30007
Machine Learning Methods Reveal Improved EEG Biomarkers in an Autism Clinical Trial
Clinical trials for autism spectrum disorder (ASD) frequently utilize electroencephalography (EEG) measurements to track and evaluate neural dynamics. The utility of these measurements is determined by the efficacy of the biomarkers (or features) extracted from the EEG. Novel methods to learn improved biomarkers from clinical trial data can increase this efficacy and statistical power. A particular challenge for machine learning is the “little big data” structure, where there are many EEG samples but only a small number of participants. We introduce a new framework to address this challenging structure in conjunction with an interpretable deep network.
Objectives:
Our objective is to evaluate whether there are significant changes in neural activity in an open-label clinical trial designed to evaluate the efficacy of a single infusion of autologous cord blood for ASD. Specifically, the goal is to uncover improved EEG biomarkers using a machine learning algorithm to capture neural differences based on EEG recordings taken while watching social and non-social stimuli in longitudinal data.
Methods: The study involves secondary analysis on 22 children with ASD from ages 3 to 7 years who participated in this clinical trial and had available EEG recordings. High-density EEG recording (EGI, Inc) were collected at baseline (T1), 6 months post treatment (T2), and 12 months post treatment (T3). The EEG was recorded while watching a total of three one-minute long videos designed to measure responses to dynamic social and nonsocial stimuli. To learn EEG biomarkers, an interpretable convolutional network was applied to clean one-second data chunks to predict whether an EEG sample was from T1, T2, or T3. The “little big data” structure of the recordings was addressed with a novel multi-domain adaptation approach by explicitly learning and using similarities between patients. By examining the features, we can track how neural changes correlate to treatment stages. The learned features were evaluated with a leave-one-patient-out cross-validation scheme.
Results:
Our learned features distinguished between trial stages at a significantly higher rate than traditional approaches. Specifically, our features determined the correct stage 67.8% of the time. We compared this to a number of existing methods, with the next best performance giving 58.4% accuracy from a Multi-Channel Deep Convolutional Neural Network, suggesting that our novel methodology captures improved EEG information (p<.001, Wilcoxon signed-rank test). One confound on predictive ability was age, but age only captured a small amount of variability. The EEG was more predictive for trial stage compared to using age alone (p<.001, paired t-test). Visualization of learned features showed that particular frequency bands and sections of the brain were important to distinguish among trial stages. Neural similarities of participants can be revealed by using a graph visualization.
Conclusions:
Our machine learning methods empirically learned improved biomarkers related to clinical trial outcomes. By enhancing discrimination by a non-trivial margin, the statistical power of EEG biomarkers was significantly enhanced. The learned EEG biomarkers are a powerful tool to identify changed neural patterns in a longitudinal study, which we will soon be evaluating in a larger, placebo-controlled study.