Both the method of simultaneous electroencephalography – functional magnetic resonance imaging (EEG-fMRI) and the method of electroencephalographic source localization are opening up new understandings in neuroscience, but both are also technically very challenging. Simultaneous EEG-fMRI combines the strengths of each modality: the temporal resolution of electroencephalography (EEG) with the spatial resolution of functional magnetic resonance imaging (fMRI). EEG source localization provides a non-invasive means of locating in the brain the sources of electrical signals measured on the surface of the head. The potential for these methods is not limited to research neuroscience; each has much to offer in fields like clinical neurology and psychiatry. In clinical neurology, for example, both methods have been used in the location of epileptic foci. Simultaneous EEG-fMRI is technically very challenging as each modality serves as a source of artifact and safety concerns for the other. Two contaminants prominent in EEG acquired in the MRI environment are the gradient artifact (GA) and the ballistocardiogram artifact (BCG). The GA occurs only during an active MRI acquisition, while the BCG is always present, as its origins arise from the cardiovascular system of the subject. The current method to remove the GA from the EEG recording is called moving windowed average template subtraction (MWATS), the windows of which are time-locked to the repetition time of the MR acquisition. The problem with MWATS is that there are a number of options on how to implement it and little consensus or consistency in the field on how it is done. Current methods of removing the BCG artifact from the EEG recording all rely on establishing its timing based upon a simultaneously acquired electrocardiogram (ECG) signal. One issue with using the ECG signal is that the differential method for acquiring it implemented by the manufactures of the MR compatible EEG systems has a high rate of failure.

Source localization is technically very challenging in part because it relies on multiple measurements for its solution. Two of the measurements required for accurate source location are the measurement of the EEG electrode positions with respect to the underlying anatomy and the measurement of the underlying anatomy itself. At present, these measurements are made separately: the underlying anatomy is measured with an anatomical MRI scan, while the electrode positions are measured with either an electromagnetic digitization device or a photogrammetry system, neither of which is MR compatible. As such the two measurements must be registered to one another, typically through the location of fiducial markers on the surface of the head in both of the measurements. By aligning the markers, the two measurements can be brought into register. The issues with this technique are two fold. First, aligning the separate measurements introduces a potential source of error. Second, both of the current methods for measuring the electrode positions are very time consuming. This thesis focuses in robust solutions to these issues.

Specifically what follows is: 1) an analysis of the consequences of all the principal options for implementing the moving windowed average template subtraction; 2) a method to eliminate the need of the separately recorded ECG signal for establishing the timing of the BCG artifact; and 3) a method for automatically locating and identifying EEG electrodes in an anatomical MRI scan. The analysis of the MWATS method permits more informed decisions to be made in its implementation. The method to eliminate the need for the ECG signal is to both simplify the experimental setup and to make the EEG-fMRI recording more robust. Automatically locating and identifying EEG electrodes in an anatomical MRI scan eliminates both a measurement step and a registration step. The elimination of the registration step eliminates a potential source of error. In addition, the combination of the elimination of the measurement step and the automatic way in which the algorithm progresses greatly reduces the time required for this process.

I explored the capability of multivariate machine learning on low-signal and bioinformatic data to distinguish between different perceptual, emotional, and behavioral outcomes. In Chapter II, I designed and executed an experiment to evoke differences in empathy on the basis of religious identities as well as on general notions of ingroup and outgroup. I discovered that the act of simply learning about another’s religious identity correlated with a detectable modulation in one’s brain activity—most notably in regions related to empathy and sympathy—which identified subjects’ religious ingroups from outgroups with 70% accuracy. In Chapter III, I used multimodal data to distinguish between control participants and patients with one of two body-image disorders: anorexia nervosa and body dysmorphic disorder. These data included psychometric scales of obsessive-compulsive behaviors, depression, anxiety, and insight; white matter connectivity patterns; and neural haemodynamic activity. My models achieved 72% accuracy in distinguishing healthy controls from patients with anorexia nervosa or body dysmorphic disorder, and 76% accuracy in distinguishing patients with anorexia nervosa from those with body dysmorphic disorder. In Chapter IV, I extended these techniques to insurance claims data to predict the actions of patients and doctors. I discovered that the pattern of care delivered by doctors predicted the hospitalizations and use of biologic drugs in patients with inflammatory bowel disease. My diagnostic model may be used by healthcare alongside various interventions to reduce hospitalizations, decrease expenditures of insurance companies, and improve patient quality of life.

The original experiments presented in this manuscript began with the objective of utilizing steady-state visually evoked potentials (SSVEPs) as a tool for the study of visual attention and of attention disorders. The early outcomes of these experiments were not successful in reproducing effects that were reported to be quite robust by the relevant literature. This failure to replicate prior results precluded further interpretation of the data and launched a reexamination of the experimental design. Efforts to improve the application of SSVEPs specifically to the study of attention also led to a questioning of the assumptions about the properties of SSVEPs and to a thorough exploration into their sources of variability. This manuscript describes the development of precise methods for generating SSVEPs, presents the results of experiments investigating poorly understood SSVEP characteristics, and reveals sources of variability that can hinder SSVEP research. The outcomes of this project demonstrate that subtle stimulus parameters such as contrast and duty cycle can considerably impact SSVEP data. The culmination of this endeavor is a multimodal imaging study illustrating the relationship between evoked potentials across a range of frequencies as well as the distinctions between responses to periodic and aperiodic stimuli.

