Personalized medicine
The whole body of my research conducted as a graduate student and a postdoctoral fellow focuses on the development of accurate and efficient algorithms for personalized medicine and healthcare data. I have worked across preclinical and clinical data to better understand the underlying experimental confounds.
Personalized prognosis of children with depression in to bipolar disorder
While psychiatrists are very good at diagnosis of disorders, several psychiatric disorders change into another diagnosis after several years. Differentiating unipolar (MDD) vs. bipolar (BP) course of pediatric mood disorders continues to be a significant struggle for clinicians treating children showing early signs of mood disorders as antidepressant treatment, while being beneficial in unipolar depression, can deteriorate children at high risk to develop bipolar disorder. In a study of 492 children with longitudinal data, our analyses focused on predicting the prognosis of these children. Using the extra trees algorithm, our machine-learning model predicted the future development of bipolar disorder 10 years earlier based on childhood characteristics at baseline with high accuracy. The key predictors, according to our model, were problematic school behaviors, conduct disorder symptoms, and emotional dysregulation
Publication
Uchida Mai*, Bukhari Qasim*, DiSalvo Maura, Green Allison, Honnold Evan, Serra Giulia, Gabrieli John, Biederman Joseph; Can Machine Learning Help Identify Childhood Risk Indicators of Future Development of Bipolar Disorder? Results from a 10 Year Longitudinal Follow-Up Study; Bipolar Disorders (submitted)
Personalized prediction of treatment response outcome
For many psychiatric disorders, including anxiety, about half of patients respond to a standard treatment. For social anxiety, the standard treatments are cognitive behavioral therapy (CBT) and pharmacotherapy, which are similarly but only moderately effective, so that a large proportion of patients remain symptomatic after one intervention. Although both treatment modalities are superior to placebo on average, no reliable predictor of treatment response is currently available to select an optimal treatment for a given patient. Individual optimization, termed precision medicine or personalized medicine, has the potential to improve care with fewer treatment trials. One big challenge of this project was, we aimed to use very limited data, only that is typically available during a normal psychiatrist visit, which excludes complex multi modal imaging data. The models we developed, achieved an impressive explained variance of above 30% for predicting the individual patient response to CBT (measured as change in Liebowitz Social Anxiety Scale (LSAS) score) on the basis of demographic information, clinician ratings, and patient self-report surveys, in such a challenging problem. These models can be further extended to other neuropsychiatric disorders for example schizophrenia, depression, Alzheimer’s disease, and autism among others. To assess generalizability, we used four distinct classes of machine learning models together with cross-validation for each model. Our findings indicated that LSAS questions could account for 30% variance in predicting outcome, which will ultimately allow us to make better decisions for treatment.
Publication
Bukhari Q, Rosenfeld D, Hofmann SG, Gabrieli J, Ghosh S; Behavioral and Psychological Predictors of Treatment Response to Cognitive Behavior Therapy in Social Anxiety Disorder: A Machine Learning Approach; Translational Psychiatry (submitted)
Computational Neuroscience
I worked on several computational neuroscience projects, that have also been discussed in other sections. Here I have listed the projects that aim to develop a new computational method for the fundamental functional connectivity analysis of brain maps.
A new method for functional connectivity estimation based on differential covariance
We introduce differential covariance analysis, a new method that uses derivatives of the signal for estimating functional connectivity. We generated neural activities from dynamical causal modeling and a neural network of Hodgkin-Huxley neurons and then converted them to hemodynamic signals using the forward balloon model. The simulated fMRI signals, together with the ground-truth connectivity pattern, were used to benchmark our method with other commonly used methods. Differential covariance achieved better results in complex network simulations.
Publication
Tiger W. Lin, Yusi Chen, Qasim Bukhari, Giri P. Krishnan, Maxim Bazhenov, Terrence J. Sejnowski; Differential covariance: A new method to estimate functional connectivity in fMRI; Neural Computation. https://pubmed.ncbi.nlm.nih.gov/32946714/
Differential covariance provides a physiologically and anatomically relevant measure of brain function and behavior
We systematically interpreted and validated functional connectivity derived from our differential covariance algorithm and identified that when applied to real datasets, our method more closely matches the underlying structural connections in the brain compared to the standard covariance matrix or precision matrix. These results were validated across several subjects in human connectome project datasets and resting state fMRI recordings from mice. In addition, our work also found that human connectome project subjects with a more integrated functional connectivity based on differential covariance tended to have shorter reaction times in several psychological tests. Our work provided a physiologically and anatomically relevant measure of brain integration and behavior.
Publication
Yusi Chen, Qasim Bukhari, Tiger Wutu Lin, Terrence Sejnowski; Functional connectivity of fMRI using differential covariance predicts structural connectivity and behavioral reaction times, Neuroimage (submitted)
Cognitive Neuroscience
I did several studies in the area of cognitive neuroscience to better understand the underlying neural processes and connections of the mental activities performed by the brain.
Social Anxiety Disorder patients changes in brain activation while viewing familiar faces
I am leading is identifying SAD patients using familiar vs. novel faces task paradigm. Using task fMRI of familiar and novel faces, I developed computational models that can accurately diagnose SAD patients.
