Our research lies at the intersection of multimodal integration, network modeling and clinical neuroscience. By combining different viewpoints, such as imaging features, behavioral characteristics and genetic markers, with key neuroscientific insights, our goal is to characterize complex processes within the brain. In particular, we focus on modeling how the brain changes given a neurological impairment, and how we might use this information to improve clinical care. Our current projects include:
Fusing Neuroimaging and Behavior to Characterize Patient Heterogeneity
Resting State fMRI is a non-invasive neuroimaging modality which has become ubiquitous for studying neurological disorders such as Autism, ADHD and Schizophrenia. The individual impairment in social and cognitive functioning among clinical populations is typically quantified by biomarkers of symptom severity. These measures are assessed and scored by a trained clinician. Owing to the well known group level confounds of rs-fMRI, a reliable characterization of behavioral heterogeneity from neuroimaging data at an individual level is extremely challenging.
In this project, we develop a novel optimization framework which jointly models the group level and patient specific resting state information through a joint matrix factorization across patients. Going one step further, we use the patient level description to predict clinical severity via a regression model. We combine the neuroimaging representation and severity prediction steps into a single Joint Network Optimization (JNO) framework, and leverage computationally efficient optimization techniques to infer the latent representations inherent to our model. This helps us isolate key resting state co-activation signatures associated with various neuropsychiatric disorders, and, at the same time, map the wide spectrum of clinical manifestations across patients.
Noninvasive Seizure Localization for Epilepsy
Epilepsy affects nearly 3.5 million people in the United States and is linked to a five-fold increase in mortality. While epilepsy is often controlled with medication, 20-40% of patients are medically refractory and continue to experience seizures in spite of drug therapies. The alternative for many of these patients is to surgically remove the areas of the brain responsible for triggering an epileptic seizure. However, the gold-standard procedure to localize the epileptic foci requires a craniotomy and implantation of electrode grids directly onto the cortical surface. This evaluation procedure is understandably traumatic and increases the patient risk for infection and injury.
Hence, our goal is to automatically and noninvasively localize the seizure onset zones, which will both streamline the pre-surgical planning process and reduce the amount of time the patients spend in the hospital. As such, we have developed a novel Coupled Hidden Markov Model to detect and localize epileptic seizures in clinical EEG recordings. This model tracks the spatio-temporal propagation of a seizure to back-infer the onset. We have validated our approach on data from the Johns Hopkins Hospital and the Children’s Hospital in Boston. Ongoing work focuses on generalizing the performance via hierarchical graphical models and deep learning approaches. Moving forward, we will combine multiple snapshots of the brain, such as structural MRI, functional MRI and EEG in order to refine the localization.
Bridging Imaging, Genetics and Diagnosis
Neuropsychiatric disorders, such as autism and schizophrenia are highly heritable, suggesting a strong genetic underpinning. These disorders are greatly characterized by behavioral and cognitive deficits linked to atypical neural functioning. The genetic variants play a crucial role in altering the neural pathways. Therefore, identifying the brain mechanisms through which the genomes confer risk is essential to uncover targeted biomarkers for these disorders. Functional MRI (fMRI) and SNPs are the most common modalities used for studying brain activity and genetic variation, respectively.
In this work we propose a joint dictionary learning framework that couples imaging and genetics data in a low dimensional subspace as guided by clinical diagnosis. The goal is to identify the potential imaging and genetic biomarkers that are related to diagnosis as a basis vector of the low dimensional space. We use a graph regularization penalty and group sparsity penalty to find the representative set of anatomical and genetic basis vectors, respectively. We have evaluated our model on two task fMRI paradigms and single nucleotide polymorphism (SNP) data from schizophrenic patients and matched neurotypical controls. We employ a ten fold cross validation technique to show the predictive power of our model. Ongoing work focuses on capturing non-linear relationships between the two imaging and genetic data using deep learning techniques.
Modeling Neural Plasticity Induced by an Acute Lesion
Neural plasticity refers to the anato-functional reorganization of the brain across the lifespan, and it plays a key role in our natural ability to compensate for injuries to the brain. For example, low grade gliomas are slow-growing tumors that originate in the central nervous system. Preoperative and intraoperative imaging suggest a progressive redistribution of the affected functional areas, from non-primary perilesional areas to homologous regions in the contralateral hemisphere. However, our understanding of functional reorganization is still largely heuristic and limited in scope.
