The opportunity to work with Danfeng (Daphne) Yao in her Human-Centric Machine Intelligence Lab at Virginia Tech led Tanmoy Sarkar Pias to the university and the Sanghani Center. A Ph.D. student in computer science, Pias said it is amazing how Yao, his advisor, has taught him to identify real problems and practical solutions.

Pias’ research focuses on advancing the trustworthiness and fairness of machine learning models by developing systematic evaluation methods to uncover critical deficiencies. 

“Once these issues are identified, I design targeted mitigation strategies to improve model performance, ensuring that artificial intelligence applications in healthcare are robust, reliable, and equitable,” he said. 

His work involves diverse data types -- including time-series, tabular, images, and text -- and leverages state-of-the-art AI/machine learning (ML) models such as Large Language Models (LLMs); transformers;  vision transformers; Long Short-Term Memory (LSTMs); as well as classic machine learning models. Pias uses techniques such as gradient-based methods; neural activation maps; response curves, SHapley Additive exPlanations (SHAP) values; and synthetic test cases to understand how models make decisions and whether they respond appropriately to important changes.

“Imagine a patient in the intensive care unit whose blood pressure is dropping rapidly and oxygen levels are dangerously low -- signs that they might go into shock or face organ failure. Ideally, an AI-powered hospital system should immediately recognize these warning signs and alert doctors so they can intervene,” he said.

But in a recent study published in Communications Medicine, Pias and his collaborators found that many existing AI models fail to respond to these kinds of critical changes.

In fact, the study showed that 66 percent of severe health conditions went unrecognized, meaning the model did not raise the alarm in time, potentially putting patients at risk. After identifying these issues, proper measures are taken to fix them, making the AI models more trustworthy, robust, and reliable, he said.

Pias has always been driven by one question: “How can I use my research to help people and improve society?” 

“Healthcare is a field where AI can truly save lives, but only if it works correctly and reliably, he said. “It should be as reliable as the doctors who depend on it,” he said.

Being a student at the Sanghani Center has provided him the opportunity of engaging with researchers across different AI/ML domains. “I enjoy attending seminars, presenting my research, and receiving valuable feedback, which enhances my work through interdisciplinary collaboration,” Pias said.

He is projected to graduate this year and wants to continue working at the intersection of AI and healthcare, either in academia or industry, where he can apply his research skills to develop trustworthy and useful AI systems.