Neuroimaging - overview
Our research focuses on mental states recognition and identifying "statistical biomarkers" of mental disorders from neuroimaging data, an exciting and rapidly growing research area at the intersection of neuroscience and machine learning. Our ultimate objective is to gain better insights about the brain functioning; thus, interpretability and reproducibility of learned models are of particular importance.
A mental state can be cognitive, such as viewing a picture or reading a sentence, emotional, such as feeling happy, anxious, or annoyed while playing a virtual-reality videogame, reflect person’s perception of pain, and so on. Examples of mental disorders our research focuses on inclide schizophrenia, drug addictions, as well as neurological disorders such as Huntington's and Parkinson's diseases.
Our primary focus is on functional and structural MRI and EEG data, including both medical-grade and wearable EEG devices.
- Spatio-temporal dynamics in brain imaging
- Schizophrenia: network-based discriminative models
- EEG-based cognitive load prediction with deep convolutional networks
- Wearable EEG-based mental state recognition
- Pain perception
- Cocaine addiction: sparse Markov networks and beyond
- Cognitive states in videogames predicted with sparse regression
- Huntington's disease: biomarkers and disease progression models
Spatio-temporal dynamics in brain imaging
The field of functional brain imaging is largely constrained by an analytic framework that focuses on the relationship between the activation of an area and the presence or absence of a task or stimulus, while ignoring temporal and spatial correlations that may influence local responses as much as the experimental paradigm. Using a variety of mathematical and computational approaches including statistical network theory, dynamical systems theory and machine learning, we are uncovering the underlying regularities spatio-temporal regularities present in functional magnetic resonance imaging (fMRI) data and electro-corticogram (ECoG) data, studying resting state and task-based activity in humans, as well as anesthesia induction and recovering in humans and primates.
Schizophrenia: network-based discriminative models
It has been long recognized that schizophrenia, unlike certain other mental disorders, appears to be delocalized, i.e. difficult to attribute to a dysfunction of a few specific brain areas, and may be better understood as a disruption of the emergent, collective properties of normal brain states that can be better captured by functional networks as opposed to failures limited to a specific brain area. However, identifying neuroimaging-based features that could serve as reliable statistical biomarkers of the disease, across multiple datasets, experimental conditions and patient cohorts, remains a challenging open problem. Our work is focused on exploring topological properties of functional brain networks derived from fMRI data in a predictive setting which emphasizes generalization (prediction) accuracy of network-based features to previously unseen subjects, as well as their stability across datasets. This methodology provides a more comprehensive and stricter evaluation of candidate biomarkers/features than the traditional approaches often limited to only statistical significance testing. Moreover, we demonstrate that network disruptions observed in schizophrenia cannot be explained by a disruption of area-based task-dependent responses, i.e. such disruptions indeed relate to the emergent properties of the brain.
EEG-based cognitive load prediction with deep convolutional networks
One of the challenges in modeling cognitive events from electroencephalogram (EEG) data is finding representations that are invariant to inter- and intra-subject differences, as well as to inherent noise associated with such data. We developed a novel approach for learning such representations from multi-channel EEG time-series which aims at preserving spatial, spectral and temporal structure in EEG data, and utilizes state-of-art deep recurrent convolutional networks to achieve highly accurate EEG-based classification of mental states, such mental load levels in a memory task.
Wearable EEG-based mental state recognition
Consumer electroencephalograms (EEGs) are much more portable and user-friendly than their medical-grade counterparts, but data from these devices lack resolution and signal quality. Therefore, their potential applications in continuous, portable monitoring operators for states inappropriate to the task (e.g. drowsy drivers), tracking mental health (e.g. anxiety) and productivity (e.g. tiredness) are fraught with difficulties. In this work, we examined responses to two different types of input: instructional (“logical”) versus recreational (“emotional”) videos, using a range of machine-learning methods. Our results demonstrate a significant potential of wearable EEG devices in differentiating cognitive states between situations with large contextual but subtle apparent differences.
Pain perception modeling
Discovering brain mechanisms underlying pain perception remains a challenging neuroscientific problem with important practical applications, such as developing better treatments for chronic pain. Herein, we focus on statistical analysis of functional MRI (fMRI) data associated with pain stimuli and apply multivariate predictive models such as sparse regression which combines model learning with variable selection, thus identifying brain areas most relevant (in terms of predictive accuracy) to pain. Furthermore, we combine such data-driven models with analytical differential equation models capturing the temporal dynamics of thermal pain perception.
Cocaine addiction: Multivariate predictive modeling
In this work, we aim to characterize multivariate patterns associated with abnormalities present in cocaine addiction, as measured by fMRI. More specifically, we developed a novel machine-learning technique for learning Markov networks that combines both node- and edge-selection, that we refer to as variable-selection sparse Markov network learning. This approach produces much more interpretable results as compared to standard sparse Markov net methods, and detects clearly identifiable network abnormalities in cocaine-addiction subjects. Furthermore, we also explore the effects of certain drugs such as methylphenidate (MPH) on the functional network features such as node degrees; our results provide a new type of evidence, based on discriminative accuracy of classifiers, that MPH tends to normalize network properties in cocaine addicts, as suggested by prior studies.
Sparse models and interpretable fMRI Analysis
This line of work spans multiple neuroimaging projects mentioned above, and started with an application of sparse regression methods, such as the Elastic Net, for mental state prediction during videogames played by subjects in fMRI scanner (PBAIC competition 2007). The advantage of using sparse multivariate models in brain imaging is two-fold: first, they use regularization, such as l1-norm constraint, that allows to handle high-dimensional but small-sample datasets typical for fMRI (e.g., 50,000 voxels/variables and only a few hundreds of samples), and second, this specific type of regularization promotes variable-selection (sparsity), and thus a better interpretability of a model (e.g., shows which brain areas can best predict a mental state). Sparse graphical models, such as sparse Markov networks we used in the cocaine addiction study, are another example of multivariate model where the sparsity constraint helps to improve model interpretabilty.
Huntington's disease: biomarkers and disease progression models
This recently started project focuses on statistical analysis of neuroimaging data (fMRI, MRI, DTI) in Huntington’s disease, which involves exploring potential statistical biomarkers of the disease and building imaging-based disease progression models.