Speaking of Health: Leveraging Large Language Models to Assess Exercise Motivation and Behavior of Rehabilitation Patients
Research Poster Health & Life Sciences 2025 Graduate ExhibitionPresentation by Suhas Bettapalli Nagaraj
Exhibition Number 53
Abstract
Cardiovascular diseases (CVDs) and chronic obstructive pulmonary disease (COPD) are major global health concerns, with CVDs causing 17.9 million deaths annually and COPD expected to rank as the third leading cause of death by 2030. Exercise rehabilitation is crucial for improving physical and mental health, yet monitoring patient motivation and behavior often relies on subjective self-reporting, which lacks precision and detail. This study investigates the use of large language models (LLMs) such as Llama-2, Meditron, and Phi-2 to analyze patient conversations and assess exercise motivation and behavior. A novel framework combines clinician-provided labels, LLM predictions, and multimodal embeddings from speech and text using a transformer-based model. The adaptive Huber loss function enhances performance by addressing the variability in rehabilitation data. Using 73 recordings from 31 patients, the framework uncovered insights into motivation, mental health, and behavior, with motivation shown to be higher during rehabilitation sessions. Compact LLMs, like Phi-2, matched or exceeded the performance of larger models in predicting patient outcomes, demonstrating their potential in real-world applications. This research highlights the value of LLMs in improving clinical decision-making and patient care, with future directions including domain adaptation, enhanced modeling techniques, and new evaluation metrics.
Importance
This study tackles a pressing challenge in healthcare: understanding and improving patient motivation during cardiopulmonary rehabilitation. Heart and lung diseases are leading causes of death worldwide, and while rehabilitation programs are essential for recovery, tracking patient progress often relies on incomplete self-reports. By using advanced ML models to analyze patient speech and behavior, this research provides a more accurate and objective approach to assessing motivation and mental health. These insights can help clinicians personalize care, improve rehabilitation outcomes, and better support patients. This work not only advances healthcare technology but also opens the door to more effective, patient-centered approaches in rehabilitation, ultimately improving quality of life for those with chronic conditions.
DEI Statement
This research addresses health disparities by focusing on cardiopulmonary rehabilitation patients, a population often underserved in terms of personalized care. Traditional monitoring methods rely on subjective self-reports, which fail to account for diverse patient experiences, including variations in age, mental health, and socioeconomic status. By leveraging ML models to objectively assess patient motivation, behavior, and mental health, this study provides a scalable solution that can accommodate diverse needs and reduce inequities in care. Additionally, the integration of speech and text analysis allows for deeper insights into communication barriers and cultural factors that influence rehabilitation outcomes. This work aims to empower clinicians with tools to deliver equitable, personalized care, improving health outcomes for traditionally marginalized or overlooked populations.