AUTHOR=Wei Yuanhui , Zhao Wei , Wu Zhen , Guo Nannan , Wang Miaoyu , Yu Hang , Wang Zirui , Shi Wenjia , Ma Xiuqing , Li Chunsun , Ren Jiabo , Yin Yue , Liu Shangshu , Yang Zhen , Chen Liang-an TITLE=Integrating multimodal features to predict the malignancy of pulmonary ground−glass nodules: a multicenter prospective model development and validation study JOURNAL=Frontiers in Oncology VOLUME=Volume 15 - 2025 YEAR=2025 URL=https://www.frontiersin.org/journals/oncology/articles/10.3389/fonc.2025.1547816 DOI=10.3389/fonc.2025.1547816 ISSN=2234-943X ABSTRACT=BackgroundThere is a clinical need for accurate noninvasive evaluation of the malignancy of pulmonary ground−glass nodules (GGNs) to reduce risks of overdiagnosis and overtreatment. This study aimed to develop and validate a clinic-biomarker-combined deep radiomic model for the prediction of GGN malignancy.Materials and methodsThis study recruited patients with GGNs from seven medical centers across five cities in China. The participants included in this study were divided into the training-validation and the test groups on the basis of the centers from which they were recruited. The malignancy of GGNs was determined based on pathological results. Clinical, radiological, and biomarker features with significant differences were used to establish predictive models. Six types of models based on different features were developed on the training-validation group: clinical-radiological (CR), biomarker-combined CR (B-CR), deep radiomic (DR), clinic-combined DR (C-DR), biomarker-combined DR (B-DR), and clinic-biomarker-combined DR (CB-DR) models. The models were then evaluated on the test group for discrimination, calibration, and clinical utility.ResultsA total of 501 participants with 571 GGNs were included in the study. Four hundred and seven participants with 454 GGNs were assigned to the training-validation group, whereas 94 participants with 117 GGNs were assigned to the test group. Significant differences were observed in sex, smoking history, triosephosphate isomerase-1 and microRNA-206 between patients with and without malignant GGNs. And size, location, and lobulation were significantly different between benign and malignant GGNs. Among all the models, the CB-DR model achieved the highest performance in classifying GGNs, with an AUC of 0.90 (95% CI: 0.81-0.97). At the optimal cutoff, the corresponding accuracy, sensitivity, and specificity were 0.89 (95% CI: 0.83–0.94), 0.90 (95% CI: 0.84–0.96), and 0.82 (95% CI: 0.62–1.00), respectively. Furthermore, malignancy evaluation based on the CB-DR model would have reduced overtreatment for 82.4% (14/17) of benign GGNs and enabled timely interventions for 90.0% (90/100) of malignant GGNs.ConclusionThe CB-DR model developed in this study exhibited satisfactory performance in predicting the malignancy of GGNs and holds potential as a valuable tool for aiding clinical decision-making in GGN management.