EEG-IP: An International Infant EEG Data Integration Platform for the Study of Risk and Resilience in Autism and Related Conditions

Background: EEG is a widely available, non-intrusive and powerful method for capturing aspects of brain behaviour. There has been an increased interest in using EEG to identify neurological markers of ASD that precede the onset of behavioral symptoms. However, EEG lacks standards in acquisition and processing, and is prone to noise contamination (Movement, EMG, EOG, ECG), particularly when working with populations with neurodevelopmental disorders and infants. Additionally, while infant samples are important in studying early risk markers of autism, most studies remain limited by small sample size.

Objectives: We established a new International infant EEG data-integration platform for the study of risk and resilience in autism (EEG-IP). The platform includes three components: (1) A data repository of >4,000 recordings on approximately 700 infants at risk and controls collected across four laboratories; (2) A generalizable and highly optimized signal extraction pipeline that maximizes the isolation of signal and noise in large-scale EEG data recordings; (3) A signal processing toolbox that optimizes the execution of analytic techniques for infant data.

Methods: Data integration of EEG recording is currently underway. In parallel, we validated the new analytic pipeline by exploring whether or not it can isolate the cortical signal to identify clearly event-related potentials (ERPs) and a reliable ICA decomposition at the individual infant level. 75 EEG recordings of 7-month old infants were drawn from EEG-IP. The visual stimulus procedure included a stimulus stream of human faces and non-face stimuli (phase scrambled version of the face stimuli). The recordings were preprocessed using the new Lossless pipeline that uses a sequence of robust measures to isolate spatially non-stationary channels and periods of time in the recording. This pipeline also performs the Adaptive Mixture Independent Component Analysis (AMICA) to isolate noise factors from cortical signal, as well as isolate the activation of specific cortical generators.

Results: Many of the procedures employed by the Lossless pipeline are computationally intensive and typical EEG pre-processing for ICA requires substantial manual interaction with the data. By optimizing this procedure for High Performance Computing (HPC) resources, and streamlining manual interaction to single annotation procedure, a sample of this size can be processed in days rather than weeks or months. From the output state of the processing pipeline we were able to extract pronounced ERPs emanating from various occipital regions of the cortex. Unlike more typical preprocessing pipelines, the output file contains all of the original data (nothing is removed that cannot be reinserted) so this signal isolation process does not restrict post-processing options.

Conclusions: Infants at risk for ASD are a traditionally difficult population to recruit, and the EEG-IP provides a sufficiently powered sample size for hypothesis testing at an unprecedented scale. However, this rests on our ability to extract valid individual-level data. The Lossless pipeline produces a state of EEG data with annotations regarding several fine-grained measures of signal quality and ICA decompositions, allowing researchers to perform complex analyses on the data, with the goal of identifying neurological differences between infants that do and don’t develop ASD.