(C-121) Comparison of Primary Care Coded Prevalence of Atrial Fibrillation and Prescribing Rates for Drugs Indicated for Atrial Fibrillation Identified Using a Large-Language Model

Background: Large language model (LLM)-driven medication-indication knowledge bases may offer disease-specific insights from prescribing data. However, it is unclear whether LLM-assisted prescribing rates for specific indications are consistent with reported prevalence.

Objectives: We used atrial fibrillation (AF) as an exemplar condition to compare LLM-assisted prescribing rates against recorded prevalence in primary care, focusing on consistency across English health geography areas.

Methods: We identified medicines indicated for AF in a validated medication-indication knowledge base developed using the GPT-4 LLM and the MedCAT clinical natural language processing model. Using data from a national prescriptions database, we compared prescription rates for the LLM-derived AF medicines (total number of items per 1,000 population) to primary-care-coded AF prevalence reported in a national quality registry (Cardiovascular Disease Prevention Audit) for all areas of England. We presented cross-sectional results for the most recent data available in both data sources (July–September 2024), and then analysed the quarterly prescribing-to-prevalence ratio over September 2021–September 2024 using a linear mixed-effects model with time (fixed effect) and health area (random intercept and random slope).

Results: During July–September 2024, LLM-indicated AF prescribing and AF prevalence showed a non-linear, positive relationship across 42 health areas (R2 = 0.52). At a national level, AF prevalence was 25.8 per 1,000 people (95%CI 25.8, 25.9), and the AF prescribing rate was 178.6 per 1,000 (95%CI 178.5, 178.7). Across 42 health areas, prevalence ranged from 11.2 (95%CI 11.1, 11.4) to 38.0 (95%CI 37.4, 38.5) per 1,000, while prescribing rates ranged from 113.5 (95%CI 113.0, 114.0) to 271.9 (95%CI 270.9, 272.9) per 1,000. The median prescribing-to-prevalence ratio (“ratio”) was 6.69 (IQR 5.74, 7.40). Longitudinally, the linear mixed-effects model showed an average baseline ratio of 7.29 (SE=0.69). The fixed effect for time was -1.6x10-5 (SE=-3.4 x10-5), suggesting a stable trend over the period. Random effects showed moderate variation in intercepts across areas (variance=1.83, SD=1.35) but minimal variation in slopes (variance=0.002, SD=0.045).

Conclusions: Prescribing rates of medicines identified by a LLM as being indicated for AF were broadly consistent with the prevalence of diagnosed AF across geography. AI-driven approach to prescription analytics is potentially scalable for other diseases to detect and monitor population-level health trends.