Semi-supervised clustering for neuro-subtyping of autism spectrum disorder

Abstract

Autism spectrum disorder (ASD) is a complex neurodevelopmental disorder with heterogeneous symptoms and neural characteristics, which has hindered the finding of reliable biomarkers. Until recently, ASD subtyping using neuroimaging has emerged to detect underlying neural features in each subgroup. However, not only the existing findings are preliminary, but also the methodologies are diverse with no conclusions on which are superior. In this study, we implemented neurosubtyping of ASD using a semi-supervised clustering method, HeterogeneitY through DiscRiminative Analysis (HYDRA), together with multi-scale dimension reduction of high-dimensional input features. We first performed orthogonal projective non-negative matrix factorization to objectively search down-sampled feature dimension, then utilized labelling information of ASD/controls to guide the clustering of ASD subjects so that the subtyping can consider the intrinsic disease-specific differentiations. Subjects were from a large ASD repository ABIDE. Functional connectivity was chosen as neural features for subtyping. Systematic evaluation of clustering performance, estimated by validity and reliability indicies, was conducted through a three-layer approach with random-stratification in the inner layer. We further compared HYDRA with unsupervised k-means, commonly-used for neuro-subtyping of mental disorders. Results showed that HYDRA with dimension reduction successfully detected two distinct clusters with high reliability. Both semi-supervised and unsupervised clustering have discovered similar connectivity patterns under each cluster. Hypo-connectivity in default mode network, hyper-connectivity in visual recognition, subcortical and cerebellum networks were observed in one cluster; and the opposite patterns were in another. Yet HYDRA produced more distinct patterns across clusters compared to kmeans. Statistical analyses suggest that neural connectivity patterns were associated with core ASD symptoms differently across each cluster. Our findings show that semi-supervised clustering is promising to subtype ASD into reliable and distinct clusters that can dissociate the heterogeneous neural markers and neural-behavioral relationships, which is one step forward for individual-based diagnosis and treatment in the future.

Applications of hypergraph-based methods in classifying and subtyping psychiatric disorders: a survey

“…These capabilities have aided in addressing the heterogeneity of psychiatric diseases and advancing the development of precision medicine in psychiatry. Many studies have demonstrated the existence of consistent functional connectivity patterns across subgroups, thus providing further support for the subdivision of psychiatric disorders into specific subtypes [34][35][36][37][38][39][40][41]. Additionally, graph theory approaches have been used to detect transdiagnostic biotypes of psychiatric disorders.…”