Adaptive heterogeneous graph neural networks for differentiating pulmonary arterial hypertension from left heart disease

Abstract

Abstract Differentiating pulmonary arterial hypertension (PAH) from pulmonary hypertension associated with left heart disease (PH-LHD) is clinically important because management differs substantially, but definitive classification requires invasive haemodynamic assessment together with clinical evaluation. We retrospectively studied 905 patients with PAH or PH-LHD treated at Shanghai Pulmonary Hospital and developed an adaptive heterogeneous graph neural network (AHGNN) that integrates contrast-enhanced thoracic CT images with noninvasive clinical variables. The model combines differentiable graph construction using Gumbel-Softmax reparameterization, hierarchical cross-modal attention, and dual-level self-supervised contrastive learning. Across 100 outer test folds from repeated patient-level five-fold cross-validation, AHGNN achieved a mean area under the receiver operating characteristic curve (AUC) of 0.946 ± 0.023 and a precision–recall AUC of 0.952 ± 0.009. At a fixed probability threshold of 0.50, sensitivity was 0.867 ± 0.070 and specificity was 0.867 ± 0.072. The Brier score was 0.139, although a calibration slope of 3.547 indicated that recalibration may be required. These findings support the potential of multimodal noninvasive data to assist referral and diagnostic triage. Prospective external validation, recalibration, and clinical safety evaluation are required before clinical use or any change to the role of right heart catheterization.