Faculty Feature: Archana Venkataraman

Dr. Venkataraman is an Associate Professor at Boston University, a Core Faculty member for the Center for Brain Recovery, and the Principal Investigator for the BU Neural Systems Analysis Laboratory.

Q&A with Archana Venkataraman

What is your current research focus, and how does it align with the Center for Brain Recovery’s mission?

My lab uses the power of data science, particularly artificial intelligence, to better extract information from noisy and complex biomedical data sets in order to improve our understanding and treatment of neurological and psychiatric disorders.

We use an interdisciplinary mindset to address questions like: ‘How do we extract better information from imaging data?’ and ‘How do we combine different sources of information to get a more complex, comprehensive patient picture?’.

Depending on the project, sometimes we’re in the early stages of research and we’d like to identify biomarkers or create diagnostic modules. In other projects, we might want to identify clinical or therapeutic targets that are difficult yet important to identify for neuroradiologists and surgeons. We use non-invasive, easily acquired data to help improve their decision making.

And then we have some exploratory work where we’re doing predictive analytics for various healthcare applications. In a completely different vein, we have also worked a little bit on speech to try and understand how to manipulate emotional cues in speech and use that as the basis for understanding how we perceive different emotional cues.

How did you initially become involved and/or interested in your field?

My path is somewhat non-traditional. I was an electrical engineering major in college with a focus on signal processing through my undergrad. During my master’s, I worked on a theoretical signal processing project and I realized that I wanted to do something more applied and something with a potential societal impact. 

From there, I moved into medical imaging for my PhD. We were focused on better understanding schizophrenia using different neuroimaging modalities that had been recently developed at that time. This included resting state fMRI, which was gaining popularity, and diffusion MRI. My focus was on designing different algorithms and models to try and integrate the two modalities to extract biomarkers for schizophrenia. 

During my postdoc at Yale, I was working in a lab that was based in radiology. There, I started working with a collaborator on autism. I really gravitated towards more interdisciplinary research. For the first time I recognized the synergy between having really active, engaged clinical partners, and the value I could bring from the analysis and technical standpoint. And I think my research has just grown from there.

What courses do you teach?

At the undergraduate level, I teach Signals and Systems. It’s a core electrical engineering course, but it is broadly applicable to a variety of domains. The course provides foundational knowledge for how one might think of and process different sources of information. It sets the foundation for more data analyses and computational modeling techniques later on.

On the graduate side, I have developed a new course called Foundations of Probabilistic Machine Learning. This course is focused on the building blocks on detection and inference, that is, trying to make decisions using data that you have that’s imperfect, or to estimate unknown quantities in the data or unknown phenomenon in the data. It’s a really foundational course in data science, but it’s also bringing data science to real world problems.

What research are you most proud of?

One work that I’m really proud of is the work that my lab has done on imaging-genetics, and specifically on analyzing human genetics using artificial intelligence. We’re really trying to understand how we can embed structure in these deep neural networks so that we’re able to learn from small sample sizes of very complex data. From the work that we’ve done, we’ve shown that there’s tremendous improvements in terms of diagnostic predictive accuracy in schizophrenia, but also that these types of models provide an interpretability that you don’t normally get with deep learning.

On the more translational front, there are two projects that I’m really proud of. First, my lab has done a lot of work on epilepsy, and specifically on trying to localize the seizure onset zone using non-invasive data. We’ve worked on clinically acquired EEG and resting state fMRI. And I’m really proud of the approach that we’ve taken to better mine information from non-invasive data that’s inexpensive, easy to acquire before the patient has to go through any traumatic procedures, and can be implemented and translated easily into a clinical setting.

And then, also on the surgical side, we’ve done work on localizing the eloquent cortex using resting state fMRI. For a bit of context, if any neurosurgical procedure needs to be done in the temporal lobe, then there is always the risk of damaging key areas of the brain known as the eloquent cortex. Unfortunately, trying to identify those areas is challenging. You can use invasive methods which are really traumatic for the patient, or you can use a traditional fMRI which some patients cannot complete if they are severely impaired. So, having the ability to find these regions using a passive modality like resting state fMRI could be very game changing within the context of preoperative mapping.

What do you consider the most pressing challenge in your field today, and how is your research addressing this challenge?

There are a couple of fundamental challenges in the field. One challenge, which is becoming increasingly obvious as much of the research community transitions to AI, is the scarcity of data. Often neuroimaging data is expensive to collect, difficult to collect, and is very segregated. Each lab or each hospital might collect data in a piecemeal fashion, and it’s just not generally accessible.

On the collaborative front, I have started to build relationships with a variety of different clinical partners who share the recognition that we need to do more to integrate data, to provide it for research, and to widely disseminate our technology.

With regards to the data issue, the way we handle it in my lab is to come up with more creative ways of designing AI models that account for the fact that we have limited sample sizes. We have used a variety of strategies: developing models for missing data, embedding structure into the models that we create, reducing the parameterization of the models very strategically, and really pressing on this idea of blending domain knowledge with AI. So, rather than just let AI run free through the data, we provide some guardrails and some guides so that it can be more efficient in how it processes.

If you could give one piece of advice to someone just starting out in your field, what would it be?

My advice would be that great research comes from great collaborations and great partners. As engineers, we often get trapped in our own little bubble. We’re surrounded by folks who have very similar skill sets, or think in a very similar way, but in order to move the medical field forward— particularly in brain imaging, neurology, neuroradiology, and psychiatry— it really needs to be interdisciplinary where everyone has a stake and everyone is bringing ideas and different perspectives to the table.

What is something unique about you that people might not know?

I’m an all weather runner, so I run in rain, sleet, snow, extreme heat, etc. I’ve been doing this for several years since college.

Is there any additional information you would like to share with the CBR community?

You can visit the Neural Systems Analysis Laboratory website to learn more about my lab’s work.

View all posts