We have begun to characterize the functional reorganization of glioma patients using a Bayesian framework. Namely, we leverage a Markov Random Field prior to smoothly refine functional network boundaries that were derived from a healthy cohort on a patient-specific basis. Ongoing work focuses on combining clinical insight with graph theoretic approaches to identify the resting-state language network of these patients.
Manipulating Emotional Cues in Human Speech
Typical human speech carries both, the semantics of language and emotion of the speaker. While semantics help one to develop algorithms to understand and produce human speech, emotional aspect is also important for its naturalness. Prosody features namely: pitch, intensity, duration and spectral tilt are known to encode the emotion information. Apart from the classical problem of recognizing emotion from human speech, another interesting problem is how to alter its perceived emotional content? This goal has profound implications for ASD. One of the earliest observations of ASD was a struggle with both the production and interpretation of prosodic information, particularly within the context of emotions. Hence, artificially amplifying the emotional content of human speech may allow individuals with ASD to perceive emotional cues in the same way that a non-autistic individual would discern them in the original signals.
Our approach is to leverage tools from computer vision to develop explicit models for the prosodic feature warping. Currently, we are developing an algorithm to predict an entire pitch trajectory based on just the acoustic information from source emotion. Another approach is to warp the time-frequency representation (spectrogram) from source emotion to target emotion via deformable image registration. The underlying assumption is that the shape of overtones in spectrograms are influenced by the speaker’s intonation. Thus, it can be used as a signature for the detection of emotion as well as its modification.
Improving Rehabilitation in Spinal Cord Injury Patients
Spinal cord injury (SCI) is a devastating trauma that can result in permanent loss of movement and sensation. While medical advancements have reduced the mortality rates for these patients, the quality of life for SCI patients remains relatively grim, as more than 50% of them will be left with severe paralysis, and the remaining will only partially recover their original functionality. Physical therapy and rehabilitation expedite recovery by retraining neural processes in the brain, a phenomenon known as functional reorganization. However, functional reorganization is not well characterized or understood, which limits both the development of more targeted therapy plans and the extraction of biomarkers for recovery.
Resting-state fMRI has the potential to detect functional reorganization due to SCI, which is necessary to identify therapeutic targets and track treatment induced changes. Combining machine learning theory with neuroimaging data and clinical insights can help explain functional reorganization of patients who sustained a spine or brain injury. We have begun to identify whole-brain connectome in these patients using prior established methods from our work. Namely, we are pinpointing specific degraded functional connections using a matrix factorization framework and characterizing subject-specific functional organization using a Bayesian approach.
Estimating the Lower Limit of Cerebral Autoregulation
Cerebral autoregulation refers to the brain’s ability to maintain a constant blood flow despite changes in blood pressure. This blood flow is crucial for oxygen delivery to maintain proper brain and body functioning. For example, abnormally high blood flow will cause edemas, while insufficient blood flow can lead to cognitive deficits and organ damage. A key parameter of cerebral autoregulation is the lower limit of autoregulation (LLA), which is the lowest blood pressure at which the brain can maintain a consistent blood flow. While the LLA has been well characterized in healthy individuals, there is evidence that it shifts in critically ill patients, and especially for individuals undergoing surgery. Therefore, determining a reliable way to measure the LLA will help improve patient management during surgery as well as long-term outcomes.
For this work, we measure arterial blood pressure (ABP) with a transducer and use near-infrared spectroscopy (NIRS) as a surrogate for cerebral blood flow. The Pearson’s correlation coefficient between ABP and NIRS measurements in 5 minute windows is used to assess whether or not cerebral autoregulation is functioning. These correlation values are binned according to the mean ABP in their corresponding window. In traditional methods, the LLA is assumed to be the lowest ABP at which the mean correlation coefficient is below a particular threshold value. The goals of this project include determining whether the threshold-based method is useful in predicting outcomes of surgery and exploring if there may be more reliable methods to compute the LLA with the correlation coefficient data.