Differentiating hypertrophy phenocopies from afterload-induced left ventricular hypertrophy using artificial intelligence-enabled electrocardiography
D Ahmetagic, B K O Arends, R R Van De Leur, P Van Der Harst, C Knackstedt, E Biagini, I Ruotolo, P P M Zwetsloot, M Michels, M I F J Oerlemans, R Van EsAbstract
Background
Left ventricular hypertrophy (LVH) is mostly caused due to increased afterload, often caused by hypertension or valvular heart disease. However, it may also be caused by LVH-phenocopies, including hypertrophic cardiomyopathy (HCM), cardiac amyloidosis (CA), or Anderson-Fabry disease (AFD). Distinguishing these etiologies is clinically crucial, as phenocopies require advanced diagnostics to enable appropriate risk stratification, genetic counseling, and timely initiation of disease-specific therapies. While existing artificial intelligence-enabled electrocardiography (AI-ECG) algorithms can detect HCM and CA individually, there remains a clinical need for a tool capable of differentiating these phenocopies from other causes of LVH in broader clinical practice.
Purpose
We aimed to develop and validate a deep-learning algorithm that differentiaties LVH-phenocopies from afterload-associated LVH using 12-lead ECG-data.
Methods
We retrospectively analysed paired ECG and transthoracic echocardiography (TTE) data from patients with afterload-associated LVH and LVH phenocopies at a Dutch academic centre. Additional LVH-phenocopy cases were included from two other Dutch academic centres and one Italian academic centre. A deep convolutional neural network, fine-tuned from the pre-trained ECGFounder model, was trained to discriminate LVH phenocopies from afterload-associated LVH. The model was evaluated in a multicentre cohort of patients with established LVH, defined as a maximal wall thickness ≥11 mm and/or an increased left ventricular mass index. This cohort solely included individuals with severe hypertension (defined by a history of hypertension-mediated organ damage or hypertensive emergency), moderate-to-severe aortic stenosis, and confirmed LVH phenocopies. All available ECGs were used for training, validation, and testing, with a strict per-patient split to prevent data leakage.
Results
The model was trained on 68,075 ECGs from 33,525 patients and validated on 5,625 ECGs from 2,245 patients with LVH (median age 58 [IQR 49–69], 58% male). In the test cohort, 860 patients (38%) had severe hypertension and 680 (30%) had moderate-to-severe aortic stenosis. LVH phenocopies included HCM (n=830, 37%), CA (n=103, 5%), and AFD (n=10, 0.4%). The algorithm achieved an area under the receiver operating characteristic curve (AUC) of 0.85 (95% CI 0.84–0.87) for distinguishing phenocopies from afterload-associated LVH.
Conclusion
Our convolutional neural network accurately distinguishes LVH-phenocopies from afterload-associated LVH within a multicentre framework. Its use may facilitate earlier identification of rare cardiomyopathies and support more efficient, precision-guided diagnostic pathways.Graphical AbstractFor image description, please refer to the figure legend and surrounding text.