Abstract 4139043: Artificial intelligence-enabled detection and phenotyping of left ventricular hypertrophy on real-world point-of-care cardiac ultrasonography and its implications for patient outcomes

Introduction: Point-of-care ultrasonography (POCUS) is routinely performed across emergency departments (EDs), but interpretation is generally restricted to acute pathology. We sought to evaluate the outcomes of individuals who had undergone an ED POCUS, but were never diagnosed with cardiomyopathy, using artificial intelligence (AI)-defined signatures of left ventricular hypertrophy (LVH) and key sub-phenotypes (hypertrophic cardiomyopathy [HCM], transthyretin amyloid cardiomyopathy [ATTR-CM], and aortic stenosis [AS]) on POCUS. Methods: First, using 261,756 videos from 9,667 standard transthoracic echocardiograms (TTEs) across a large, diverse health system, we trained a view quality-adapted, video-based deep learning model to detect a) LVH, representing the mean of a multi-label classifier for i) moderate or greater nominal severity as reported by the echocardiographer; ii) left ventricular posterior wall thickness [LVPWd] of ≥1.3 cm, and/or iii) interventricular septum thickness [IVSd] of ≥1.3 cm, and b) known cardiomyopathy defined by non-mutually exclusive labels of i) ATTR-CM, ii) HCM, and/or iii) AS ( Fig. 1A ). We deployed these tools among adult patients without known cardiomyopathy who underwent clinical POCUS across EDs (2013-2023) linked to in-hospital and out-of-hospital death data. We explored the association between distinct label output probabilities and all-cause mortality in age- and sex-adjusted Cox regression models ( Fig. 1B ). Results: Among 24,448 individuals (median age 58, [IQR 40-73] years; 13,478 [55.1%] women) followed over 2.2 [IQR: 1.1-5.8] years, higher AI-POCUS probabilities for LVH were associated with worse long-term prognosis, with a 29% higher mortality risk in the highest vs lowest AI-defined quintile (adj. HR 1.29 [95%CI: 1.13, 1.46]) ( Fig. 2A ). When stratifying based on the probability of distinct phenotypes, an ATTR-CM-like phenotype in the highest (vs lowest) quintile conferred a 39% higher adjusted risk of death (adj. HR 1.39 [95%CI: 1.22, 1.59]) ( Fig. 2B ). Similarly, there was a 14% (adj. HR 1.14 [95%CI: 1.01, 1.30]) and 15% (adj. HR 1.15 [95%CI: 1.02-1.29]) higher risk of death in the highest (vs lowest) AS ( Fig. 2C ) and HCM ( Fig. 2D ) phenotypic quintiles, respectively. Conclusions: AI-enabled automated identification and phenotyping of LVH is feasible on routine POCUS studies and identifies individuals who are at risk of premature mortality, potentially due to undiagnosed cardiomyopathy.