Identifying Neuropsychiatric Disorder Subtypes and Subtype-dependent Variation in Diagnostic Deep Learning Classifier Performance

Abstract

Clinicians and developers of deep learning-based neuroimaging clinical decision support systems (CDSS) need to know whether those systems will perform well for specific individuals. However, relatively few methods provide this capability. Identifying neuropsychiatric disorder subtypes for which CDSS may have varying performance could offer a solution. Dynamic functional network connectivity (dFNC) is often used to study disorders and develop neuroimaging classifiers. Unfortunately, few studies have identified neurological disorder subtypes using dFNC. In this study, we present a novel approach with which we identify 5 states of dFNC activity and 4 schizophrenia subtypes based on their time spent in each state. We also show how the performance of an explainable diagnostic deep learning classifier is subtype-dependent. We lastly examine how the dFNC features used by the classifier vary across subtypes. Our study provides a novel approach for subtyping disorders that (1) has implications for future scientific studies and (2) could lead to more reliable CDSS.

Neuropsychiatric Disorder Subtyping Via Clustered Deep Learning Classifier Explanations

“…These findings indicate that SZ seems to mainly affect the interactions of the VSN with the SCN, SMN, and ADN and of the CBN with the CCN and SCN to varying degrees. Many of these interactions were also identified in previous SZ subtyping efforts with the FBIRN dataset [17].…”

“…SPEC was about 5% higher than SENS, and ACC was near 80%. Our architecture obtained higher performance than some existing dFNC SZ classifiers [17][14].…”

“…In this study, we used the Functional Imaging Biomedical Informatics Research Network (FBIRN) rs-fMRI dataset. It has been used in multiple clustering [17][16] and classification studies [6] and can be made available upon a request to the authors. The dataset has 160 healthy controls (HCs) and 151 individuals with SZ (SZs).…”

“…Efforts involving SZ dFNC clustering have often been more focused on extracting novel insights [6], [7] or presented methods that could provide novel insights [15], [16]. However, dFNC clustering studies have generally not focused on SZ subtyping, and the small number of studies [17] that have focused on SZ subtyping have done so without really considering which brain networks most differentiate each subtype from SZ. It is possible that explainable deep learning approaches for SZ dFNC classification could be combined with clustering to identify novel SZ subtypes based on their differences with healthy individuals (HCs).…”

A Convolutional Autoencoder-based Explainable Clustering Approach for Resting-State EEG Analysis

“…This is partly due to the limited number of clustering approaches for lengthy EEG time-series and is problematic because clustering EEG offers key opportunities. Namely, clustering can identify novel disorder-related dynamics or subtypes, as has been shown in EEG [3] and functional magnetic resonance imaging (fMRI) [5][6] analysis. Moreover, most EEG clustering methods do not give insight into the features important to their clusters (i.e., they are not explainable).…”