Artificial intelligence-guided detection of under-recognised cardiomyopathies on point-of-care cardiac ultrasonography: a multicentre study

Abstract

Background: Point-of-care ultrasonography (POCUS) enables cardiac imaging at the bedside and in communities but is limited by abbreviated protocols and variation in quality. We aimed to develop and test artificial intelligence (AI) models to screen for under-diagnosed cardiomyopathies from cardiac POCUS.

Methods: In a development set of 290 245 transthoracic echocardiographic videos across the Yale-New Haven Health System (YNHHS), we used augmentation approaches, and a customised loss function weighted for view quality to derive a POCUS-adapted, multi-label, video-based convolutional neural network that discriminates hypertrophic cardiomyopathy and transthyretin amyloid cardiomyopathy from controls without known disease. We evaluated the model across independent, internal, and external, retrospective cohorts of individuals undergoing cardiac POCUS across YNHHS and the Mount Sinai Health System (MSHS) emergency departments (between 2012 and 2024) to prioritise key views and validate the diagnostic and prognostic performance of single-view screening protocols.

Findings: Between Nov 1, 2023, and March 28, 2024, we identified 33 127 patients (mean age 58·9 [SD 20·5] years, 17 276 [52·2%] were female, 14 923 [45·0%] were male, and for 928 [2·8%] sex was recorded as unknown) at YNHHS and 5624 patients (mean age 56·0 [20·5] years, 1953 [34·7%] were female, 2470 [43·9%] were male, and for 1201 [21·4%] sex was recorded as unknown) at MSHS with 78 054 and 13 796 eligible cardiac POCUS videos, respectively. AI deployed to single-view POCUS videos successfully discriminated hypertrophic cardiomyopathy (eg, area under the receiver operating characteristic curve 0·903 [95% CI 0·795-0·981] in YNHHS; 0·890 [0·839-0·938] in MSHS for apical-4-chamber acquisitions) and transthyretin amyloid cardiomyopathy (0·907 [0·874-0·932] in YNHHS; 0·972 [0·959-0·983] in MSHS for parasternal acquisitions). In YNHHS, 40 (58%) of 69 hypertrophic cardiomyopathy cases and 22 (46%) of 48 transthyretin amyloid cardiomyopathy cases would have had a positive screen by AI-POCUS at a median of 2·1 (IQR 0·9-4·5) years and 1·9 (0·6-3·5) years before diagnosis. Moreover, among 25 261 participants without known cardiomyopathy followed up over a median of 2·8 (1·2-6·4) years, AI-POCUS probabilities in the highest (vs lowest) quintile for hypertrophic cardiomyopathy and transthyretin amyloid cardiomyopathy conferred a 17% (adjusted hazard ratio 1·17, 95% CI 1·06-1·29; p=0·0022) and 32% (1·39, 1·19-1·46; p<0·0001) higher adjusted mortality risk, respectively.

Interpretation: We developed and validated an AI framework that enables scalable, opportunistic screening of under-recognised cardiomyopathies through simple POCUS acquisitions.

Funding: National Heart, Lung, and Blood Institute, Doris Duke Charitable Foundation, and BridgeBio.

Figure 1:. Study overview

Figure 2:. Video-level performance of a view-agnostic multi-label POCUS classifier in the Yale–New Haven Health System

Video-level performance (AUROC curve with 95% CIs) for discrimination of hypertrophic cardiomyopathy, transthyretin amyloid cardiomyopathy, and severe aortic stenosis, by deploying a POCUS-adapted model to different echocardiographic views obtained across the emergency departments of the Yale–New Haven Health System. Results for all patients (A, B) and for those without known heart failure at the time of their assessment (C, D) are presented, further stratified by the confidence of the automatic view classifier in detecting the corresponding view (all videos [A, C] vs view confidence probability of ≥0·5 [B, D]). The numbers at the bottom of each bar denote the counts of cases out of all eligible video counts in this group. All 95% CIs are derived from bootstrapping with 1000 replications. AUROC=area under the receiver operating characteristic. POCUS=point-of-care ultrasonography.

Figure 3:. Saliency maps

Activation maps for hypertrophic cardiomyopathy (A, B), transthyretin amyloid cardiomyopathy (C, D), and severe aortic stenosis (E, F) across parasternal long-axis and apical-4-chamber views obtained at the point of care in the emergency department. The colour scale denotes the relative importance of different areas, averaged across time, for each individual label.

Figure 4:. Density plot of time between positive artificial intelligence-POCUS screen and eventual confirmatory testing

Density plot summarising the time difference between a positive POCUS screen and confirmatory testing by cardiac magnetic resonance or nuclear cardiac amyloid testing for 40 patients and 23 patients with an eventual diagnosis of hypertrophic cardiomyopathy or transthyretin amyloid cardiomyopathy, respectively. POCUS=point-of-care ultrasonography.

Figure 5:. Hypertrophic cardiomyopathy and transthyretin amyloid cardiomyopathy-specific probabilities and overall survival among individuals without documented cardiomyopathy in the YNHHS

Kaplan–Meier survival curves across quintiles (Q1–5) of the probabilities for hypertrophic cardiomyopathy (A) and transthyretin amyloid cardiomyopathy (B) on point-of-care ultrasonography in the YNHHS. Results are presented for n=25 261 eligible individuals who were never diagnosed with cardiomyopathy during the follow-up period (median of 2·8 [IQR 1·2–6·4] years). We report the p values for comparison of the Kaplan–Meier curves, and also Cox regression-derived HRs (95% CIs) adjusted for age, sex, hypertension, diabetes, ischaemic heart disease, chronic kidney disease, and peripheral arterial disease. HR=hazard ratio. Q=quintile. YNHHS=Yale–New Haven Health System.