Functional network and spectral analysis of clinical EEG data to identify quantitative biomarkers and classify brain disorders
Matlis, Sean Eben Hill
MetadataShow full item record
Many cognitive and neurological disorders today, such as Autism Spectrum Disorders (ASD) and various forms of epilepsy such as infantile spasms (IS), manifest as changes in voltage activity recorded in scalp electroencephalograms (EEG). Diagnosis of brain disease often relies on the interpretation of complex EEG features through visual inspection by clinicians. Although clinically useful, such interpretation is subjective and suffers from poor inter-rater reliability, which affects clinical care through increased variability and uncertainty in diagnosis. In addition, such qualitative assessments are often binary, and do not parametrically measure characteristics of disease manifestations. Many cognitive disorders are grouped by similar behaviors, but may arise from distinct biological causes, possibly represented by subtle electrophysiological differences. To address this, quantitative analytical tools - such as functional network connectivity, frequency-domain, and time-domain features - are being developed and applied to clinically obtained EEG data to identify electrophysiological biomarkers. These biomarkers enhance a clinician’s ability to accurately diagnose, categorize, and select treatment for various neurological conditions. In the first study, we use spectral and functional network analysis of clinical EEG data recorded from a population of children to propose a cortical biomarker for autism. We first analyze a training set of age-matched (4–8 years) ASD and neurotypical children to develop hypotheses based on power spectral features and measures of functional network connectivity. From the training set of subjects, we derive the following hypotheses: 1) The ratio of the power of the posterior alpha rhythm (8–14 Hz) peak to the anterior alpha rhythm peak is significantly lower in ASD than control subjects. 2) The functional network density is lower in ASD subjects than control subjects. 3) A select group of edges provide a more sensitive and specific biomarker of ASD. We then test these hypotheses in a validation set of subjects and show that both the first and third hypotheses, but not the second, are validated. The validated features successfully classified the data with significant accuracy. These results provide a validated study for EEG biomarkers of ASD based on changes in brain rhythms and functional network characteristics. We next perform a follow-up study that utilizes the same group of ASD and neurotypical subjects, but focuses on differences between these two groups in the sleep state. Motivated by the results from the previous study, we utilize the previously validated biomarkers, including the alpha ratio and the subset of edges found to be a sensitive biomarker of ASD, and test their effectiveness in the sleep state. To complement these frequency domain features, we also investigate the efficacy of several time domain measures. This investigation did not lead to significant findings, which may have important implications for the differences between sleep and wake states in ASD, or perhaps generally for clinical assessment, as well as for the effect of noise on signal in clinically obtained data. Finally, we design a similar analysis framework to investigate a set of clinical EEG data recorded from a population of children with active infantile spasms (IS) (2-16 months), and age-matched neurotypical children, in both wake and sleep states. The goal of this analysis is to develop a quantitative biomarker from the EEG signal, which ultimately we will apply to predict the clinical outcome of children with IS. In addition to spectral and functional network analysis, we calculate time domain features previously found to correlate with seizures. We compare the two populations by each feature individually, test the effects of age on these features, use all features in a linear discriminant model to categorize IS versus neurotypical EEG, and test the findings using a leave-one-out validation test. We find almost every feature tested shows significant population differences between IS and control groups, and that taken together they serve as an effective classifier, with potential to be informative as to disease severity and long-term outcome. Furthermore, analysis of these features reveals two groups, indicating a possibility that these features reflect two distinct qualitative characteristics of IS and seizures.