Learning outcome-discriminative dynamics in multivariate physiological cohort time series

Area of Science:

• Physiological modeling
• Time series analysis
• Machine learning

Background:

• Physiological model identification is challenging due to changing operating regimes and measurement artifacts.
• Existing methods struggle to accurately capture dynamic physiological states and artifacts simultaneously.

Purpose of the Study:

• To develop a novel learning algorithm for Switching Linear Dynamical Systems (SLDS) to identify physiological time series dynamics.
• To improve the prediction of physiological regimes and outcomes by modeling switching dynamics and artifacts.

Main Methods:

• Utilized the Switching Linear Dynamical Systems (SLDS) framework to model physiological time series.
• Developed a novel learning algorithm for SLDS tailored for physiological data analysis.
• Applied the method to a simulation study and a real-world physiological classification task.

Main Results:

• Demonstrated significant improvement in classification accuracy compared to expectation maximization-based feature learning.
• Successfully decoded postural changes from heart rate and blood pressure data.
• The proposed algorithm effectively identifies underlying physiological dynamics and distinguishes them from artifacts.

Conclusions:

• The novel SLDS learning algorithm offers a robust solution for identifying dynamics in physiological time series.
• This approach enhances the accuracy of physiological state classification and outcome prediction.
• The general framework can be extended to various state-space modeling applications in biomedical research.