Final Report: Improved Discrimination of Asperger Patients using fMRI and Machine Learning

Asperger Syndrome is an autism spectrum disorder that reduces a patient’s ability to interact socially, and restricts their interests and abilities. Using functional magnetic resonance imaging (fMRI), it has been shown that Asperger patients exhibit reduced activity between the nodes of a resting-state neuronal network comprised of the posterior cingulate cortex (PCC), the medial prefrontal cortex (MPFC), and the lateral parietal cortex, compared to healthy controls [1]. The study of Ref. [1] proposed a technique to discriminate Asperger patients and healthy controls using the self-organizing map (SOM) algorithm [2] to automatically generate a cluster representing this resting-state network from restingstate fMRI data. If robust, such a method could have important applications in screening for this disorder, both in clinical and research settings. However, this method has a few potential weaknesses. First, it performs discrimination using a single statistic. Second, the SOM algorithm clusters voxel timecourses based on Euclidean distances, and therefore may not robustly cluster voxels that are functionally connected but possess inter-voxel delays. Third, the original method required user interaction to choose the cluster that best represents the resting-state network, which may bias the results. Therefore, we are proposing three innovations to this algorithm. To address the first and last issues, we will apply filter and wrapper methods directly to the fMRI images, bypassing SOM and the need for user interaction. We have also implemented several different discrimination methods to compare to the simple z-score threshold test implemented in