Detecting dynamical changes in vital signs using switching Kalman filter.

Area of Science:

• Biomedical Engineering
• Time Series Analysis
• Clinical Monitoring

Background:

• Vital signs are crucial for monitoring patient health in general wards.
• Detecting changes in vital sign dynamics can indicate disease progression.
• Time series data of vital signs often exhibit non-stationarity, complicating analysis.

Purpose of the Study:

• To quantify the bias introduced by non-stationarities in switching models.
• To evaluate the impact of spectral, symbolic, and entropy indices as inputs.
• To compare switching Kalman filter (SKF) models trained on stationary versus non-stationary data.

Main Methods:

• Utilized a test to differentiate stationary and non-stationary periods based on mean and variance stability.
• Trained SKF models using indices derived from both stationary and non-stationary vital sign periods.
• Compared model performance based on input data characteristics.

Main Results:

• Significant differences were observed between indices measured during stationary and non-stationary periods.
• Switching model outcomes were highly dependent on the input indices used.
• Multi-scale entropy (MSE) demonstrated the highest correlation (38%) between non-stationary and stationary switches.

Conclusions:

• Non-stationarities in vital sign data can introduce bias into switching models.
• The choice of input indices significantly impacts the reliability of switching models.
• MSE shows potential for robustly capturing vital sign dynamics across stationary and non-stationary states.