Why the study?
Major HF registries are limited by manual chart abstraction and restriction to hospitalized patients.
Does an AI-based phenotyping approach identify a comparable or larger cohort of heart failure patients compared to an ICD code-based approach in the VA healthcare system?
Cohort (n=1,416,512)
Yes
Does an AI-based phenotyping approach identify a comparable or larger cohort of heart failure patients compared to an ICD code-based approach in the VA healthcare system?
Answers are grounded in this paper and cite their sources.
An AI-based phenotyping approach identified a substantially larger cohort of heart failure patients in the VA system compared to traditional ICD coding, with broadly consistent clinical characteristics.
BACKGROUND: Major heart failure (HF) registries are limited by manual chart abstraction and to hospitalized patients. Using artificial intelligence (AI) approaches and electronic health record (EHR) data from the Veterans Affairs (VA) healthcare system, we have assembled the National DCVA-HF Registry, which includes both ambulatory and hospitalized HF patients. In the current study, we descriptively compared these patients with those identified using International Classification of Diseases (ICD) codes. METHODS: We identified 1,416,512 Veterans with at least one ICD code for HF in the VA national EHR from 1999 to 2017. The first date of an ICD code for HF was considered the index HF date. We used validated encounter-based AI approaches based on machine learning and natural language processing models, and ICD code-based approaches based on ≥1 hospitalization or ≥2 outpatient encounters due to HF, to assemble the AI-HF and ICD-HF cohorts, respectively. The two cohorts were compared using absolute standardized differences, with values ≥10% indicating clinical significance. All analyses were descriptive, and no inferential or causal claims were made. RESULTS: The AI and ICD approaches assembled 1,031,970 and 614,828 patients, respectively, after a mean of 0.4 and 1.0 years from the index HF date. Patients in the AI-HF (vs. ICD-HF) cohort had a mean age of 71.4 (vs. 70.5) years, 98.1% (vs. 98.0%) were men, 13.9% (vs. 16.1%) were African American, and 7.7% (vs. 13.4%) had index HF hospitalization, with absolute standardized differences of 8%, 1%, 6%, and 18%, respectively, which were <10% for 66 other baseline characteristics. One-year post-index HF hospitalization occurred in 10.7% and 15.6% of patients in the AI-HF and ICD-HF cohorts, respectively. One‑year mortality was lower in the ICD-HF cohort, reflecting expected immortal‑time bias due to later cohort qualification. CONCLUSIONS: The findings from this descriptive study demonstrate that the AI approach assembled a substantially larger cohort that included most patients identified using the ICD code approach, suggesting broad consistency between the two approaches. Future external validation is needed to determine its potential utility as a robust tool for improving patient care, health services operations, and clinical research in HF.
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2026) conducted a cohort in Heart failure (n=1,416,512). Artificial intelligence (AI) phenotyping approach vs. ICD code-based approach was evaluated on 1-year post-index HF hospitalization.
Synapse has enriched 10 related papers. Consider them for comparative context: