Multitask Gaussian processes for multivariate physiological time-series analysis.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Time Series Analysis

Background:

• Gaussian process (GP) models are powerful for nonparametric Bayesian regression but are often limited to single time series (STGPs) in healthcare.
• The proliferation of sensors in healthcare necessitates advanced tools for analyzing multiple, correlated physiological signals.

Purpose of the Study:

• To propose and evaluate multitask Gaussian process models (MTGPs) for simultaneous analysis of multiple correlated physiological time series.
• To demonstrate the capability of MTGPs to handle signals sampled at different frequencies and with varying data availability.
• To integrate prior knowledge about signal relationships, such as delays and temporal behavior, into the modeling framework.

Main Methods:

• Developed a flexible multitask Gaussian process (MTGP) framework for multivariate time-series analysis.
• Introduced a novel normalization technique for improved hyperparameter interpretability within MTGPs.
• Applied MTGPs to synthetic datasets and two real-world patient monitoring and radiotherapy scenarios.

Main Results:

• MTGPs successfully learned correlations between multiple physiological time series, even with differing sampling rates and data intervals.
• The proposed MTGP framework demonstrated superior performance compared to standard Gaussian processes and existing methods in the evaluated biomedical applications.
• A novel normalization method enhanced the interpretability of MTGP hyperparameters.

Conclusions:

• Multitask Gaussian process models offer a robust and efficient solution for analyzing complex, multivariate physiological time series in healthcare.
• MTGPs provide a flexible framework for integrating prior knowledge and handling heterogeneous data, outperforming current state-of-the-art methods.
• This approach advances physiological monitoring capabilities through advanced machine learning techniques.