Applying Regionalized Tessellation to Detect Diagnostic Markers of ASD in Resting EEG Data

Thursday, May 11, 2017: 5:30 PM-7:00 PM
Golden Gate Ballroom (Marriott Marquis Hotel)
A. Atyabi1,2, T. McAllister3, S. Hasselmo4, S. A. A. Chang5, M. J. Rolison4, T. Halligan4, B. Lewis4, T. C. Day4, K. A. McNaughton4, K. S. Ellison4, J. Wolf6, K. Stinson7, S. M. Malak8, J. A. Trapani8, E. Jarzabek8, J. McPartland8 and A. Naples4, (1)Pediatrics, University of Washington, Seattle, WA, (2)Seattle Children's Innovation & technology Lab, Seattle Children's Research Institute, Seattle, WA, (3)Child Study Center, Yale University School of Medicince, New Haven, CT, (4)Child Study Center, Yale University School of Medicine, New Haven, CT, (5)Yale University, New Haven, CT, (6)Yale Child Study Center, New Haven, CT, (7)Yale University- Child Study Center, Milford, CT, (8)Child Study Center, Yale School of Medicine, New Haven, CT
Background: Atypical functional connectivity in electrophysiological data is widely reported in individuals with ASD. Existing methods for connectivity analysis include investigating: 1) electrode-to-electrode functional differences and 2) electrode-groupwise differences (cluster based permutation and mass univariate analysis). The first approach results in dense connectivity networks, from which it is difficult to identify regional differences that reliably distinguish diagnostic groups. The second approach relies on finding subsets of electrodes that distinguish groups across empirically simulated distributions. However, this approach, primarily applied to frequency and time-frequency representations of EEG data, is not generalizable to arbitrary features of the EEG, such as variations of phase, complexity, entropy and so on.

Objectives:  Our goal was to support identification of diagnostic markers of ASD by developing and applying a permutation-based connectivity approach that was both capable of incorporating arbitrary EEG features and robust to family-wise error.

Methods:  Analyses were conducted on two minutes of resting state EEG from participants with ASD (N=73;age=14.2) and TD (N=32;age=12.4). Data were acquired at 1000Hz using a 128-channel Hydrocel Geodesic Sensor Net. Electrodes were assigned to regions based on scalp location, and summarized using any chosen EEG feature. For each investigated (red) region, pair-wise tests on these summaries identify a set of (green) regions that statistically differ from both said region and all non-significant (black) regions. The process is repeated with larger regions until no significant differences are identified.

Results:  Our analyses focused on alpha power as the EEG feature of interest. Individuals with ASD showed reduced alpha correlation between right occipital and frontal electrodes, while TD individuals showed a pattern of reduced connectivity among frontal regions. This is shown by clusters of electrodes in individuals with ASD and TD that were significantly different from the region containing electrode E25 (marked as green cube in figures). In ASD, the trio of electrodes E90,E91, and E96 was found to be the main contributor to the observed regionalized significance, while in TD the trio of E62,E67, and E71 was the main source of the statistically significant difference. Two additional regions defined by single electrodes (E61 in ASD and E85 in TD) were also significantly different from the region containing electrode E25.

Conclusions:  Our preliminary results support prior findings of reduced long-range but increased short-range connectivity in ASD, and show the utility of the regionalized tessellation methodology for automatically identifying scalp regions that distinguish between diagnostic groups. The strength of this approach lies in its ability to utilize a variety of arbitrary features of the EEG in our analyses. Our ongoing analyses focus on metrics of EEG complexity at the regional and electrode level. These analyses show that regionalized tessellation is capable of identifying established patterns of brain activity that vary between groups, and holds promise for novel investigations of large, high-dimensional data sets.