Week 0

Motivation for doing postdoctoral work as an ICoN fellow

The projects that I undertook during my PhD were aligned with the goals of basic science research. Specifically, I used task-based fMRI data to build predictive models of the cerebellum, with the goal of understanding its role in cognitive function. I primarily used neuroimaging methods and techniques to build these models, however, along the way, I recognized the utility of incorporating other modalities, namely transcriptomics, behavior, and neuropsychology to address my research questions. For example, I worked with gene samples from the Allen Human Brain Atlas, and although the data were sparse for the cerebellum, it allowed me to imagine another lens by which I could investigate cerebellar functional organization. I was also involved in two patient studies that were particularly impactful for me. Both projects were collaborations between multiple research groups invested in testing a novel theory of cerebellar function. In multiple experiments, we tested patients with cerebellar degeneration (e.g. spinocerebellar ataxia) on a series of tightly controlled behavioral experiments instead of simple clinical batteries. It was my first experience interacting with patient populations for the purposes of scientific research, and it motivated me to seriously consider the potential of translating basic science research, across multiple modalities, into patient-centered healthcare outcomes.
I am excited at the prospect of pursuing an ICoN computational fellowship in the Ghosh and Gabrieli research groups for multiple reasons. First, regardless of where my career takes me, I am keen to use the expertise I cultivated during my PhD to work on large-scale and innovative projects that exist at the intersection of health sciences, technology, and data science. Second, I am eager to work in a highly collaborative environment under the leadership of ICoN center faculty members. One of the appeals of the fellowship is that there are very clear overall objectives that are shared across the faculty, research scientists, and postdocs, all of whom are contributing unique skill sets and expertise to the work. Third, I will have the opportunity to extensively build my knowledge across a wide range of areas: clinical assessment, endophenotyping, advanced machine learning models, and cross-integration of multiple modalities (e.g. behavior, transcriptomics, neuroimaging, video/speech analysis).
Finally, it is an exciting time to work in precision medicine, and although there is an abundance of data and computing power, what is lacking are models of individual human behavior that can inform risk predictors of health. Therefore, the prospect of working in collaboration with the Children’s Hospital of Philadelphia is incredibly compelling as it facilitates the development of behavioral tests whilst simultaneously evaluating diagnostic models in a continuous fashion. Importantly, any collaboration with clinicians will keep patient wellbeing at the forefront of our work.

Project Goals

One of the goals of the Ghosh/Gabrieli lab is to develop biologically-driven risk predictors and monitors of mental health. This will be done using multi-modal datasets that correlate genetic information with clinical evaluations through endophenotypes learned from neuroimaging, behavior, actigraphy signals and speech/vocal recordings. Mental health disorders encompass an enormous landscape of behavioral symptoms and clinical assessment. To start, I would be interested in narrowing the scope to a set of symptoms that can be assessed across multiple modalities. Perserverative behavior is a symptom that underlies numerous mental health disorders, including ADHD, OCD, autism, and schizophrenia, and it is measurable via multiple modalities. For example, perseveration has been studied using fMRI and is thought to be the consequence of frontal failure to modulate striatal outputs during controlled, purposive action, resulting in the continued reselection of previously activated outputs (Shallice, 1988). Another modality that has been used to study preservation is behavior. Cognitive tasks including the Wisconsin Card Sorting task and the Trail Making task, have been leveraged to study perseveration in typical and atypical development (Benitz et al. 2016). More recently, novel tools have been developed that automate the labeling of effector movement, gaze, and facial expression in real time using technology like deeplabcut and mediapipe. This technology could potentially allow us to identify perseverative behaviors across adult and developmental populations. It is possible to imagine combining live and longitudinal information from large-scale video and speech datasets to build prognostic computational models that identify and predict perseverative behaviors across a large range of mental health disorders. Furthermore, based on what we already know from neuroimaging and behavior, we could validate these models using biologically-driven data that include large-scale networks of the cortex and subcortex, and behavioral variation.
In the context of this project, I am particularly interested in the following questions (posed in the original proposal):

Considering the practicalities of precision healthcare, should we expect neuroimaging data to be incorporated into every individual’s health care assessment? To what extent does brain imaging even improve prognostic applications? Large scale cortical and subcortical networks have provided us with valuable insights about brain organization and topography in both healthy and patient populations, however, it is not clear what role these data will play in diagnostic applications. Will we be able to reliably identify macro circuits that drive behavioral (and genetic) variation? And if so, will we be able to use these circuits to inform predictive and diagnostic models of mental illness?
How good are behavioral assessments? Are there cognitive tasks that better capture symptoms compared to self-report clinical questionnaires. This is a question that we have been exploring in the motor world, specifically in the context of identifying cognitive impairments in patients with cerebellar degeneration. Subtle deficits are often masked by self-report questionnaires, and are then only revealed in carefully controlled behavioral experiments that test very specific cognitive processes.
I also believe there is great scope to automate the analysis of human behavior to aid psychiatric diagnoses. Some of this work is already underway in the motor world (e.g., automating clinical assessment of spinocerebellar ataxia and Parkinson’s disease). I can imagine building models that identify symptoms of neurodevelopmental and neuropsychiatric disorders (e.g., perseveration, hyperfocus, task switching etc.) on continuous and longitudinal video data collected in naturalistic settings. These models of complex behavior could then be validated by biologically-driven data and used to inform clinical tools that are then adopted by healthcare professionals.

Other considerations

Finally, there are a number of ethical considerations that I would like to consider as I venture into the realm of precision medicine. Translational work is a new area of research for me and I am aware that precision medicine is a quickly evolving field that raises important questions about ethical responsibility:
Data privacy: Who stands to benefit, and who is harmed? What possible solutions help protect individuals’ and communities’ ownership over their own data? I think this question is especially relevant for neural data that is not (as of yet) being used to track, manipulate, or identify individuals.
AI and healthcare: with such heavy reliance on technology in developed nations, I think it’ll be important to consider the concerns associated with bias and morality of AI, and how it may actually challenge equitable access to healthcare.
Communicating uncertainty: How do we communicate uncertainty about our translational results, especially when models are continuously updated with new information. It’ll be important to understand the uncertainty arising from technical procedures: small/incomplete sample sizes, sparse longitudinal datasets, instrument variability etc.
To what extent can we engage with diverse communities in translating neuroscience research to the health space. Can we bring together neuroscientists, physicians, ethicists, engineers, mental health workers, and caregivers to understand the ethical impact of data- and model-driven precision medicine?