Quantifying brain-functional dynamics using deep dynamical systems: Technical considerations.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Computational Psychiatry

Background:

• Mental health and illness dynamics are complex and unpredictable.
• Artificial intelligence (AI) and dynamical systems reconstruction offer novel approaches to characterize brain dynamics.
• Understanding these dynamics is crucial for identifying potential biomarkers for mental illness.

Purpose of the Study:

• To analyze computational challenges in applying deep learning to model individual-level dynamical systems.
• To demonstrate the impact of these challenges on classifying psychiatric disorders using functional magnetic resonance imaging (fMRI) data.
• To guide the development of reproducible, individual-level generative models for mental illness biomarkers.

Main Methods:

• Utilized generative modeling of brain dynamics from fMRI data as a case study.
• Performed an extensive analysis of challenges associated with deep learning parameter optimization.
• Investigated the impact of model variability on the classification of schizophrenia and major depression.

Main Results:

• Deep learning models exhibit a tendency to find unique solutions during parameter optimization.
• This variability significantly hinders the reproducibility of downstream predictions.
• The identified challenges impact the clinical utility of AI-driven biomarkers.

Conclusions:

• Reproducibility is a critical hurdle for individual-level generative models in neuroscience.
• Deep learning's tendency for unique solutions challenges the development of reliable biomarkers for mental illness.
• Future research should focus on developing methods for reproducible AI in clinical neuroscience.