Background: Nearly a third of people using cannabis meet criteria for cannabis use disorder based on ICD-10 codes (CUD). The prevalence of CUD among people with HIV (PWH) is 4-times that of people without HIV (PWoH). Yet, cannabis use is understudied and not routinely collected in healthcare settings. Our objective was to improve methods of identifying cannabis use. Methods: We used data from Veterans Aging Cohort Study (VACS), a well characterized, national cohort with electronic health record data on PWH and demographically similar PWoH. We implemented a natural language processing (NLP) algorithm on unstructured clinical notes for over 160,000 people to identify cannabis use. Using a publicly available lexicon, initial results were fine-tuned by filtering relevant key-terms on notes from years 2000 to 2025. To further validate our search, we compared NLP model prediction with manual chart review. A sub-study, Medications, Alcohol, and Substance use (MASH) collected self-report cannabis use and biomarker data for 531 PWH from VACS. We also compared biomarker data to text search results. Two reviewers performed manual chart reviews to compare output of text search with chart review results. Results: The NLP was initially run on 100K notes, we validated 200 notes with manual annotation with an accuracy of 92%, precision of 93% and F1-score of 95%. Of 531 PWH in MASH, 81% reported ever using cannabis. Twenty-nine percent self-reported cannabis use at least monthly or had positive cannabis biomarker; agreement between self-report and biomarker was excellent (kappa = 0.72). Only 10% were identified with CUD based on ICD-10 codes. Over 600 snippets were created from 26,905 notes on 102 individuals. Based on text search results, 31% had current use and 30% past use. Of those with text indicating current use, 47% had CUD. Of the 102, compared to biomarker, sensitivity was 88% for self-report, 44% for CUD, 69% for text indicating current cannabis use, and 75% for combined measure of text and CUD. The NLP algorithm on the entire cohort run over a million notes was validated with a 85% accuracy by chart review. Implications: Improved identification of cannabis use with NLP in healthcare settings is needed to better understand impact of cannabis use and CUD on overall health and disease management. We propose cannabis use data collection in primary care settings and leveraging existing data to understand long term implications of CUD.
Dr. Kidwai-Khan's is a senior Data Scientist whose work involves research related to applying theoretical data engineering and computer science knowledge to develop algorithms related to epidemiological and clinical research. This includes developing data modeling processes to create predictive modeling for tailored analysis related to epidemiological and clinical studies and development of complex databases and informatics tools in the domain of clinical and health services research. Dr. Kidwai-Khan received a Doctorate in Engineering from GWU.