Natural language processing–assisted chart review to assess documented cannabis use in electronic medical records (EMR) of adults undergoing cancer infusion therapy.

Abstract

11125 Background: Patients’ consumption of cannabis for symptom management is well-documented, yet prevalence of cannabis use, and its documentation among patients undergoing active cancer treatment is understudied. Our goal was to assess the prevalence of documented cannabis use in the EMR of patients undergoing infusion cancer therapy. Methods: We conducted a cross-sectional analysis of adult members of Kaiser Permanente Southern California undergoing infusion therapy between February 16 to 28, 2025. Natural language processing (NLP) of search terms indicative of cannabis use was used to identify any EMR documentation of such use in unstructured clinical notes in the past year. Each matched query included 15-20 words before and after the keyword and was manually reviewed for clinical context and confirm cannabis use. Results: We identified N = 690 cancer survivors undergoing infusion therapy during the two-week period. Of these, N = 354 (51%) had a cannabis notation in their EMR after applying the NLP program. Half were female. Altogether 1,010 snippets from N = 354 patients were output into a spreadsheet for manual review; each snippet required about 30 seconds to determine cannabis use status. Of the N = 354 patients, overall prevalence of cannabis use was 29% (N = 101). The highest prevalence was in patients undergoing first line palliative treatment (38%); followed by palliative after first line (25%); adjuvant (16%); then curative first line therapy (13%). Cannabis use was lower in among those undergoing curative after first line therapy (6%) and neoadjuvant therapy (3%). Highest cannabis use was observed in those with cancer of the breast (15%); lung (14%); prostate (10%); colorectal (10%) and gynecologic (8%). Regarding demographics, odds of cannabis use was greater in males (OR = 1.34, 95% CI: 0.82-2.21) vs. females (ref). Compared with Asians (ref), odds of cannabis use was about two-fold greater in Hispanic (OR = 1.95, 95% CI: 0.78-5.35), Black (OR = 2.47, 95% CI: 0.80-7.87), and White (OR = 2.30, 95% CI: 0.95-6.15) patients. Odds of cannabis use was 24% greater in younger individuals ( < 50 years, OR = 1.76, 95% CI: 0.84-3.61), but all confidence intervals crossed the null. Conclusions: NLP-assisted manual review of chart notes is a time-efficient method to collect information on unstructured data. This methodology determined prevalence of cannabis use in patients undergoing cancer infusion therapy is 29%. Cannabis use was greater among patients being treated in palliative settings compared to adjuvant/curative settings. A limitation of this study is that cannabis use may be underestimated because we assume that not all discussions were recorded in the EMR. Information garnered from this cross-sectional data exploration will be used to plan a prospective study of cancer survivors to examine the risks and benefits of cannabis use.