DECODING DISCORDANT CARDIOVASCULAR OUTCOMES IN ATRIAL FIBRILLATION AND AUTOIMMUNE DISEASE: AN NLP AND MACHINE LEARNING APPROACH
Author(s)
Yue Cheng, MSc1, Jessica Walsh, MD2, Siyuan Ye, Undergraduate Student3, Daniel Witt, PharmD1, Xiangyang Ye, PhD1;
1University of Utah, Department of Pharmacotherapy, Salt Lake City, UT, USA, 2University of Utah, Department of Internal Medicine, Salt Lake City, UT, USA, 3University of Washington, Information School, Seattle, WA, USA
1University of Utah, Department of Pharmacotherapy, Salt Lake City, UT, USA, 2University of Utah, Department of Internal Medicine, Salt Lake City, UT, USA, 3University of Washington, Information School, Seattle, WA, USA
OBJECTIVES: This study aimed to identify clinical and treatment factors associated with unexpected major adverse cardiovascular events (MACE) in patients with atrial fibrillation (AFib) and autoimmune disease (AID).
METHODS: Using ICD codes, 2,887 patients (median follow-up: 4.6 years) with AFib and AID were identified from the University of Utah Electronic Health Record Systems. MACE was defined as heart failure, stroke, acute MI, unstable angina, or death. Risk was assessed via a consensus of 13 models (5 machine learning models, 8 validated scores, e.g., ASCVD, Reynolds, QRISK3). Medications were mapped to 60 treatment categories based on therapeutic classes in Medication Reference Terminology. A customized pipeline was developed to integrate structured and unstructured data using Natural Language Processing (NLP) and large language models to analyze contributing and protective factors in medical records.
RESULTS: The cohort (mean age 67.9 ± 14.0; 44.9% male; 88.1% White; 4.2% Hispanic) included 1,073 (37.2%) MACE events. While aggregated treatment rates were similar (26.0% MACE vs. 25.7% event-free), statistically significant differences (p < 0.05) occurred in 36% of therapeutic classes. The MACE group exhibited markedly higher use of anti-anginal agents (61.4% vs. 19.6%) and cardiovascular agents (95.0% vs. 71.5%), while the event-free group showed significantly higher treatment proportions for antidepressive agents (41.8% vs. 18.8%) and anticonvulsants (29.7% vs. 9.9%). Among "discordant" cohorts, 4 out of 158 (2.5%) low-predicted-risk patients experienced MACE, while 31 out of 101 (30.7%) high-predicted-risk patients remained event-free. The discordance between predicted risk and actual MACE was primarily driven by treatment regimens dominated by age and comorbidities.
CONCLUSIONS: Medication profiles revealed distinct clinical phenotypes in risk-discordant patients. Unexpected MACE cases demonstrated intensive cardiovascular treatment patterns, while event-free high-risk patients showed higher neuropsychiatric medication use. This exploratory work suggests that NLP-enhanced machine learning may improve phenotype identification in complex AFib patients with AID. Future large-scale studies are needed for validation.
METHODS: Using ICD codes, 2,887 patients (median follow-up: 4.6 years) with AFib and AID were identified from the University of Utah Electronic Health Record Systems. MACE was defined as heart failure, stroke, acute MI, unstable angina, or death. Risk was assessed via a consensus of 13 models (5 machine learning models, 8 validated scores, e.g., ASCVD, Reynolds, QRISK3). Medications were mapped to 60 treatment categories based on therapeutic classes in Medication Reference Terminology. A customized pipeline was developed to integrate structured and unstructured data using Natural Language Processing (NLP) and large language models to analyze contributing and protective factors in medical records.
RESULTS: The cohort (mean age 67.9 ± 14.0; 44.9% male; 88.1% White; 4.2% Hispanic) included 1,073 (37.2%) MACE events. While aggregated treatment rates were similar (26.0% MACE vs. 25.7% event-free), statistically significant differences (p < 0.05) occurred in 36% of therapeutic classes. The MACE group exhibited markedly higher use of anti-anginal agents (61.4% vs. 19.6%) and cardiovascular agents (95.0% vs. 71.5%), while the event-free group showed significantly higher treatment proportions for antidepressive agents (41.8% vs. 18.8%) and anticonvulsants (29.7% vs. 9.9%). Among "discordant" cohorts, 4 out of 158 (2.5%) low-predicted-risk patients experienced MACE, while 31 out of 101 (30.7%) high-predicted-risk patients remained event-free. The discordance between predicted risk and actual MACE was primarily driven by treatment regimens dominated by age and comorbidities.
CONCLUSIONS: Medication profiles revealed distinct clinical phenotypes in risk-discordant patients. Unexpected MACE cases demonstrated intensive cardiovascular treatment patterns, while event-free high-risk patients showed higher neuropsychiatric medication use. This exploratory work suggests that NLP-enhanced machine learning may improve phenotype identification in complex AFib patients with AID. Future large-scale studies are needed for validation.
Conference/Value in Health Info
2026-05, ISPOR 2026, Philadelphia, PA, USA
Value in Health, Volume 29, Issue S6
Code
CO164
Topic
Clinical Outcomes
Topic Subcategory
Comparative Effectiveness or Efficacy
Disease
SDC: Cardiovascular Disorders (including MI, Stroke, Circulatory), SDC: Systemic Disorders/Conditions (Anesthesia, Auto-Immune Disorders (n.e.c.), Hematological Disorders (non-oncologic), Pain)