Neural basis of intelligence
The goal of this project is to identify the neural basis of intelligence and define intelligence through brain matrices better than the IQ measure. We aim to compare how well a typical IQ type measure predicts performance on non-IQ cognitive tests (that’s largely why IQ matters) vs. best multi-modal brain measure (brain size, cortical thickness or area, resting state, DTI). Essentially through this project, I asked whether brain measures could outperform IQ in predicting other cognitive measures. This has implications on identifying the basis of intelligence in human brain, which can be used to improve cognitive skills. Using the models I have developed, I got promising results that are helping us improve our understanding of the basis of intelligence by identifying functional brain regions and their connectivity across the brain
Semi-prosopognosia – an analysis of a man who see half of every face as melting
The representation of faces in early stages of visual processing depends on retino-centered reference frames, but little is known about the reference frames that code the high-level representations used to make judgements about faces. Here, we focus on a rare and striking disorder of face perception—hemi-prosopometamorphopsia (hemi-PMO)—to investigate these reference frames. After a left splenium lesion, Patient A.D. perceives features on the right side of faces as if they had melted. The same features were distorted when faces were presented in either visual field, at different in-depth rotations, and at different picture-plane orientations including upside-down. A.D.’s results indicate faces are aligned to a view- and orientation-independent face template encoded in a face-centered reference frame, that these face-centered representations are present in both the left and right hemisphere, and that the representations of the left and right halves of a face are dissociable.
Publication
Jorge Almeida, Andreia Freixo, Miguel Tábuas-Pereira, Sarah B. Herald, Daniela Valério, Guilherme Schu, Diana Duro, Gil Cunha, Qasim Bukhari, Brad Duchaine, Isabel Santana; Face-Specific Perceptual Distortions Reveal A View- and Orientation-Independent Face Template. Current Biology. 2020;S0960-9822(20)31092-7. doi:10.1016/j.cub.2020.07.067
Previous research (from last decade)
I worked on several interesting during my Masters and PhD work. Some of the selected work I have listed below
Preclinical Imaging
During my PhD at ETH Zurich, I mainly worked with preclinical imaging mouse models. This experience helped me experiment personalized medicine and computational tools on knock out mice with the freedom to vary large ranges of drug dose.
Personalized dose effects: Dose dependent effects of anesthetics
In this project, I evaluated the dependence of mouse resting-state fMRI patterns on the dose of isoflurane using pseudo-stationary and dynamic functional connectivity analysis and found that at higher isoflurane levels spatial segregation among the brain regions is lost. Also the functional connectivity between homotopic regions was found to be lost at higher dose of isoflurane. This was the first time that dynamic functional connectivity was analyzed in mice fMRI data. I was able to extract extra information using dynamic functional connectivity, which was not apparent in the static functional connectivity analysis. My static and dynamic functional connectivity analysis explained two long standing theories on the work of anesthesia on brain networks by showing that loss of modular structure of functional connectivity and loss of connectivity between the homotopic brain regions at higher dose of anesthesia are essentially the same observed at different time scales.
Publications
Bukhari, A. Schroeter, and M. Rudin, Increasing isoflurane dose reduces homotopic correlation and functional segregation of brain networks in mice as revealed by resting-state fMRI. Nature Scientific Reports 8 (2018) 10591. https://www.nature.com/articles/s41598-018-28766-3
Effects of different types of anesthetics on brain functional networks
In this work, I applied network modeling and dual regression approach to the mice fMRI data to study the effects of different anesthetic regimens including isoflurane, medetomidine and the combination of the two, on brain networks. We showed that cortico-thalamic interactions depend on the type and depth of anesthesia and by combining the two anesthetics at low dose, we can superpose the interactions observed for each anesthetic alone. In this work I suggested that iso-med combination anesthesia is more suitable for future studies as it combines the benefits of individual anesthetics. This work was later used by several researchers around the world to guide their experiments.
Publications
Bukhari, A. Schroeter, D.M. Cole, and M. Rudin, Resting State fMRI in Mice Reveals Anesthesia Specific Signatures of Brain Functional Networks and Their Interactions. Front Neural Circuits 11 (2017) 5.
Identifying response to treatment using machine learning
I also applied machine learning to investigate the neurobiological differences due to alterations in brain functional and structural characteristics. At Children Hospital Boston together with Dr. Lino Becerra and Dr. David Borsook, I investigated the differences in the pharmacological MRI (phMRI) response to treatment with an analgesic drug (buprenorphine) as compared to control (saline) in rats using machine learning.
Publications
Bukhari, D. Borsook, M. Rudin, and L. Becerra, Random Forest Segregation of Drug Responses May Define Regions of Biological Significance. Front Comput Neurosci 10 (2016) 21.
http://journal.frontiersin.org/article/10.3389/fncom.2016.00021/full
Computer Vision
3D tracking bacteria using 2D microscopic images
In 2011, for my master thesis work at Light Microscopy lab of ETH Zurich under the supervision of Dr. Peter Horvath, I modelled the depth of moving biological objects (salmonella) using the point spread function of the microscope, providing a unique computer vision and artificial intelligence based methodology to track objects in 3D space using only 2D images. Our results showed that the motion of the bacteria can be approximated by the piece-wise linear regression model and tracked in 3D using Hungarian algorithm. We successfully tracked the salmonella bacteria from the given set of images in this work presenting a novel method based on point spread function matching to perform and analyze live cell imaging of the bacteria cells and track it in 3D using only 2D images.