Machine Learning With Neuroimaging: Evaluating Its Applications in Psychiatry.

Area of Science:

• Neuroscience
• Psychiatry
• Machine Learning

Background:

• Psychiatric disorders exhibit complex, heterogeneous neurobiology, challenging traditional research methods.
• Multivariate pattern classification and supervised machine learning (ML) are increasingly used with neuroimaging data to understand these complexities.
• These ML approaches present unique challenges in study design and interpretation.

Purpose of the Study:

• To establish best practices for evaluating ML applications in psychiatric research.
• To guide the assessment of ML for clinical prediction (diagnosis, prognosis, treatment) and mechanistic insights.
• To use functional connectivity magnetic resonance imaging (fc-MRI) as a specific example.

Main Methods:

• Review of current ML methodologies applied to psychiatric neuroimaging.
• Discussion of evaluation criteria for clinical prediction models.
• Examination of interpretability and reliability of neuroimaging features derived from ML.

Main Results:

• ML classification for individual-level predictions requires clinical informativeness, independence from confounds, and rigorous performance/generalizability assessment.
• Understanding psychiatric mechanisms via ML necessitates careful consideration of feature utility, interpretability, and reliability.
• Large, public datasets enhance the potential utility of ML in psychiatry.

Conclusions:

• Adherence to best practices is crucial for the translational utility of ML in psychiatry.
• Robust evaluation frameworks are needed for both predictive and mechanistic ML studies.
• Future research should leverage large datasets and focus on interpretable and reliable neuroimaging features.