Tech Talk Tuesdays: AI and Healthcare
Reshaping the Future of Physical Therapy Through the Metaverse and Multimodal Biofeedback
About 800,000 Americans suffer from stroke every year, and also more than 10 million people live with Parkinson’s Disease (PD) worldwide. The economic and societal burden of post-stroke, PD, or other neuromuscular disorders is very high. Immersive technologies, including augmented or virtual reality (AR/VR)—and eventually Metaverse— as new rehabilitation paradigms, can potentially revolutionize and enhance motor learning of patients in need by offering a safe environment and by replicating real-life scenarios. In this talk, I will present our recent work on virtual therapy and artificial intelligence for neuromotor disorders and beyond.
About The Presenter
Leila Barmaki, Assistant Professor of Computer and Information Sciences
Dr. Leila Barmaki is an Assistant Professor of computer and Information Sciences and a resident faculty member of the Data Science Institute at the University of Delaware. She directs the Human-Computer Interaction lab (HCI@UD). She holds a PhD, and MSc in Computer Science. Before joining UD, she was a postdoctoral fellow at Johns Hopkins University. In her research, Dr. Barmaki combines augmented and virtual reality, multimodal and applied machine learning, and human-computer interaction to design high-impact medical and educational interventions. Her work has been supported by NSF, NIGMS via DE INBRE, USDA, NASA EPSCoR, Amazon, and UDRF.
Methods to address bias in large language models and their relevance to healthcare applications
As the application of large language models (LLMs) and other foundation models continues to expand across our society, concerns about the possibility of worsening biases against different individuals and groups also continue to grow. Such concerns have been especially highlighted following the success of deep learning-based methods during the past decade. However, the widespread use of generative AI-based methods (including LLMs) deepens such concerns even further.
In this talk, I plan to briefly cover some of the present methods to address such concerns, which include the methods aimed at working with the input data, the designed model, or the inference procedure. As an important case study, I will then discuss the application of LLMs for clinical tasks, covering both present applications (like clinical note summarization) and future applications (like differential diagnosis). Referring to a couple of the projects in our research lab, I will discuss the trends in mitigating biases in LLMs used for biomedical and health applications.
About The Presenters
Dr. Rahmat Beheshti, Assistant Professor of Computer and Information Sciences
Rahmat Beheshti works in the area of machine learning with a special focus on health and medical applications. He has a unique background through formal training in engineering and medical schools and has extensive experience in leading interdisciplinary teams consisting of computer scientists, clinicians, health IT, and policymakers. His research lab primarily works on developing advanced machine learning (like deep learning and generative AI) algorithms to make new discoveries from large and complex health datasets.
