Identifying sub-populations via unsupervised cluster analysis on multi-edge similarity graphs

Area of Science:

• Neuroscience
• Computational Biology
• Psychiatry

Background:

• Autism spectrum disorder (ASD) and schizophrenia (SCZ) are complex psychiatric disorders characterized by significant heterogeneity.
• Traditional supervised classification methods struggle to capture the underlying etiological diversity within these diagnostic categories.
• Unsupervised approaches are needed to explore and define patient subgroups based on comprehensive data.

Purpose of the Study:

• To develop and validate an unsupervised clustering framework using multi-edge similarity graphs.
• To combine multimodal data (structural networks, cognitive scores) for enhanced patient stratification.
• To investigate the heterogeneity within autism spectrum disorder (ASD) and schizophrenia (SCZ) populations.

Main Methods:

• Utilized unsupervised cluster analysis on multi-edge similarity graphs integrating information from different modalities.
• Developed a novel 'holding power' concept to optimize edge weights, reflecting subject certainty within clusters.
• Applied the framework to clinical populations of ASD and SCZ, combining structural network and cognitive data.

Main Results:

• The method successfully clustered the ASD-control population into two classes with 92% accuracy against diagnostic labels.
• The SCZ-control population was clustered into four distinct groups, achieving 78% accuracy with diagnostic labels.
• Results indicate significant heterogeneity within the SCZ population, more so than in the ASD population.

Conclusions:

• The multi-edge similarity graph approach offers a powerful tool for unsupervised stratification of complex psychiatric disorders.
• This method effectively elucidates underlying patient heterogeneity, outperforming traditional supervised methods.
• The findings underscore the distinct subgroup structures within ASD and particularly SCZ, paving the way for personalized medicine.