Biological processes often produce signals detectable by multiple means, but with entirely different information content. For both research and clinical applications concurrent multimodality data collection plays an important role in understanding the signal sources. Electroencephalography (EEG) and functional magnetic resonance imaging (fMRI) are particularly interesting examples, in that each offers largely independent yet complementary information on neuronal activity. While we and others have made great strides in making concurrent EEG-fMRI recordings possible, the EEG data, in particular, still contain signal artifacts of cardiac origin (ballistocardiogram) that make EEG analysis difficult, or even impossible.

To date, no satisfying means to separate brain EEG signal and ballistocardiogram (BCG) exist especially for non-event-related-potential experiments and under 3-T MR scanner. The BCG presents high temporal non-stationarity due to variation in cardiac cycles, and its amplitude scales with magnetic field strength. This explains the considerable variation of success levels among studies, with more successful applications achieved at lower field strength. Previously published methods used one of blind source separation methods to remove the BCG. All such blind source separation approaches are limited to performing component extraction based on the contaminated data alone, agnostic of the structural difference between BCG and EEG. Another kind of approach is to utilize reference signals for the artifact itself. However, this requires purpose-built hardware and exploits no further denoising step besides a simple subtraction.

We have developed three algorithms to separate EEG signal and BCG artifacts. Firstly, we have designed a Direct Recording - Prior Encoding (DRPE) method to maximally incorporate prior knowledge of BCG/EEG subspaces described by bases learned from a modified recording configuration, and of the group sparsity characteristics in the signal. To further promote subspace separability, a Direct Recording Joint Incoherent Basis (DRJIB) method is proposed to learn a representative and sparse set of BCG and EEG bases by minimizing a cost function consisting of group sparsity penalties for automatic dimension selection and an energy term for encouraging incoherence. Reconstruction is subsequently obtained by fitting the contaminated data to a generative model using the learned bases subject to regularization. The third algorithm takes advantage of currently available high-density EEG cap, to reliably estimate the full-scalp BCG contribution from a near-optimal small subset (20 out of 256) of channels and a corresponding weight through our modified experimental setup using Orthogonal Matching Pursuit (OMP).

We show in carefully constructed simulations that the residual artifacts are reduced by several orders of magnitude to a tiny fraction of the true signal. In human studies we show that the methods work effectively, and validate our quantitative results. Beyond the application to the EEG-fMRI challenge, we expect that our algorithmic methods will have impact in many other domains where signal and contaminant have distinct information structures. Digital signal, imaging, or even financial data are also contaminated with noises so that studying and characterizing the information structures of desired and undesired signals would greatly improve the modeling power.

White matter development in the human brain undergoes a uniquely extended developmental trajectory, and the maturation of this complex network of connections is broadly relatable to real-world measures of cognitive ability. We begin this dissertation with a review of the literature on structural brain development, in order to provide a useful background layer for our discussion (Chapter 1). Next, we report a case study investigating white matter abnormalities in the context of prenatal methamphetamine exposure, and their cognitive correlates (Chapter 2). Then we describe our successfully-funded NIH grant proposal to 1) develop novel image analysis techniques for investigating white matter connectivity, and 2) apply them to the study of a) typical frontal lobe white matter maturation, b) its relation to executive functioning, and c) how these processes are affected by prenatal alcohol exposure (Chapter 3). The results of this effort are described in Chapter 4, Chapter 5, and Chapter 6. Finally, we supplement the work on our UCLA cohort with exciting results from two large, multi-site, collaborative efforts: 1) ADHD-200, an initiative to employ brain mapping findings in a machine learning environment for the diagnostic classification of individual subjects with attention deficit hyperactivity disorder (ADHD) (Chapter 7), and 2) PING (the Pediatric Imaging, Neurocognition, and Genetics study), where we are leading an effort to provide the most comprehensive mapping to-date of typical white matter development (Chapter 8).

The focus of this graduate thesis is the development and optimization of clinically applicable computer-aided diagnostic tools (CADTs) for seizure disorder. This thesis is comprised of two parts (1) development of unimodal and multimodal CADTs for seizure disorder and (2) a novel method for optimization of hyperparameters in machine learning models. The aims of CADTs are to address key challenges in the diagnosis and treatment of seizure disorder, including reducing the time to an accurate diagnosis, improving the sensitivity and specificity of diagnostic neuroimaging, and the understanding of the diagnostic value of interictal scalp electroencephalography (EEG). This could improve the long-term prognosis of patients with non-epileptic seizures (NES) and candidates for potentially curative resective surgery for epilepsy because treatment earlier in these patients'�� disease course has been shown to be more effective. Our novel method for optimizing hyperparameters has the potential to slightly improve the accuracy of machine learning models, while substantially increasing the interpretability of learned estimates and reducing computational cost. We define hyperparameters as variables that contribute to machine learning models but are not optimized jointly with parameters inherent to the model. When viewed as a whole, this body of work represents contributions both to the statistical development and application of machine learning to important clinical challenges.