Temporal generative models for learning heterogeneous group dynamics of ecological momentary assessment data

Area of Science:

• Computational psychiatry
• Machine learning for healthcare
• Psychiatric data science

Background:

• Precision psychiatry seeks individualized mental disorder characterization using dynamic processes.
• Ecological momentary assessment (EMA) via mobile tech yields high-frequency, multi-dimensional, correlated, and hierarchical data.
• Existing models like mixed-effects models and standard RTRBMs have limitations in handling EMA data complexity and heterogeneity.

Purpose of the Study:

• To introduce a novel temporal generative model, the Heterogeneous Dynamic Restricted Boltzmann Machine (HDRBM).
• To address the limitations of existing models in capturing heterogeneous group dynamics within populations using covariates.
• To demonstrate the HDRBM's effectiveness on simulated and real-world EMA datasets.

Main Methods:

• Development of the HDRBM, a temporal generative model incorporating covariates to learn heterogeneous group dynamics.
• Application and evaluation of the HDRBM on simulated and real-world ecological momentary assessment datasets.
• Comparison of HDRBM performance against existing modeling approaches for temporal, hierarchical data.

Main Results:

• The HDRBM significantly improves accuracy and interpretability in modeling EMA data compared to existing methods.
• Incorporating covariates within the HDRBM framework allows for exploration of underlying drivers of participant group dynamics.
• The HDRBM effectively serves as a generative model for complex EMA studies.

Conclusions:

• The HDRBM offers a powerful new approach for analyzing complex, heterogeneous EMA data in precision psychiatry.
• This model enhances our ability to understand individual and group dynamics in mental health research.
• HDRBM facilitates more accurate and interpretable insights into mental disorders by leveraging covariate information.