Diagnostic Classification for Autism in Male Adults Based On Resting State fMRI Fractal Connectivity

Background:    Much emphasis in autism research has been placed on systems-level atypical brain connectivity. Almost all published work thus far has measured ‘functional connectivity’ via Pearson’s correlation coefficient or independent components analysis over a specific low-frequency band. Both do not take into account the long memory fractal properties that are inherent within BOLD fMRI time-series data. New measures are now available that account for such fractal properties, but it is an open question as to whether such measures of ‘fractal’ connectivity are more sensitive than more conventional measures of ‘functional’ connectivity. Here, we investigate this question via testing the predictive utility of ‘fractal’ versus conventional measures of functional connectivity for making diagnostic predictions on adult males with and without autism.

Objectives:   To assess the predictive utility of patterns of resting state ‘fractal’ connectivity and conventional measures of functional connectivity for diagnostic classification of adult males with and without autism.

Methods:   Thirty neurotypical and thirty high-functioning adult males with a diagnosis of an autism spectrum condition (ASC), matched on age (18-45) and IQ, were scanned with fMRI at 3T (TR=1302 ms; 512 whole-brain volumes) while being instructed to stay awake and keep their eyes closed. BOLD time-series were extracted from 110 regions from a standard anatomy-based brain atlas and the mean time-series from each region was decomposed with a maximum overlap discrete wavelet transform. Conventional measures of ‘functional’ connectivity were computed using Pearson’s correlation on wavelet scales 2 (0.096-0.192 Hz), 3 (0.048-0.096 Hz) and 4 (0.024-0.048 Hz). Long memory ‘fractal’ connectivity measures were estimated via asymptotic wavelet correlation matrices that measure convergence of the wavelet correlation spectrum on an asymptotic value across a range of low frequency scales (Achard, Bassett, Meyer-Lindenberg, & Bullmore, 2008, Phys Rev E). Classification analyses were performed with a linear support vector machine and a leave-one-subject-pair-out cross validation scheme (k=30). Classification performance measures were evaluated for statistical significance with permutation tests (10,000 iterations).

Results:   Whole-brain fractal connectivity matrices provided sufficient pattern-information for above chance classification of diagnostic status (Accuracy=81.67%, Sensitivity=80%; Specificity=83.33%; PPV=82.76; NPV=80.65%; all p<9.99x10^-5).  This performance stands in stark contrast to conventional measures of functional connectivity measured by Pearson correlation matrices across each wavelet scale (all Accuracy<73%).

Conclusions:   This work demonstrates enhanced sensitivity for ‘fractal’ connectivity measures for characterizing atypical connectivity in autism compared to conventional correlation measures that are currently in wide use (Pearson’s correlation). ‘Fractal’ connectivity measures diverge from conventional measures of functional connectivity by taking into account the long memory fractal properties of BOLD time-series data. One unresolved question for future work lies in understanding the contributions of neural and non-neural sources influencing fractal behavior in BOLD time-series data.  Finally, this work observes levels of classification performance that are statistically significant for deeming important for hypothesis testing, but are not high enough to be useful in a clinical setting or in settings where the base rates of ASC approach the population prevalence. Increases in sample size as well as use of other methods for whole-brain parcellation will also be important for future work.