Identifying Boys With Autism Spectrum Disorder Based on Whole-Brain Resting-State Interregional Functional Connections Using a Boruta-Based Support Vector Machine Approach

Abstract

An increasing number of resting-state functional magnetic resonance neuroimaging (R-fMRI) studies have used functional connections as discriminative features for machine learning to identify patients with brain diseases. However, it remains unclear which functional connections could serve as highly discriminative features to realize the classification of autism spectrum disorder (ASD). The aim of this study was to find ASD-related functional connectivity patterns and examine whether these patterns had the potential to provide neuroimaging-based information to clinically assist with the diagnosis of ASD by means of machine learning. We investigated the whole-brain interregional functional connections derived from R-fMRI. Data were acquired from 48 boys with ASD and 50 typically developing age-matched controls at NYU Langone Medical Center from the publicly available Autism Brain Imaging Data Exchange I (ABIDE I) dataset; the ASD-related functional connections identified by the Boruta algorithm were used as the features of support vector machine (SVM) to distinguish patients with ASD from typically developing controls (TDC); a permutation test was performed to assess the classification performance. Approximately, 92.9% of participants were correctly classified by a combined SVM and leave-one-out cross-validation (LOOCV) approach, wherein 95.8% of patients with ASD were correctly identified. The default mode network (DMN) exhibited a relatively high network degree and discriminative power. Eight important brain regions showed a high discriminative power, including the posterior cingulate cortex (PCC) and the ventrolateral prefrontal cortex (vlPFC). Significant correlations were found between the classification scores of several functional connections and ASD symptoms (p < 0.05). This study highlights the important role of DMN in ASD identification. Interregional functional connections might provide useful information for the clinical diagnosis of ASD.

Evaluation of Risk of Bias in Neuroimaging-Based Artificial Intelligence Models for Psychiatric Diagnosis

“…A total of 517 studies presenting 555 development-purpose AI models were eligible for inclusion in the analysis…”

“…All models were rated for poor reporting quality in the technical assessment domain across 4 signaling questions. Of the 555 models, 544 (98.0%; 95% CI, 96.9%-99.2%) lacked reports for how to evaluate AI performance publicly …”

“…The study protocol was registered at PROSPERO (CRD42022340624) and followed the Preferred Reporting Items for Systematic Reviews and Meta-analyses (PRISMA) reporting guideline (Figure 1…”

“… a Of the 555 artificial intelligence (AI) models, 469 were tested at the internal sample and 86 at the independent external sample, including intellectual disability, autism spectrum disorder, attention-deficit/hyperactivity disorder,…”

Classification of Alzheimer's disease: application of a transfer learning deep Q‐network method

“…The number of times that the accuracy was greater than or equal to the original test set accuracy was counted and divided by the total number of permutations. If the p ‐value is less than 0.05, the original accuracy is considered significant (Zhao et al, 2022). The specific process of the transfer learning algorithm is shown in Figure 1.…”