AHA Scientific Sessions 2026 Abstract · Nov 2, 2026

High-Yield Echocardiography Triage Using a Longitudinal ECG Transformer Model

By Walker J, Tennyson J, Cheema B et al.

Background: AI-ECG models for structural heart disease are often framed as broad screening tools, but high alert volumes may limit clinical adoption. These tools predominantly use single ECGs for screening. We developed a longitudinal ECG model (i.e., using serial ECGs for a given patient) to identify a small, high-risk group likely to have actionable findings on echocardiography. Methods: Using the Northwestern Medicine Enterprise Data Warehouse, we assembled a retrospective cohort of 6,349 adults with at least two ECGs, each linked to a TTE performed within 90 days. Patients were split into training (N = 4,444), validation (N = 952), and held-out test sets (N = 953). We trained a transformer model using longitudinal ECG waveforms and structured clinical covariates to predict a prespecified actionable TTE composite. The composite was considered positive if the linked TTE demonstrated new or worsening evidence of at least one of the following: newly reduced LVEF ≤40%; further LVEF decline >5 percentage points among patients with prior LVEF ≤40%; moderate or severe valvular disease; pulmonary hypertension; diastolic dysfunction with preserved LVEF plus evidence of elevated filling pressures, left atrial enlargement, or increased LV mass; or moderate or severe LV hypertrophy. Model performance was evaluated in independent validation and held-out test sets using AUROC, positive predictive value (PPV), sensitivity, and specificity. We also calculated number needed to echo (NNE = 1/PPV) to detect one actionable finding at fixed alert rates corresponding to the top 1%, 5%, and 10% of predicted risk. Results: The prevalence of an actionable composite among these patients (mean age 69.9 years, 54.5% female, 12.1% Black) was 26.4% in the validation set and 28.5% in the test set. The model achieved an AUROC of 0.76 (95% CI: 0.73-0.79). Using fixed alert thresholds flagging the top 1%, 5%, and 10% of predicted risk, PPV was 90%, 58%, and 49%, corresponding to NNE of 1.1, 1.7, and 2.0 ( Figure ). Sensitivity at these alert rates was 3%, 10%, and 17%, and specificity was 100%, 97%, and 93%. Conclusion: A transformer model assessing longitudinal change in ECGs identified patients at high likelihood of actionable structural heart disease on TTE, achieving 90% PPV at a 1% fixed alert rate (NNE = 1.1). This low fixed alert rate approach and use of serial ECGs over time may improve diagnostic yield and help target TTE to patients most likely to have clinically actionable findings while maintaining a low alert rate.

doi.org/10.1161/circ.154.suppl_1.6361Biomarkers

AgingResearch.News