A comparison between discrete and continuous time Bayesian networks in learning from clinical time series data with...

Area of Science:

• Machine Learning
• Biomedical Informatics
• Time Series Analysis

Background:

• Mobile devices are increasingly used for healthcare data collection, generating complex, irregularly spaced clinical time-series data.
• Analyzing such data, particularly for home monitoring in chronic obstructive pulmonary disease (COPD), presents challenges for machine learning.
• This study focuses on evaluating temporal Bayesian network models for this type of data.

Purpose of the Study:

• To determine which temporal Bayesian network models best handle irregularly spaced multivariate clinical time-series data.
• To compare the performance of dynamic Bayesian networks (discrete-time) and continuous time Bayesian networks.

Main Methods:

• Two models were studied: dynamic Bayesian networks (discrete time) and continuous time Bayesian networks.
• Performance was evaluated on regularly and irregularly spaced synthetic data, including data missing completely at random and not at random.
• Real-world, unevenly spaced COPD patient data was also used for validation.

Main Results:

• Dynamic Bayesian networks outperformed continuous time models for regularly spaced data and data missing completely at random.
• For data not missing completely at random, continuous time Bayesian networks performed better, especially without prior knowledge of missingness.
• Both models performed similarly on real-world unevenly spaced COPD data, with continuous time models offering more detailed predictions.

Conclusions:

• Discrete-time models are suitable for regularly spaced data and data missing completely at random.
• Continuous-time models are competitive or superior for complex, irregularly spaced clinical time-series data.
• Continuous-time models offer practical advantages in clinical applications due to their ability to provide fine-grained predictions.