Cluster-aware machine learning of robust brain-behavior associations for precision neuropsychiatry.

1931 

Abstract Submission 

Amanda Buch¹, Conor Liston¹, Logan Grosenick¹

¹Weill Cornell Medicine, Cornell University, New York, NY

Amanda Buch  
Weill Cornell Medicine, Cornell University
New York, NY

Conor Liston  
Weill Cornell Medicine, Cornell University
New York, NY

Logan Grosenick  
Weill Cornell Medicine, Cornell University
New York, NY

Explainable machine learning of complex multimodal data in neuroscience research is revolutionizing precision neuropsychiatry. Interpretable clustering of patients into distinct subtypes can enhance personalized prognosis, diagnosis, and treatment. However, training on biomedical data poses challenges due to high dimensionality, clustering, and limited sample size. To address this, we propose a scalable approach for cluster-aware embedding, incorporating a convex clustering penalty. This approach facilitates hierarchical clustering of principal component analysis (PCA), locally linear embedding (LLE), and canonical correlation analysis (CCA). Our method improves upon existing techniques and offers a modular framework for interpretable biomarker discovery in precision medicine. We apply this approach to identify neurocognitive subtypes in the Adolescent Brain Cognitive Development (ABCD) and Autism Brain Imaging Data Exchange (ABIDE) datasets.

Clustering algorithms are often considered difficult optimization problems. However, by relaxing certain constraints, clustering algorithms can be reformulated as convex optimization problems by relaxing the hard clustering constraint ("convex clustering") [3]. Here, we introduce Pathwise Clustered Matrix Factorization (PCMF) and its multi-view extension, incorporating a convex clustering penalty to make embedding methods cluster-aware. PCMF does not require pre-determining the number of clusters and can generate a dendrogram for subtype discovery in biomedical datasets. It fits a path of solutions along a sequence of parameters and estimates split points based on model fit improvement. We extend our PCMF approach to include canonical correlation analysis (CCA) within clusters, introducing pathwise clustered canonical correlation analysis (P3CA). We apply P3CA to discover neurocognitive subtypes in the Adolescent Brain Cognitive Development (ABCD) [5] and Autism Brain Imaging Data Exchange (ABIDE) [6-7] datasets. In the first case study, we use resting state functional connectivity (RSFC) and behavioral (ADOS-2 and verbal IQ) data from N=299 individuals with autism spectrum disorder (ASD) in the ABIDE dataset. In the second case study, we use RSFC and behavioral (NIH Toolbox) data from N=490 adolescents in the ABCD dataset. We implemented a stringent preprocessing of both datasets following or exceeding well-established guidelines [8–9]. Prior to fitting P3CA to the ABIDE and ABCD datasets, we perform robust feature selection as previously described [2]. We evaluated the stability of neurocognitive subtypes by randomly holding out 30% of the data and recalculating P3CA.

Applying P3CA to the ABIDE dataset, we found strong differences in associations of ASD subtype embeddings with behavior and RSFC (Fig. 2), consistent with known autism subpopulation differences of behaviors with prefrontal cortex to somatosensory cortex, posterior parietal cortex, and middle temporal gyrus [2]. Subject-level P3CA embedding coefficients were robust to data perturbation (cosine similarity: 0.93 ± 0.05 for U estimates; 0.97 ± 0.03 for V estimates). Our method provides clearer cluster separation and improved interpretability. Applying P3CA to the ABCD dataset, we identified cluster-specific brain-behavior embeddings that described two neurocognitive phenotypes (separating high/low crystallized vs. fluid intelligence individuals), which were robust in replication in a second partition of the ABCD dataset.

Overall, our approach provides a flexible and effective method for clustering and subtype discovery in biomedical datasets with neuroimaging and behavioral data. We have introduced an interpretable joint clustering and embedding strategy using a modular convex clustering penalty that can be applied in multimodal datasets. We showcased our method in two case studies using the ABCD and ABIDE datasets, which revealed robust neurocognitive subtypes in both adolescents and individuals with autism spectrum disorder.

Neurodevelopmental/ Early Life (eg. ADHD, autism) ²

Connectivity (eg. functional, effective, structural)

Methods Development ¹

Multivariate Approaches

Autism

Cognition

Computational Neuroscience

Data analysis

Development

Machine Learning

Modeling

Statistical Methods

Other - neuropsychiatric subtypes