Using Information-Based Functional Brain Mapping to Detect Biomarkers of Autism In Adults

Background:  Highlighting objective biomarkers at the neural level are of paramount importance in autism research. In addition to providing insight into the mechanisms contributing to the clinical phenotype, from a practical standpoint, they may also aid in the diagnostic process or be used for evaluating treatment outcome. Machine learning techniques applied to massively multivariate biological datasets have particularly important utility here as they may be able highlight such biomarkers.

Objectives:  Here we investigate whether individuals can be classified according to diagnostic status solely based on information present in fMRI data when individuals are engaged in self-reflective thought. We use a support vector machine (SVM) and an information-sampling technique known as ‘searchlight’ mapping to hone in on specific brain regions whose fMRI response patterns contain information for making a distinction between those with and without a diagnosis.

Methods:  Twenty eight adult males (18-45 years old) with an ADI-R confirmed diagnosis of an autism spectrum condition (ASC) and 28 age-, sex-, and IQ-matched Control adults were scanned with fMRI at 3T while making mentalizing or physical judgments about themselves or a non-close other (i.e. the British Queen). A linear SVM was applied over 8mm radius spherical ‘searchlights’ centered on every voxel in the brain. Classification accuracy, sensitivity, and specificity was obtained with 28-fold cross validation (i.e. leave-one-subject-pair-out) on the unsmoothed Self>Other t-maps. P-values were assigned to each searchlight after simulating the null hypothesis distribution of cross validated classification accuracies using permutation testing (1,000,001 iterations). FDR control was used for thresholding at the whole-brain level and within an a priori ROI in ventromedial prefrontal cortex constructed from a quantitative meta-analysis of normative studies reporting a Self>Other contrast. To investigate the utility of local multivariate information-sampling over other approaches, searchlight mapping results were compared to a whole-brain SVM and classification based on univariate ‘activation’ data (i.e. averaged t-values within the searchlight).

Results:  Whole-brain SVM yielded classification accuracy no better than chance (50%). With searchlight mapping, no searchlights survived stringent whole-brain FDR correction. However, within the vMPFC ROI known to be sensitive to self-relevant information processing, one cluster survived FDR small-volume correction. The peak searchlight within vMPFC was 87.5% accurate in predicting diagnostic status and gave 92.85% sensitivity and 82.14% specificity. This classification accuracy was very improbable to have occurred at chance (p = 6 x 10^-6). This kind of multivariate classification outperformed univariate SVM classification on the same searchlight (66.07% accuracy), demonstrating that the added information from multivariate response patterns aid classification over and above univariate approaches.

Conclusions:  These results are a proof of concept that local pattern-information from vMPFC during self-referential thought can classify ASC from Controls with high accuracy, sensitivity, and specificity. In addition to providing a potential biomarker for ASC in male adults, these results also demonstrate the utility for using local information-based functional brain mapping rather than multivariate whole-brain classification or massively univariate activation-mapping in autism neuroimaging research.