Episode Summary
Elizabeth Skidmore, a Professor of Occupational Therapy at the University of Pittsburgh and Associate Dean for Research at Pitt’s School of Health and Rehabilitation Sciences, speaks with Pitt HexAI podcast host Jordan Gass-Pooré about Beth’s work as an occupational therapist and rehabilitation scientist with expertise in neurological rehabilitation. Beth and Jordan discuss rehabilitation sciences, intervention program design, program implementation, fidelity, applications of artificial intelligence and the training of occupational therapists in the AI age.
Episode Audio
Episode Transcript
Welcome to the University of Pittsburgh’s Health and Explainable AI Podcast. I’m your host, Jordan Gospa, a health and science reporter. Join me as we cover advancements being made in health informatics and explainable AI for students, researchers, and healthcare practitioners interested in applications of artificial intelligence and machine learning.
This podcast is produced by the University of Pittsburgh’s Health and Explainable AI Research Laboratory at the University of Pittsburgh’s School of Health and Rehabilitation Sciences, Department of Health Information Management, headed by Ahmedi Pitts. The HEX Lab cultivates extramural collaborations with academic institutions both nationally and internationally through its research, educational contributions, and this podcast series.
Hello and welcome back to Pitt HEX AI, a podcast series produced by the University of Pittsburgh’s Health and Explainable AI Research Laboratory. I’m Jordan Gospa, your host, and today we’re going to speak with Elizabeth R. Skidmore, a professor in the Department of Occupational Therapy at the University of Pittsburgh and Associate Dean for Research at Pitt’s School of Health and Rehabilitation Sciences. Welcome, Elizabeth.
One of the first things I wanted to ask—because folks might not know exactly what Rehabilitation Sciences is—could you give us a brief explanation of what that actually means and what you do day to day to give some context?
Rehabilitation Sciences is a broad discipline that incorporates many health disciplines interested in studying how injury and illness occur, how people recover from illness and injury, what interventions make an impact on that recovery, and how we can best deploy those interventions within health systems and communities. It’s a very broad field.
Within the School of Health and Rehabilitation Sciences at the University of Pittsburgh, we have over 16 disciplines represented in that broad category of science. Many people think of occupational therapy, physical therapy, and speech-language pathology as examples, but it also includes colleagues in health informatics, rehabilitation science and technology, and a variety of other disciplines that all work together toward this larger goal.
You focus on neurological rehabilitation, is that correct?
That is correct. My clinical background and research focus specifically on how we can deliver interventions to better meet the needs of people who’ve had changes in cognition after brain injury, stroke, or other conditions that affect cognitive function. We study how to optimize interventions and reduce disability among those populations.
My day-to-day is complex because I wear many hats. As an occupational therapist in the clinic, you would find me in an inpatient rehab hospital or in the community delivering interventions. As a professor at the University of Pittsburgh, I oversee research in my own lab and train students and faculty in conducting research. As Associate Dean for Research, my job is to oversee the infrastructure that helps all my colleagues conduct their research—making sure we have the right personnel, systems, and financial management in place to support their work.
I work in large multidisciplinary teams. As an occupational therapist, I collaborate closely with neuroscientists, neurologists, neurosurgeons, physiatrists (rehab physicians), physical therapists, speech-language pathologists, rehabilitation nurses, and psychologists. We bridge traditional sciences like neuroscience and physiology with clinical sciences.
As for my background, I was always interested in health sciences and originally thought I was destined for medical school. I participated in several pre-med programs, but I wanted to work more closely with people addressing the problems they encounter in everyday life. Occupational therapy drew my attention because it focuses on helping people develop creative solutions to live fulfilled and productive lives.
I earned a bachelor’s degree in occupational therapy and practiced for several years in inpatient rehabilitation with people who had stroke and brain injury. I began to ask questions: Why do some patients do better than others? What interventions work best? How can I better integrate evidence into practice? And how can I overcome system-level barriers to delivering the right intervention? Those questions drove me to earn my PhD in Rehabilitation Sciences and become a full-time scientist working closely with clinicians.
Regarding AI, the term has been circulating in our field for about 10 years, but much more prominently in the last couple of years with commercial applications like ChatGPT. Many in my field don’t have deep training in AI, so it often feels like a broad, somewhat mysterious concept.
My direct experience with AI began through collaborations with Dr. Yanan Shan Wong and Dr. Liming Xia at Pitt. We were trying to solve a real clinical problem in one of our studies, and they introduced us to the potential value of machine learning methodologies.
I initially thought of AI as a “black box”—you put information in and get an answer out—which made me skeptical as a scientist trained in hypothesis testing and confirmatory analysis. But my clinical trials research faced a challenge: assessing intervention fidelity.
Over 15 years, we’ve developed an intervention for people with cognitive impairments that improves engagement in rehab and reduces long-term disability. However, we must assess fidelity—ensuring the intervention is delivered as intended. Highly trained raters watch about 20% of our therapy session videos and score them using a standardized checklist. This is expensive and time-consuming.
As we scale to 50 rehab hospitals nationwide, this manual approach becomes untenable. I joked with my AI colleagues: “If you could take AI and analyze thousands of hours of videos to determine whether intervention sessions meet fidelity criteria, that would be great.” They said it was possible, and that’s how we began exploring machine learning models.
We annotated videos and compared machine learning outputs to our trained raters’ scores. We found high coherence between the two approaches. The long-term goal would be a system where clinicians upload session videos and receive automated fidelity feedback.
There are concerns about AI, of course. Like any analytic method, it’s “garbage in, garbage out.” Poor assumptions or biased data can lead to faulty conclusions. There’s also national discussion about transparency and bias in AI models. In rehabilitation research, health equity is central, so we must ensure data represent the populations we serve and remain mindful of biases in measurement, provider decision-making, and electronic health records.
From a student perspective, there’s excitement and overwhelm. Many students are clinicians who think in terms of individual patients rather than large datasets. Both students and faculty need more training in responsible AI use. At present, training in occupational therapy around AI is limited, though other rehabilitation disciplines may be further along.
Regarding fidelity: when testing an intervention, it’s not enough to know whether it works—we must know whether it was delivered as intended and at sufficient “dose.” For pharmacological interventions, dose can be chemically measured. For interactive, non-pharmacological interventions, we must break down components and evaluate delivery through structured checklists.
In our case, we require at least 80% fidelity for positive outcomes. But manual video review is time-intensive and costly. Machine learning offers a scalable alternative for assessing fidelity across large patient populations.
Good study design starts with a well-defined research question, measurable variables, reliable and unbiased measures, and appropriate analytic methods. Bad study design includes flaws in any of those elements. Bias can never be fully eliminated, but it can be minimized through careful design and randomization.
For students interested in research ideas, one promising area is analyzing non-verbal communication in rehabilitation therapy. We’ve identified verbal markers of guided discovery versus directive teaching, but non-verbal communication—facial expressions, posture, gestures—remains difficult to systematically classify. Computer vision is advancing, but non-verbal behavior does not easily fit into discrete categories. Understanding and modeling these elements could significantly strengthen rehabilitation research.
Elizabeth, thank you so much for being on the show.
Thanks to everyone for tuning in and following the show. The Health and Explainable AI Podcast is produced by the University of Pittsburgh’s Health and Explainable AI Research Laboratory at the University of Pittsburgh’s School of Health and Rehabilitation Sciences, Department of Health Information Management.
I’m Jordan Gospa. Thanks for listening.
