Targeted AI-ECG screening for structural heart disease reduced false positives by 87.8% and improved F1-score by 0.25 median compared to untargeted screening.
Does targeted deployment of AI-ECG guided by EHR data improve diagnostic performance and reduce false positives for structural heart disease screening compared to untargeted deployment?
Using longitudinal EHR data to guide targeted AI-ECG screening for structural heart disease significantly improves diagnostic precision and reduces false positive rates compared to untargeted screening.
Absolute Event Rate: 0% vs 0%
Abstract Background Artificial intelligence (AI) can enable scalable screening of structural heart disease (SHD) from routine electrocardiograms (ECGs), yet broad clinical adoption remains limited due to high false positive rates and a lack of targeted deployment strategies. Purpose To develop and validate a multi-modal approach that leverages existing patient data in the electronic health record (EHR) to guide the targeted deployment of AI-ECG for SHD screening. Methods Our development dataset included 159,322 individuals (median age 68 IQR: 57-78 years, 50.4% women) across a large U.S.-based health system between (2013-2021), with 118 million coded EHR events, as well as 754,533 pairs of temporally linked ECG images and echocardiography reports (≤90 days between studies). We designed an algorithmic pipeline that leverages joint EHR and ECG embeddings for targeted SHD screening. First, we used a validated foundation model (CLMBR-T) to produce longitudinal EHR representations, which we fine-tuned to identify signatures of 27 SHDs, optimized to a sensitivity of ≥90%. This EHR representation represents the population with a high pre-test probability of these SHDs (Fig. 1a). Next, we used an ECG image vision transformer (ViT) developed using contrastive language-image pre-training against linked echocardiogram reports, optimized to detect SHDs with a balance of precision and recall (Fig. 1b). This targeted (sequential) screening approach was compared against an untargeted approach of opportunistically deploying AI-ECG across all individuals in i) a temporally distinct dataset of 5,198 individuals who had their first TTE in 2022-2023 within 90 days after their ECG, and ii) in a geographically distinct test set of 33,518 individuals enrolled in the UK Biobank with concurrent ECG and cardiac magnetic resonance imaging. Results Using training embeddings as reference, our foundational ECG image encoder successfully discriminated a representative sample of 27 SHD labels during testing, including left ventricular systolic dysfunction (AUROC of 0.90), and severe aortic stenosis (AUROC 0.85), among others (Fig. 2a). Targeted (vs opportunistic) deployment of AI-ECG led to a median 0.25 IQR: 0.19-0.43 absolute increase in F1-score, which corresponded to a 87.8% IQR: 82.4%-98.2% relative decrease in false positive screens across the population (Fig. 2b). These gains were replicated in a geographically distinct test set, with a 36.8% median IQR 16.3%-54.4% relative increase in F1-scores, when examining AI-ECG performance in an enriched versus untargeted ("screen-all") population. Conclusion Using a deep learning representation of existing longitudinal EHR data can guide the efficient use of AI to screen for SHD from routinely available ECG images, thus minimizing false discovery and the possibility of unnecessary downstream testing.Figure 1. Figure 2.
Oikonomou et al. (Sat,) reported a other. Targeted AI-ECG screening for structural heart disease reduced false positives by 87.8% and improved F1-score by 0.25 median compared to untargeted screening.