A meta-analysis of machine learning classification tools using rs-fmri data for autism spectrum disorder diagnosis

Abstract

The Autism Spectrum Disorder (ASD) is a complex and heterogeneous neurodevelopmental condition characterized by cognitive, behavioral, and social dysfunction. Much effort is being made to identify brain imaging biomarkers and develop tools that could facilitate its diagnosis - currently based on behavioral criteria through a lengthy and timeconsuming process. In particular, the use of Machine Learning (ML) classifiers based on resting-state functional Magnetic Resonance Imaging (rs-fMRI) data is promising, but there is an ongoing need for further research on their accuracy. Therefore, we conducted a systematic review and meta-analysis to summarize and aggregate the available evidence in the literature so far. The systematic search resulted in the selection of 93 articles, whose data were extracted and analyzed through the systematic review. A bivariate randomeffects meta-analytic model was implemented to investigate the sensitivity and specificity across the 55 studies (132 independent samples) that offered sufficient information for a quantitative analysis. Our results indicated overall summary sensitivity and specificity estimates of 73.8% (95% CI: 71.8-75.8%) and 74.8% (95% CI: 72.3-77.1%), respectively, and Support Vector Machine (SVM) stood out as the most used classifier, presenting summary estimates above 76%. Studies with bigger samples tended to obtain worse accuracies, except in the subgroup analysis for Artificial Neural Network (ANN) classifiers. The use of other brain imaging or phenotypic data to complement rs-fMRI information seem to be promising, achieving specially higher sensitivities (p = 0.002) when compared to rs-fMRI data alone (84.7% - 95% CI: 78.5-89.4% - versus 72.8% - 95% CI: 70.6-74.8%). Lower values of sensitivity/specificity were found when the number of Regions of Interest (ROIs) increased. We also highlight the performance of the approaches using the Automated Anatomical Labelling atlas with 116 ROIs (AAL116). Regarding the features used to train the classifiers, we found better results using the Pearson Correlation (PC) Fishertransformed or other features in comparison to the use of the PC without modifications. Finally, our analysis showed AUC values between acceptable and excellent, but given the many limitations indicated in our study, further well-designed studies are warranted to extend the potential use of those classification algorithms to clinical settings.