Huang Zhihua, Diao Xiaolin, Huo Yanni, Zhao Zhihui, Geng Jiahui, Zhao Qing, Liu Jia, Xi Qunying, Xia Yun, Xu Ou, Li Xin, Duan Anqi, Zhang Sicheng, Gao Luyang, Wang Yijia, Li Sicong, Luo Qin, Liu Zhihong, Zhao Wei
Center for Respiratory and Pulmonary Vascular Diseases, Department of Cardiology, Fuwai Hospital, National Clinical Research Center for Cardiovascular Diseases, National Center for Cardiovascular Diseases, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing.
Department of Information Center, Fuwai Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing.
Chest. 2025 Jun 18. doi: 10.1016/j.chest.2025.06.008.
Pulmonary hypertension (PH) is a complex, life-threatening condition requiring noninvasive, accessible, and accurate diagnostic tools, particularly in resource-limited settings. Early and precise identification of PH and its subtypes is critical for effective management and timely intervention.
Can deep learning (DL) methods applied to chest radiography (CXR) accurately detect PH and its subtype, congenital heart disease-associated pulmonary arterial hypertension (CHD-PAH)?
A retrospective cohort study was conducted with 4,576 patients, including 2,288 patients with PH, who underwent CXR followed by right heart catheterization (RHC) or transthoracic echocardiography. DL models were developed and validated for detecting PH (CXR-PH-Net model) and CHD-PAH (CXR-CHD-PAH-Net model). Internal testing used a data set of 2,140 patients (1,070 patients with PH), and additional validation included an RHC-confirmed internal cohort (1,158 patients) and an external RHC cohort (90 patients) from 2 independent hospitals. Model performance was evaluated primarily using the area under the receiver operating characteristic curve (AUC), sensitivity, and specificity.
The CXR-PH-Net model achieved a sensitivity of 0.902 and an AUC of 0.964 for PH detection in the internal test set. In the RHC-confirmed cohort, sensitivity was 0.902 (AUC, 0.872) internally and 0.803 (AUC, 0.811) externally. The CXR-CHD-PAH-Net model demonstrated sensitivities of 0.859 and 0.870 with AUCs of 0.908 and 0.860 in the internal and external data sets, respectively. Meanwhile, the CXR-CHD-PAH-Net model showed favorable sensitivity in detecting CHD-PAH among patients with mild PH, with values of 0.813 and 0.846 in the internal and external datasets, respectively.
The CXR-PH-Net and CXR-CHD-PAH-Net models demonstrated high sensitivity as screening tools for PH and CHD-PAH, potentially facilitating early detection and triage for further evaluation, particularly in resource-limited settings. Further validation in diverse populations is warranted to enhance clinical generalizability.
ClinicalTrials.gov; No.: NCT05566002; URL: www.
gov.
Zotero 插件