Deep Learning-Enhanced Noninvasive Detection of Pulmonary Hypertension and Subtypes via Chest Radiographs, Validated...

Area of Science:

• Artificial Intelligence in Medical Imaging
• Cardiology
• Radiology

Background:

• Pulmonary hypertension (PH) is a critical condition needing accessible diagnostic tools.
• Early PH detection is vital for management and intervention, especially in resource-limited settings.

Purpose of the Study:

• To assess the accuracy of deep learning (DL) models in detecting PH and CHD-PAH using chest radiography (CXR).

Main Methods:

• Developed and validated DL models (CXR-PH-Net, CXR-CHD-PAH-Net) on a large retrospective cohort (4,576 patients).
• Internal and external validation datasets included RHC-confirmed cohorts.
• Performance evaluated using AUC, sensitivity, and specificity.

Main Results:

• CXR-PH-Net achieved high sensitivity (0.902) and AUC (0.964) for PH detection.
• CXR-CHD-PAH-Net demonstrated strong performance for CHD-PAH detection (AUCs up to 0.908).
• Models showed good sensitivity for detecting CHD-PAH in mild PH cases.

Conclusions:

• DL models using CXR show promise as sensitive screening tools for PH and CHD-PAH.
• These models could facilitate early detection and triage, particularly in underserved regions.
• Further validation in diverse populations is recommended for broader clinical application.