Abstract: Background 
In previous studies we have shown the ability of a novel method of machine learning system named MS-ROM/IFAST to extract interesting features in digital EEG with standard 19 electrodes montage that allow very good distinction of ASD children from those who are developing typically and from those affected by other neuro-psychiatric disorders. If this signature is already present already at birth, then a screening program could be afforded in general hospitals registering EEG signals in the newborn. Since the equipment routinely available in neonatology units employ often few channels (2-8 electrodes), we were curious to check if features extracted from just two channels were enough to allow a good diagnostic performance in the same cases of the above-mentioned studies. 
Aim 
The aim of this study is to evaluate the information load present in just two EEG channels to distinguish autistic subjects from typically developing ones and from those affected by other neuro-psychiatric disorders. 
Methods 
C3 and C4 time-series were isolated from EEG data sets used in two previous studies, the first carried out in Italy on 25 subjects (15 ASD and 10 typicals) and the second carried out in US on 40 subjects (20 ASD and 20 with other neuropsychiatric disorders). 
A continuous segment of artifact-free EEG data lasting 10 minutes in ASCCI format was used to compute multi-scale entropy values and for subsequent analyses. 
A Multi-scale ranked organizing map (MS-ROM), based on the self-organizing map (SOM) neural network, coupled with the TWIST system (an evolutionary system able to select predictive features) created an invariant features vector input of EEG on which supervised machine learning systems acted as blind classifiers. 
Results 
After MS-ROM/I-FAST preprocessing, ninety features were extracted from C3-C4 timeseries of study 1 and of study 2 representing the EEG signature. Acting on these features the overall predictive capability of different machine learning systems in deciphering autistic cases from typicals (study 1) and from other NP disorders (study 2) ranged between 93% and 94. % (study 1) and from 80 and 88% (study 2) These results were obtained at different times in separate experiments performed on the same training and testing subsets. 
Conclusion 
The results of this study suggest that also a minor part of EEG contain a precious information useful to detect autism if treated with advanced computational algorithms. This could allow in the future to use standard EEG from newborn to check if ASD signature is already present at birth.

Notes: 

(1) Autism Research Unit, Villa Santa Maria Foundation, Tavernerio, Italy
(2) Semeion Research Centre, Roma, Italy
(3) Tarnow Center for Self-Management, Huston, TX