Understanding heterogeneity in psychiatric disorders: A method for identifying subtypes and parsing comorbidity

Area of Science:

• Neuroscience
• Psychiatry
• Machine Learning

Background:

• Psychiatric and neurodevelopmental disorders are highly heterogeneous, presenting challenges in identifying distinct subtypes and comorbidities.
• Distinguishing disorder-related neural variations from normal variation is difficult in patient populations.

Purpose of the Study:

• To evaluate the capability of Contrastive Variational Autoencoders (CVAEs) in identifying disorder-related individual differences, capturing disease subtypes, and disentangling comorbidity.
• To compare CVAE performance against baseline models under varying hyperparameters and data availability.
• To introduce and test a novel architecture and training procedure for CVAEs to enhance reproducibility and model comorbidity.

Main Methods:

• Utilized synthetic neuroanatomical MRI data with known ground truth for shared and disorder-specific effects.
• Assessed CVAE and non-contrastive baseline models for detecting disorder subtypes and disentangling comorbidity.
• Introduced a new CVAE architecture for comorbid disorders and a novel training procedure.

Main Results:

• CVAE models significantly outperformed non-contrastive methods in correlating with disorder-specific effects and discovering subtypes.
• The novel CVAE architecture successfully disentangled neuroanatomical differences associated with comorbid disorders.
• The improved training procedure reduced result variability by up to 5.5 times, enhancing reproducibility.

Conclusions:

• CVAEs provide a robust framework for precision psychiatry, enabling reliable detection of interpretable neuromarkers.
• The study demonstrates CVAEs' effectiveness in discovering psychiatric disorder subtypes and disentangling complex comorbidities.
• The enhanced CVAE approach offers improved reliability and accuracy for neuroimaging research in psychiatry.