Interdisciplinary Initiatives Program Round 12 - 2024
Project Investigators:
Jonathan H. Chen, Medicine - Biomedical Informatics Research
Michael Bernstein, Computer Science
Robert Tibshirani, Biomedical Data Science and Statistics
Mary Kane Goldstein, Health Policy
Abstract:
Over 25 million people in the United States have deficient access to medical specialty expertise, disproportionately affecting marginalized communities. Delays in access can lead to worsened health outcomes and higher mortality rates. Current electronic consultation (eConsult) systems can improve access by reducing time and space barriers for specialty medical advice but remain constrained by specialist labor shortages. Advanced clinical decision support (CDS) systems and foundation models in healthcare offer the potential to scalably meet the needs of more patients, but they face challenges such as privacy concerns, data security, and integration into clinical workflows. To address these issues, this project aims to develop and evaluate advanced artificial intelligence (AI) technologies, specifically a Retrieval-Augmented Language Model (RALM) using state-of-the-art large language models (LLMs) combined with real medical records and evidence-based reference sources. To concretely demonstrate and evaluate the potential of such an approach, we will develop a system to predict diagnostics and medications that a subsequent (specialist) physician would order for a patient requesting a consultation. This could translate into increased capacity and access for patients by creating a new scalable channel for consultative suggestions that could eliminate unnecessary or ineffective in-person visits that only result in clinical orders that could have been addressed beforehand. The RALM-based Digital Medical Consultation system will integrate real-time patient-specific data with LLMs to generate structured lists of recommended tests and medications for patients along with generated explanatory consultation notes.
Our multidisciplinary team is uniquely positioned to translate AI applications into clinical practice. The system will be evaluated through iterative testing, validation phases, and human-computer interaction (HCI) usability testing with primary care and specialty physicians. This project will demonstrate that digital consultation tools can empower personalized specialty-level care in the primary care context through discrete evaluation of predicted vs. actual clinical orders as well as human expert evaluation of the quality of generated explanatory consultation text. In so doing, it will increase the efficiency and capacity of an overwhelmed healthcare workforce. Moreover, it will promote access and mitigate disparities in populations historically marginalized by the healthcare system. The products of this project will stretch beyond peer-reviewed publications advancing knowledge in the design of these systems. The results will be shared through peer-reviewed publications, interactive prototypes, and comprehensive clinical knowledge graphs, paving the way for future research and large-scale implementation of AI-driven healthcare solutions. These advances will only become more important with the escalating complexity of medicine, making high-quality, timely, equitable, and economical healthcare more accessible for all. Ultimately, we hope our innovations will fulfill a vision for clinical decision support that empowers our most valuable medical resource, people, to make healthcare more scalable in reach, responsiveness, and reproducibility.
