Understanding disease processes by partitioned dynamic Bayesian networks.

Area of Science:

• Computational Biology
• Medical Informatics
• Statistical Modeling

Background:

• Clinical data often exhibits temporal changes due to interventions, yet time-homogeneous models like dynamic Bayesian networks fail to capture these dynamics.
• Existing models lose crucial pathophysiological process specificities when applied to evolving clinical conditions.

Purpose of the Study:

• To introduce Partitioned Dynamic Bayesian Networks (PDBNs) for modeling time-varying clinical data and capturing distribution regime changes.
• To develop a heuristic algorithm for learning non-homogeneous models that balance specificity and simplicity.

Main Methods:

• Proposed Partitioned Dynamic Bayesian Networks (PDBNs) for representing time non-homogeneity.
• Developed a heuristic algorithm for learning PDBNs, prioritizing model simplicity.
• Conducted extensive simulation experiments to evaluate the algorithm's performance.

Main Results:

• The heuristic algorithm successfully constructed models that accurately reflect underlying data-generating processes, whether homogeneous or non-homogeneous.
• Simulated experiments demonstrated good fit and statistical accuracy of the learned models.
• The approach effectively captured distribution regime changes inherent in clinical data.

Conclusions:

• Partitioned Dynamic Bayesian Networks offer an intuitive and theoretically sound method for modeling time-varying pathophysiological processes.
• The developed heuristic algorithm provides a practical solution for learning complex, non-homogeneous models from clinical data.
• Application to psychotic depression yielded clinically relevant insights into the disorder's dynamics.