An intelligent clinical decision support system for peripheral arterial disease

/ ABSTRACT Unstructured clinician notes contain “80% of medical information” in the electronic health record (EHR) but are cumbersome to search and analyze. Multiple national thought leaders in health care, including the National Academy of Science, Engineering and Medicine have called for improved “access, tools, and capacity for [clinical] data analysis.” Yet, our inability to search, summarize, and extract knowledge from clinical notes remains a major technical impediment to replace the tedious, time-consuming task of human chart review. Large language models (LLMS) are advanced deep learning algorithms capable of the understanding of concepts and context from text. They facilitate several tasks relevant to digital chart review such as advanced information search and retrieval. Furthermore, LLMs can generate coherent prose to form context specific summarizations. Together, these two abilities hold the potential of realizing the vision of a world where most clinical decisions are cost-effectively “supported by accurate, timely, and up-to-date clinical information.” Nevertheless, real world deployment of digital chart review must ensure performance (i.e. accuracy, reliability) and clinicians’ compatibility to facilitate effective and sustainable interactions with clinicians within the context of their clinical workflows. Our research goal is to develop a digital chart review for providing clinicians with accurate and actionable information at the point of care. Our central hypothesis is that a domain adapted LLMs can deliver a decision focused chart summary which is comparable to human chart review by vascular surgeons. We will test this hypothesis using real world data from a large statewide health information exchange on more than 124,000 patients with peripheral arterial disease i.e. PAD (the leading cause of amputation in the US). The mentorship panel developed a set of research activities to execute the following aims that are aligned with the candidate’s overall goal of becoming an independent clinical investigator capable of using advances in health information technology to transform vascular care, and thus the health outcomes of underserved communities living with PAD. In Aim 1 we design and validate a large Language Model capable of extracting and summarizing PAD information from plain text notes. We will evaluate the outputs of the large Language model for accuracy, completeness, relevance, and uncertainty using industry standard metrics. We hypothesize that the LLM text extraction and summarization is comparable to human chart review. In Aim 2, we Co-design with clinicians a minimally viable prototype of a digital chart review system. We will incorporate feedback from the clinicians using the iterative Agile Innovation process. We hypothesize that different clinical scenarios will require emphasis of different information types. Project Number: 1K23HL181388-01 | Fiscal Year: 2025 | NIH Institute/Center: National Heart Lung and Blood Institute (NHLBI) | Principal Investigator: Andrew Gonzalez | Institution: INDIANA UNIVERSITY INDIANAPOLIS, INDIANAPOLIS, IN | Award Amount: $175,049 | Activity Code: K23 | Study Section: NHLBI Mentored Patient-Oriented Research Study Section[MPOR (MA)] View on NIH RePORTER: https://reporter.nih.gov/project-details/1K23HL18138801