From an institutional or public health viewpoint, the clinical note is where information goes to die — absent increasingly powerful biomedical informatics solutions for fishing it out and putting it to use (while also protecting patient privacy).
We’re talking progress notes, letters, notes from telephone encounters, history and physical, clinical plans, and assessments.
To improve automated clinical decision support for statin therapy, Siru Liu, PhD, assistant professor of Biomedical Informatics, Adam Wright, PhD, professor of Biomedical Informatics, and colleagues sought to identify patients documented in clinical notes as having intolerance or contraindications to statins or having turned down statin therapy. They developed an efficient three-part artificial intelligence (AI) framework, reporting the project in the International Journal of Medical Informatics.
In a massive test, their framework found documented intolerance in 6.4% of all adult patients, contraindications in 0.7%, and statin deferral in 2.9%.
The framework includes a rule-based natural language processing filter (sensitivity 100%), a large language model (LLM)-based refinement filter (specificity 97.3%), and an LLM-based multicategory classifier. The project used Chat GPT-4o from Open AI. In testing, they fed 197,761 clinical notes from 47,192 adult patients into their framework — a month’s worth at Vanderbilt Health. Filtered notes were manually labeled to allow scoring of the classifier. F1 scores range from zero to 100 and are sensitive to both false positives and false negatives, with 70 to 89 considered good and 90 or above excellent: The classifier scored 99 in classifying intolerant patients, 81 in classifying contraindications, and 86 in classifying patient deferrals. Computation costs were chicken feed.
Liu and Wright, who holds the DBMI Directorship in Clinical Informatics, were joined by Allison McCoy, PhD, and, from Yale University, Qingyu Chen, PhD. The study was supported by the National Institutes of Health under award R00LM014097.
