Single-view echocardiography-based deep learning for pulmonary hypertension detection: a multi-level supervised and explainable approach

Background: Early identification of pulmonary hypertension (PH), particularly mild disease, remains challenging. We aimed to develop and validate a deep learning (DL) model for noninvasive PH screening using single-view transthoracic echocardiography (TTE), without reliance on Doppler measurements.

Methods: In this retrospective study, apical four-chamber (A4C) TTE cine loops were collected between January 2023 and June 2025. A total of 135 patients with PH confirmed by right heart catheterization (RHC), 17 RHC-confirmed non-PH patients, and 95 non-PH controls defined by TTE under strict inclusion criteria were included. An additional temporally independent validation cohort of 40 RHC-confirmed subjects was enrolled between December 2025 and May 2026. Multiple DL architectures were evaluated for detecting PH. Model performance was assessed at both sample and patient levels using the area under the receiver operating characteristic curve (AUROC), sensitivity, and specificity.

Results: A total of 256 patients with 318 TTE examinations were included for model development and independent validation, and an additional 40 RHC-confirmed subjects with 40 TTE examinations were included for temporal validation. The multi-level supervised model achieved the best performance among the evaluated architectures, with a patient-level AUROC of 1.000 [95% confidence interval (CI): 0.998–1.000], sensitivity of 0.954, and specificity of 1.000 in the independent validation cohort. In the temporal validation cohort, the model achieved an AUROC of 0.932 (95% CI: 0.843–0.992), with sensitivity of 0.875, specificity of 0.813, and accuracy of 0.850. Among 50 RHC-confirmed PH patients, the DL model achieved an AUROC of 0.937 (95% CI: 0.865–0.987) for distinguishing mild PH from non-PH, with sensitivity of 0.857 and specificity of 0.864, compared with AUROCs of 0.607–0.825 for conventional echocardiographic parameters.

Conclusions: We developed an interpretable DL framework based on single-view A4C echocardiography for noninvasive PH assessment. The model showed promising performance, including preserved sensitivity for mild PH. Further multicenter prospective validation is required before clinical implementation.