• The proposed model offers a promising advancement for real-time patient monitoring and intervention in critical care.