Modeling the Connections of BRAIN Regions in Children with Autism Using Evolutionary Algorithms and Electroencephalography Analysis.

Poster Presentation
Saturday, May 4, 2019: 11:30 AM-1:30 PM
Room: 710 (Palais des congres de Montreal)
E. Grossi, Autism Research Unit, Villa Santa Maria Foundation, Tavernerio, Italy
Background: Recent studies with neuroimaging methods like diffusion tensor imaging, functional connectivity, and graph theoretic methods have showed atypical development of neural connectivity in ASD, with excessive local connectivity within neural assemblies and deficits in long-range connectivity between functional brain regions. In this paper we present a new pre-processing approach of EEG data based on a novel algorithm applied to raw data and to quantitative EEG features able to pick-up abnormal connections.

Objectives: The aim of this study is to focus brain connections abnormalities in ASD using novel algorithms applied to EEG data.

Methods: Twenty children diagnosed with ASD (DSM-V criteria) and 20 children diagnosed with NPD (ADHD –N.16, mood disorders –N.2, anxiety disorders –N. 2) matched identically for age and male/female ratio, were entered into the study. A continuous segment of artifact-free EEG data lasting 10 minutes in ASCCI format were entered in Cin-Cin algorithm, a new pre-processing method to treat multichannel time series related to brain activity. The algorithm is based on an input vector characterized by a linear composition of city-block matrix distances between 19 electrodes. In this way, each EEG is transformed in a vector of 171 numbers expressing all the one by one distances among the 19 electrodes. Each distance value is assumed to express the connection among the two brain areas below corresponding electrodes. An evolutionary algorithm (a TWIST system based on KNN algorithm) was used to subdivide the dataset into training and testing set and select connections yielding the maximal amount of information. After this pre-processing different machine learning systems were used to develop a predictive model based on a training testing crossover procedure applied to selected connections distances.

Results: The connections subset involving 11 electrodes with nine connections (T4_F3, O2_F4, P3_T3, P3_C3, O1_C4, P3_T5, P4_T5, O1_T5, O1_P3) allowed the maximum degree of predictive performance by Machine Learning Systems used as classificators. Four of these connections are long range (three inter hemispherical and one intrahemispherical) and five short range. Long-range mean distances values resulted higher in ASD group while the opposite was true for short-range distances. The best machine learning system (three-layer feed- forward neural network with 8 hidden nodes) obtained a global accuracy of 96.2% (96.4 % sensitivity and 96.0 % specificity) in differentiating ASD subjects from NPD subjects.

Conclusions: The results of this study indicate the existence of brain connections abnormalities in ASD detected with evolutionary algorithms and Electroencephalography Analysis applied on a linear composition of city-block matrix distances between 19 electrodes. In addition, the model could distinguish the autistic children from the control children with an accuracy rate of 96.2%.