AI-Identified Reasons for Nonescalation of Lipid-Lowering Therapy for Secondary Prevention

To determine whether large language models (LLMs) can identify a treat-to-target strategy use and extract reasons for LLT nonescalation from clinical notes when LDL-C concentration is above goal.

This was a retrospective cohort study of patients from January 1, 2019, through May 15, 2025. Notes from most recent clinical encounters were retrieved and split into development, test, and inference sets.

The setting was a large, integrated health system in Northern California. Adults aged 18 to 90 years with 2 or more clinician-billed encounters for ASCVD and with LDL-C level greater than or equal to 70 mg/dL at the most recent encounter were included in note-based analyses.

Model performance was evaluated on LDL-C target documentation, LDL-C target value, LLT adjustment, and reasons for nonescalation against clinical annotations. Outcome prevalence was assessed on the inference set.

A total of 28 113 unique patients with ASCVD-related encounters were identified, of whom 13 810 (49.1%; mean [SD] age, 70.9 [11.4] years; 7904 male [57.2%]) had above-goal LDL-C level. Model-annotator and interannotator reliability were not statistically different across outcomes, and micro-weighted F1 scores ranged from 0.80 to 0.98.

Of the 13 394 patients with above-goal LDL-C level and 13 354 relevant notes in the inference set, a treat-to-target strategy was identified in 3756 (28.1%) of all relevant notes, most commonly targeting an LDL-C level less than 70 mg/dL (2280 [a clinically meaningful level]). LLT was not escalated in 11 842 of all relevant notes (88.7%), with 4788 (40.4%) containing at least 1 LLT nonescalation reason, most often due to patient-reported adverse effects (1683 [12.6%]), LDL-C level close to goal (922 [6.9%]), patient preference to avoid LLT (914 [6.8%]), preference for lifestyle changes (811 [6.1%]), and deferral to repeat laboratory tests (617 [4.6%]).

Results of this study show that after using a state-of-the art LLM, nearly 9 in 10 ASCVD encounters with above-goal LDL-C level did not have patient LLT escalated, most often because of adverse effects, patient preference, preference for lifestyle changes, therapeutic inertia (eg, repeat laboratory tests), and near-goal LDL-C level.

These findings highlight an avenue for personalizing system-level digital health interventions to help discover and minimize guideline-discordant care gaps in secondary prevention of ASCVD.