AHA Scientific Sessions 2026 Abstract · Nov 2, 2026
Agentic EHR Phenotyping for Familial Hypercholesterolemia (FH): Development and Implementation of a Local Multi-Agent Language Model System
By Abou Alaiwi S, Ram B, Batinica B et al.
Introduction: FH affects ~1 in 250 individuals and confers increased risk of atherosclerotic cardiovascular disease (ASCVD) yet remains underrecognized. EHR phenotyping relies on structured fields, but key Dutch Lipid Clinic Network (DLCN) criteria (family history (FHx), genetic evidence, and physical exam (PE)) are frequently embedded in clinical narratives, limiting scalable, high-fidelity detection. Hypothesis: An evidence-based, locally deployed multi-agent language-model system can recover DLCN criteria missed by structured EHR data alone and enable scalable, interpretable FH detection. Methods: In a large, diverse US health system, we identified an adult cohort (age ≥18 years) enriched for FH: LDL-C ≥160 mg/dL off lipid-lowering therapy or ≥110 mg/dL on therapy. Structured data included diagnosis codes, labs, medications, demographics, and structured FHx fields. Unstructured notes were processed using a local multi-agent language-model architecture built on Llama 3.1 via Ollama. Three domain-specialized agents extracted DLCN-domain evidence: FHx of premature ASCVD or hypercholesterolemia; FH-related genetic testing (LDLR, APOB, PCSK9, or LDLRAP1); and PE findings. Evidence was mapped to DLCN criteria and scored by structured-only and combined approaches, classifying patients (pts) as unlikely FH (<3), possible FH (3-5), probable FH (6-8), or definite FH (>8). Results: We identified 135,014 pts with available LDL-C: mean age, 64.9 ± 15.1 years; 10.5% Hispanic/Latino; and 12.1% Black/African American. Among 10.7 million notes, 4.3 million were prioritized by author and note type. Compared with diagnosis codes, agents identified substantially more PE stigmata: corneal arcus, 626 vs 61; tendon xanthoma, 454 vs 73; and xanthelasma, 789 vs 222. The genetic-testing agent identified 34 pathogenic or likely pathogenic monogenic variants, all confirmed by expert chart review. The combined approach reclassified 408 pts upward, increasing definite FH from 92 to 187 pts and probable FH from 1,059 to 1,328 pts. Reclassification was driven primarily by PE stigmata (349 of 408), with smaller contributions from FHx (44) and genetic data (25); 10 pts were upgraded by more than one agent-derived domain. Across increasing DLCN categories, pts were progressively younger and had greater ASCVD burden: any ASCVD, 42% to 76%; premature coronary disease, 3% to 48%. Despite lipid-lowering therapy in 96% and high-intensity statin use in 62%, pts with definite FH had a median LDL-C of 157 mg/dL, and only 17% had LDL-C <100 mg/dL, indicating marked undertreatment. Conclusion: A local, interpretable multi-agent language-model system transformed longitudinal clinical narrative into computable DLCN evidence, reclassified pts into higher-confidence FH categories, and exposed an undertreated, high-risk population suitable for targeted outreach and lipid-lowering intensification.