Accurately diagnosing neurological disorders through non-invasive, clinically accessible methods is a broad challenge facing researchers today. Specifically, the approach has been to extract features to study the functionality of the brain as a whole, rather than trying to identify changes in structure due to physical degradation or injury. For example, the use of cerebrospinal fluid and neuroimaging biomarkers to diagnose neurological disease are expensive, invasive, and difficult to implement in a clinical environment. With EEG data being cheap, widely available, and non-invasive, it is an attractive resource for identifying biomarkers for early diagnosis, disease tracking, and therapy monitoring of patients with neurological conditions.
Current research in this field is aimed at comparing how topographic EEG markers compare between healthy controls and a particular patient population. However, the clinical usefulness of these findings as a diagnostic tool of a specific disease can be assessed by ruling out abnormalities that are common to multiple disorders. Specifically, the abnormalities detected must be able to distinguish one disease from another. In addition to a control data set, we were provided with a test data set containing patients in a depressed level of consciousness (DLOC), also known as a coma, as a result of acute severe brain injury. This data set contained patients that may have suffered from diffuse brain injuries or from focal injuries that affect cerebral function in wider scale. Due to ambiguity of the exact diagnoses of each of the patients in the test set, the objective of our project was not to identify biological markers for use as a potential diagnostic feature, rather to generate networks using these biological markers and identify abnormalities and network dynamics common to coma patients.
The functional organization of the brain, consisting of interconnected local networks of neurons, enables it to be studied as a dynamical system. Studying high-level interactions in the brain at a macroscopic scale overlooks the microscopic interactions at the neuronal level to make meaningful conclusions regarding the dynamics among spatial regions of the brain. As a system, the brain encompasses many stable and metastable states, which enables it to self-organize by modulating its internal structure (Friston, 2000). As a result of neurological disease, these modulations that reorganize the internal structure of the brain can be interpreted as changes in the network dynamics of the brain. These network dynamics can be inferred by measuring the functional connectivity between different regions of the brain. Modern imaging techniques, such as EEG, can provide macroscopic data that enables the observation of time-dependent electrical activity. As such, the statistical dependencies between EEG signals can be quantified to measure functional connectivity. However, the variability of EEG signals and the incomplete body of research on neuronal processes limits the quality of our analysis. Characterizing dynamic patterns is complicated by the intrinsic complexity and nonlinearity of underlying neuronal processes (Abarbanel and Rabinovich, 2001). We primarily focused on linear connectivity measures which are simple to calculate and sufficient in detecting signal coupling although they can be insensitive to nonlinear coupling when compared to nonlinear connectivity measures (David et. al., 2003).
The primary objective of this project was to study the effects of traumatic brain injury on network interactions within the brain as well as their time-evolution characteristics in distinct spatial regions. The EEG data contained a set of 10 healthy, control patients, and 10 test patients in a coma state who had suffered traumatic brain injury. Our data driven approach in network construction aimed to identify changes in brain network dynamics due to traumatic brain injury. The following list identifies the individual objectives in this project:
- Identify abnormalities and characterize dynamics of brain network interactions for control and coma patients.
- Develop a data-driven, graph-based model of the brain to make global observations about the changes in spatial dynamics as a result of traumatic brain injury.
- Utilize existing data sets in selecting sampling parameters to optimize model selection.
- Evaluation of traditional and new methods used in the field for time series analysis
- Explore existing techniques long utilized by research.
- Explore emerging techniques currently being explored by research.
The clinical usefulness of EEG data relies on the identification of biomarkers of disease as a potential diagnostic tool. To explore the potential of using EEG data as a diagnostic tool, two methods of EEG data analysis were proposed by this project: network analysis and time-series analysis. First, network analysis was conducted to compare the spatial interactions in the brain among a control group and test group which contained subjects in a depressed state of consciousness (DLOC). The brain was modeled using an unweighted, undirected graph with nodes corresponding to EEG electrode positions and binary edges determined by calculating the cross-correlation of signals collected from each pair of electrodes and comparing this value to a threshold. The use of multiple thresholds provided an additional dimension with which to compare network interactions since the graphs were constrained to be binary and unweighted for simplicity. In the context of network analysis, potential classification features were identified by calculating network metrics after generating graphs at various threshold values. The key features that were investigated in network analysis were the percentage and frequency of hubs and islands, average number of edges per node, characteristic path length, and average clustering coefficient. Comparisons were drawn between the control and test sets using a combination of these network statistics and spatial connectivity patterns. The calculated network metrics were assessed in the context of the spatial observations made using head-maps generated for each subject and threshold value. In the second part of this project, different time-series techniques used in current EEG research were explored, namely, threshold analysis, autocorrelation analysis, wavelet analysis. After processing the data using these techniques, observations were drawn regarding the distinctions highlighted between the control and experimental groups. Identifying motifs or patterns in the time-evolution of EEG signals unique to a specific patient population can serve as potential biomarkers of disease.
Through a primarily data-driven approach, we assessed different sets of sampling parameters before analyzing the data as a whole. Specifically, the data selection process was characterized by assessing different combinations of parameters, namely, such epoch length, number of epochs sampled, and value of the edge threshold. After producing multiple graphs for each patient, their relative sparseness and connectivity was compared in order to select a moderate model to generate a global graph for both the control and test data set. Through comparison of the global graphs and the global control graph to each of the test patients’ graphs, we observed qualitative and statistical differences, revealing how network dynamics are affected by neurological disease.
We explored three methods used in research in time series analysis. The first was a time threshold analysis. Time threshold analysis effectively allows researchers to filter through data and highlight trends in data points that represent stronger signals. The second method was autocorrelation analysis. Autocorrelation is a well-known signal analysis technique that allows a researcher to understand how correlated data is to itself. The third method that we explored was wavelet analysis. Wavelet analysis was developed in response to a growing need to analyze both time and frequency trends in EEG data. We used the EEG data provided to us by Dr. Ching’s lab to evaluate these methods.