Using Resting State Functional MRI to Build a Personalized Autism Diagnosis System

Background:

Autism spectrum disorder (ASD) is a neurodevelopmental disorder affecting approximately 1 in 68 people. Yet, there is no confirmed cause identified for ASD. Studying brain functional connectivity via resting state functional MRI (rs-fMRI) is one of the trending techniques used in understanding ASD.

Objectives: We tested the feasibility of using rs-fMRI data to build a computer-assisted diagnosis (CAD) system for ASD. From the clinical point of view, such a system could be of great value as it helps to predict and understand behavior of autistic individuals from very early, since rs-fMRI may be acquired even from infants. This is an important step towards advancing personalized medicine in autism, which is the ultimate goal of our group’s research efforts in this area.

Methods:

Resting state functional MRI data for 202 subjects were obtained from the National Database for Autism Research (NDAR). These were used to train and test a deep learning-based system to automatically detect rs-fMRI features potentially diagnostic for ASD. Power spectral density of time courses corresponding to the spatial activation areas were raw features input to a stacked autoencoder network. The higher-order representations produced by the network were used to build a classifier based on probabilistic support vector machines.

Results:

Hyperparameter optimization for the stacked autoencoder resulted in the selection of sparseness ρ = 0.3, sparse penalty β = 1.0, and weight regularization μ = 10⁻⁵. With these training parameters, classification accuracy was estimated via two-fold cross validation to be 0.856, with sensitivity of 0.984 and specificity of 0.798. Area under the ROC curve was 0.962. Using the ASD prevalence of the general population (1:68), the corresponding positive predictive value (PPV) is 0.44 and negative predictive value (NPV) is 0.998. In a high risk population where an older sibling has ASD, the prevalence is 18.7% (Pediatrics 128:e488, 2011), and the corresponding PPV is 0.925 whereas NPV is 0.982.

Conclusions:

This fMRI algorithm may be more informative in high-risk cases than in the general population. In a study of high-risk infants (Sci Transl Med 9:eaag2882, 2017), rs-fMRI networks at 6 months of age correctly predicted an ASD diagnosis at 24 months in 9 out of 11 who converted to ASD, and correctly predicted 48 of 48 who did not convert. Few rs-fMRI networks were correlated with social communication and cognitive ability in high-risk infants, while many more were correlated with repetitive behaviors. This suggests a developmental context since striatal and brainstem neural networks tend to mature earlier than cortically based networks. In the older population of the present study, we identified similar regions with altered connectivity previously noted in ASD including the pre-motor, supplementary motor, dorsal lateral, medial prefrontal, and sensorimotor cortex and regions involved in language. Previous rs-fMRI studies have noted that some of these regions, as parts of the Default Mode Network and interhemispheric connectivity networks, have reduced connectivity in ASD (Front Psychiatry 7:205, 2017). The data presented suggest the algorithms, especially when combined in a multimodal approach, have the potential to inform a diagnosis of ASD.