Targeted deployment of AI-ECG for efficient screening of transthyretin amyloid cardiomyopathy using deep learning representations of longitudinal electronic health records

Abstract Background Transthyretin amyloid cardiomyopathy (ATTR-CM) remains largely underdiagnosed despite the availability of novel disease-modifying therapies. Artificial intelligence-enabled electrocardiography (AI-ECG) has shown potential in the screening of ATTR-CM, but its optimal deployment across large health systems remains unclear. Deep learning-enabled phenotypes derived from longitudinal electronic health records (EHR) may enable the targeted deployment of AI-ECG, minimizing false discovery and unnecessary downstream testing. Purpose To define and validate a multimodal strategy that uses longitudinal EHR representations followed by targeted AI-ECG deployment to enable high-precision, high-throughput phenotyping of ATTR-CM across a large and diverse U.S.-based health system. Methods We developed a multimodal pipeline that integrates longitudinal representations from an individual’s EHR, spanning demographics, diagnoses, encounters, laboratory measurements, medications, and procedures, to define optimal intersections of clinical phenotypes and healthcare encounters that are associated with the highest precision for AI-ECG-guided screening of ATTR-CM (Fig. 1A-B). In a development set of 724,426 ECGs from 150,390 individuals seeking care across a large U.S.-based health system between 2013 and 2021, we derived reference EHR representations for ATTR-CM cases (1,675 0.2% ECGs from 188 0.1% individuals) versus controls without known disease. In a temporally distinct cohort from 2022-2023, we evaluated whether targeted sequential screening by deep learning applied to EHR and AI-ECG improved the precision of system-wide screening vs the opportunistic deployment of AI-ECG across all unselected individuals. Results Our temporally distinct test set consisted of 48,373 ECGs from 17,074 unique individuals (median age 70 IQR: 59-79 years, 54.6% men, 15.9% Black). Among these, 127 (0.3%) ECGs from 49 (0.3%) were linked to abnormal cardiac nuclear amyloid imaging consistent with ATTR-CM (Fig. 1C). Based on optimal decision thresholds derived in the training set (90% sensitivity for EHR, and maximal F1-score for AI-ECG), we observed that targeted, vs opportunistic (untargeted), deployment of AI-ECG was associated with a 62.4% 95%CI: 54.6%-70.8% reduction in positive screens (from 133 to 50), with PPV increasing by 2.3-fold (from 0.05 to 0.12; delta 0.07 95%CI: 0.01-0.13) and NPV remaining relatively unchanged at 0.997 (from 0.998). Conclusion Our proposed strategy defines an optimal implementation path for the deployment of AI-ECG-guided ATTR-CM screening across large health systems, leveraging highly interoperable, deep learning-enabled longitudinal EHR representations. This may improve the cost-effectiveness of AI-ECG screening and prevent unnecessary downstream testing from system-wide screening.Figure 1