Community Spotlight: Sriharsha Mopidevi

Can you tell us about your current research and what inspired you to pursue this area?

My work in the AI-4-AI Lab sits at the intersection of software engineering, machine learning, and healthcare informatics. I focus on two main areas: building privacy-preserving AI systems for clinical data and developing platforms that make research data more accessible and useful.

One of my primary projects is MedVidDeID, an open-source pipeline that removes protected health information from multimodal medical data, including video, audio, and text. Clinical encounters captured on video hold tremendous research value, but they also contain sensitive patient information that needs to be carefully handled. MedVidDeID integrates computer vision models for face detection, speech recognition for audio transcription, and NLP-based tools for identifying and redacting personal health information across all three modalities. The goal is to make it possible for researchers to work with rich clinical data without compromising patient privacy.

I also lead the development of the Observer platform, a clinical research data management system that supports the lab’s observational studies. It handles everything from data collection and storage to exploration and analysis, and I am currently building out a FAIR (Findable, Accessible, Interoperable, Reusable) data repository on top of it to facilitate collaboration and data sharing across research teams.

What drew me to this area was the realization that there is a real gap between what AI and software engineering can do and what is actually available to clinical researchers. Building the tools that bridge that gap felt like meaningful work.

How does your research intersect with real-world challenges?

Almost every part of my work is shaped by real-world constraints. Privacy is the most obvious one. Healthcare data is heavily regulated, and for good reason. But that regulation creates a tension: researchers need access to detailed, multimodal clinical data to do meaningful work, while patients deserve strong privacy protections. Building systems like MedVidDeID that can reliably de-identify data across video, audio, and text is a direct response to that challenge.

Infrastructure is another big one. Running deep learning models for video processing at scale is computationally expensive. I have spent a lot of time optimizing our Databricks workflows, from reducing costs to working through memory constraints in audio and video processing pipelines, so that these tools are practical for routine use and not just proof-of-concept demos.

Usability matters just as much. The best pipeline in the world is useless if the clinicians and researchers who need it cannot actually use it. So a significant part of my work involves building intuitive interfaces, designing data management workflows that fit into how people actually work, and making sure the systems I build are reliable and maintainable.

Finally, validation is critical. In healthcare, you cannot just deploy a model and hope for the best. Every component needs to be rigorously tested and its performance clearly documented, which is why our published work includes detailed evaluations of each stage of the pipeline.

What do you find most rewarding about working in the Department of Biostatistics, Epidemiology, and Informatics?

The interdisciplinary nature of the work is what I value most. In the AI-4-AI Lab, I get to work alongside clinicians, informaticists, and researchers who bring completely different perspectives to the problems we are tackling. That kind of collaboration pushes me to think beyond just the technical solution and consider how the tools I build will actually be used in clinical and research settings.

I also appreciate that DBEI is a place where applied work is taken seriously. Building production-grade software for healthcare research is not always the most glamorous work, but it has a direct impact on what researchers can accomplish. Being in an environment that values that kind of contribution is really motivating.

Working with Dr. Kevin Johnson has been especially formative. His vision for the lab, using AI and digital health tools to improve the science and practice of ambulatory care, gives all the technical work a clear purpose.

Can you share a significant recent project, publication, or professional recognition that you are particularly proud of contributing to or achieving?

The project I am most proud of is MedVidDeID and the publications that came out of it. Our paper on the audio-video de-identification pipeline was published in the Journal of Biomedical Informatics, where I was first author. It was a significant effort that involved integrating multiple AI models (WhisperX for speech recognition, YOLOv11 for face detection, PHIlter for text de-identification) into a cohesive, production-ready system and rigorously evaluating its performance.

We also published related work in JAMIA (Journal of the American Medical Informatics Association), and had papers accepted at AAAI and the Pacific Symposium on Biocomputing. Having our work recognized across both clinical informatics and AI venues felt like validation that we are building something that matters to both communities.

Beyond publications, seeing MedVidDeID used in practice and knowing that it enables research that would not be possible otherwise is deeply satisfying.

What advice would you give to students or early-career professionals in your field?

A few things I have learned along the way:

Be adaptable with technology. Frameworks and tools change constantly. What matters more than mastering any single stack is developing the ability to learn quickly and pick up whatever the problem requires. I have worked across Python, Node.js, Go, React, Django, and more, and the willingness to be framework-agnostic has been one of the most valuable skills in my career.

Build things that actually run in production. There is a big difference between a proof-of-concept notebook and a system that works reliably every day. Learning how to deploy, monitor, and maintain software will set you apart, especially in a research environment where that expertise is not always common.

Collaborate across disciplines. Some of the best insights in my work have come from conversations with clinicians and researchers who see the problems from a completely different angle. Do not stay siloed in your technical comfort zone.

And finally, do not be afraid to tackle ambitious problems. When I started working on multimodal de-identification, it felt like a massive challenge. But breaking it down into manageable pieces and iterating made it achievable. The problems worth solving are usually the ones that seem a little too big at first.

What is an interest or pastime that you enjoy outside of academia?

I have a deep interest in astrophysics and cosmology, and I am currently taking astrophysics classes at Penn. There is something grounding about thinking about the scale of the universe after spending a day debugging code or optimizing pipelines. I also enjoy keeping up with developments in space exploration and physics, and watching sci-fi movies (the more thought-provoking, the better).

Outside of that, I love cooking, hitting the gym, and spending time outdoors. I am a regular on the Schuylkill River Trail for long walks. When the weather is right, you will probably find me at the beach or out on a hike somewhere.