Nationwide US federated deployment of artificial intelligence for multimodal screening of ATTR cardiomyopathy: first multicenter analysis from the TRACE network

Background Transthyretin amyloid cardiomyopathy (ATTR-CM) remains underdiagnosed, and reliance on advanced nuclear imaging limits early detection. Given that artificial intelligence (AI) models can detect electrical and mechanical signatures of ATTR-CM from electrocardiograms (AI-ECG) and echocardiograms (AI-Echo), we set up a large, nationwide US network to demonstrate federated deployment of AI-ECG and AI-Echo and to examine the burden of probable undiagnosed ATTR-CM. Purpose To examine the prevalence of probable AI-defined ATTR-CM and assess the generalizability of these models across a diverse network of US-based healthcare systems. Methods We established the largest multicenter AI-driven ATTR-CM discovery network (TRACE Network) in the US with federated deployment of multiple AI tools at 15 US healthcare systems. Individuals aged 50–95 years with an ECG and echocardiogram within 1 year of each other were eligible for inclusion. We developed and distributed format-agnostic software packages for local inference of AI-ECG and AI-Echo. Screen-positive rates at predefined 85% specificity thresholds were examined along with the models' ability to discriminate individuals with diagnosed ATTR-CM from age- and sex-matched controls. The association between screen positivity and incident mortality among those without diagnosed ATTR-CM was assessed using Cox proportional hazards models. Results Seven healthcare systems contributed data to this first TRACE Network analysis, encompassing 203,655 unique individuals (age 70±10.8 years, 105,969 [52.0%] women), 2,432 PYP tests, and 276 individuals with confirmed ATTR-CM. AI-ECG and AI-Echo inference were successfully performed on 201,475 and 6,692 individuals, respectively (Fig. 1). Of these, 9,656 (4.8%) were AI-ECG-positive for probable ATTR-CM, and 428 (6.4%) were AI-Echo-positive. In a pooled cohort analysis, models demonstrated strong performance in discriminating individuals with ATTR-CM against age- and sex-matched controls (AI-ECG: AUC = 0.785 [95% CI: 0.753 – 0.817]; AI-Echo: AUC = 0.823 [95% CI: 0.769 – 0.873]). Models generalized across sites [AUC range: 0.758 – 0.881] despite the different modalities and ECG formats examined. Finally, individuals who screened positive with AI-defined ATTR-CM but had not undergone confirmatory testing for ATTR-CM (n = 9,408) had a significantly increased risk of death compared with age- and sex-matched controls (AI-ECG: adjusted hazard ratio = 1.74 [95% CI: 1.66 – 1.81], p < 0.01; AI-Echo: adjusted hazard ratio = 1.28 [95% CI: 1.05 – 1.57], p = 0.01) (Fig. 2). Conclusion In this largest-ever deployment of AI-ECG and AI-Echo models for opportunistic screening of individuals at risk of ATTR-CM, we demonstrate a large burden of probable undiagnosed ATTR-CM with prognostic implications. Leveraging this framework for screening holds promise for enabling broad, timely identification of patients to maximize the overall benefit of new therapies.