Abstract 4365660: Multimodal Artificial Intelligence Improves the Yield of Nuclear Cardiac Amyloid Testing in Suspected Transthyretin Amyloid Cardiomyopathy: a Report from the TRACE-AI Network

Abstract

Introduction: Artificial intelligence (AI) applied to single cardiovascular diagnostic modalities is an emerging approach for the scalable screening of transthyretin amyloid cardiomyopathy (ATTR-CM) offering the promise of timely detection and intervention. However, broad clinical deployment strategies may be limited by high rates of false positive screens. Hypothesis: We hypothesize that a multimodal approach that jointly evaluates AI-detectable electrocardiographic (AI-ECG) and echocardiographic (AI-Echo) phenotypes will improve the precision of AI-guided ATTR-CM screening as confirmed on nuclear cardiac amyloid testing. Methods: This was a retrospective analysis of 1,165 unique patients referred for nuclear cardiac amyloid testing for suspicion of ATTR-CM across two independent health systems participating in the TRACE-AI Network Study (Yale-New Haven Health System [YNHHS] and the Houston Methodist Hospitals [HMH]). We retrieved the last 12-lead ECG and transthoracic echocardiogram (TTE) performed in the year before nuclear cardiac amyloid testing and deployed previously validated models trained to discriminate ATTR-CM from age/sex-matched controls. We evaluated the diagnostic performance of unimodal screening strategies using a) AI-ECG alone, or b) AI-Echo alone vs c) a multimodal, joint AI-ECG/AI-Echo strategy. Results: Our study included 656 (73±12 years, 307 [46.8%] female, 50 [7.6%] ATTR-CM) and 509 (70±13 years, 188 [36.9%] female, 96 [18.9%] ATTR-CM) individuals who were referred for nuclear cardiac amyloid testing across YNHHS and HMH, respectively ( Fig. 1a ). At validated thresholds (≥0.15), 314 (47.9%) and 323 (63.5%) individuals screened positive on at least one modality and 69 (10.5%) and 74 (14.5%) screened positive on both modalities across YNHHS and HMH, respectively ( Fig. 1b ). Double positivity on multimodal screening resulted in specificity of 0.93 to 0.94 and positive predictive value of 0.36 to 0.66, across sites. At the same probability threshold, this translated into a 78.3% [95%CI 72.7%-83.6%] (YNHHS) to 88.5% [95%CI: 83.2%-92.7%] (HMH) reduction in false positives vs AI-ECG alone, and 51.1% [95%CI: 38.0%-65.1%] (HMH) to 60.3% [95%CI: 52.2%-67.0%] (YNHHS) reduction in false positives vs AI-Echo alone ( Fig. 2 ). Conclusion: Multimodal AI-enabled screening of suspected ATTR-CM significantly decreases false positive screens from unimodal AI models and represents a promising strategy for system-wide screening programs.

Ähnliche Arbeiten