SpaDE: Semantic Locality Preserving Biclustering for Neuroimaging Data

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Machine Learning

Background:

• Neuroimaging studies often miss subtle patterns in smaller subgroups, reducing specificity in neuropsychiatric conditions.
• Subject heterogeneity challenges traditional clustering methods in high-dimensional neuroimaging data.
• Existing biclustering methods face difficulties with sparse data and interpreting attribute groupings.

Purpose of the Study:

• To introduce a deep neural network, SpaDE, for unsupervised feature learning and biclustering in neuroimaging.
• To enhance neurobiological interpretability by preserving semantic locality in subject and feature subgroups.
• To regularize for sparsity for improved representation learning in brain connectome data.

Main Methods:

• Developed a deep neural network, semantic locality preserving auto decoder (SpaDE), for biclustering.
• Employed SpaDE on human brain connectome data from schizophrenia (SZ) and healthy control (HC) subjects.
• Compared SpaDE against state-of-the-art biclustering methods.

Main Results:

• SpaDE successfully identified coherent subgroups of subjects and neural features, outperforming existing methods.
• The model extracted modular neural communities exhibiting significant group differences between HC and SZ subjects.
• Bi-clustered connectivity substructures showed strong correlations with cognitive measures like attention and working memory.

Conclusions:

• SpaDE offers a powerful tool for unsupervised learning and biclustering in neuroimaging, enhancing interpretability.
• The method reveals specific brain network alterations in schizophrenia linked to cognitive deficits.
• SpaDE advances the analysis of complex neurobiological data and its relation to cognition.