Background: In the past years, a growing number of studies have addressed the predictive power of brain biomarkers through application of machine learning (ML) to brain data. Emerging literature reports a loss of brain asymmetry in autistic brains, even before the full-blown expression of symptoms, particularly in brain circuits involved in communication functions. It is not yet clear whether this lack of lateralization is present at a very early age and could be detected using machine learning applications on EEG data. ??

Objectives: The study objective is to test the hypothesis that applying artificial neural networks to data derived from symmetric couples of EEG electrodes in infants at high risk of autism at 12 months of age can predict later diagnosis at 24 months of age. Clinical variables were further included in the analysis to add predictive power.

??Methods: As part of a larger project (NET-2013-02355263-3, funded by Italian Ministry of Health, entitled Early Bird Diagnostic Protocol for ASD), a cohort of infants at risk for autism (presenting with early signs or siblings of older autistic children) has been studied longitudinally with EEG and clinical evaluations. Thirty-five infants (M: F=23:12) at risk for autism performed 12 minutes of HD-EEG registration (128 channels) at 12 months of age in a rest condition; 12 out of 35 infants received an autism diagnosis at 24 months of age. Nineteen electrodes commonly used in the 10-20 assembly were extrapolated for each subject; the linear correlation indexes and the Manhattan distances among the 8 homologous electrode pairs and the diagnoses at 24 months served as input for artificial neural networks (16 input variables and 2 outputs). ??

Results: Analysis of the average EEG distance values between the 8 electrode couples revealed a statistically significant difference between the two groups for FP1-FP2, T3-T4, and 01-02 correlation values, by the first year of life. The application of TWIST software allowed the isolation of 7 out of 16 variables (FP1-FP2 correlation; F3-F4 correlation; C3-C4 correlation; P3-P4 correlation; T5-T6 correlation; O1-O2 correlation; FP1-FP2_Manhattan) and ML classifiers applied to these variables reached a predictive performance of 90.44% accuracy (Table 1). In a second experiment, 13 clinical variables referring to the child's birth were added to the EEG variables. The application of TWIST software allowed the isolation of Gestational age, Cesarean Section, Birth weight, Father age, VABS Deviation IQ, which, in addition to EEG variables previously described, allowed ML to further increase the predictive performance, reaching 97.92% of accuracy. ??

Conclusions: The 90 % accuracy rate of predictive results obtained supports the idea that left and right hemispheres communicate atypically in infants at risk for autism who later develop autistic symptoms. This finding is in line with previous research reporting a reduction of brain asymmetry in autistic brains, especially in fronto-temporal circuits. The possibility to extrapolate this altered feature from the EEG application using only 19 electrodes paves the way to the possible application of this tool in at-risk populations of neonatology units, and opens to very early intervention strategies in infants at risk.

Notes: 

(1)IRCCS Stella Maris Foundation, Pisa, Italy, (2)Department of Clinical and Experimental Medicine, University of Pisa, Pisa, Italy, (3)Autism Research Unit, Villa Santa Maria SCS, Tavernerio, Como, Italy, (4)Information Technology, Bracco Imaging, MIlano, Lombardia, Italy