Abstract:

Objective ·To verify the performance and explore the clinical application value of a multi-modal pulmonary nodule diagnosis model combined with metabolic fingerprints, protein biomarker CEA and Image-AI via random forest (MPI-RF). Methods ·This study enrolled 289 patients with pulmonary nodules who were admitted to the Shanghai Chest Hospital, Shanghai Jiao Tong University School of Medicine and were detected by low-dose helical computed tomography (LDCT). The patients were divided into malignant nodule group ( n=197) and benign nodule group ( n=92) based on postoperative pathological results, and the basic information of the two groups was collected and compared. Electrochemiluminescence was used to detect the preoperative serum CEA levels of the patients in the two groups, matrix-assisted laser desorption/ionization mass spectrometry (MALDI-MS) was used to detect the serum metabolic fingerprints, and the CT image artificial intelligence model Image-AI was used to calculate the image scores. CEA data, serum metabolic fingerprints data and image scores were integrated and input into MPI-RF to calculate the malignant probability score of each patient. The receiver operator characteristic curve (ROC curve) and area under the curve (AUC) were used to evaluate the performance of different models, and the DeLong test was used for comparative analysis, including the diagnostic performance of MPI-RF in different types (solid nodule, pure ground-glass nodule and part-solid nodule) and sizes (diameter<8 mm and diameter≥8 mm) of pulmonary nodules, the diagnostic performance comparison of MPI-RF with Mayo Clinic model, veterans administration (VA) model and Brock model, and the diagnostic performance comparison of MPI-RF with lung imaging reporting and data system (Lung-RADS) in benign and malignant nodules. Results ·MPI-RF had good diagnostic performance in the differentiation of benign and malignant pulmonary nodules (AUC=0.887, 95% CI 0.848?0.925, sensitivity 81.22%, specificity 83.70%). Among them, the AUC of MPI-RF for solid nodules was 0.877 (95% CI 0.820?0.934), for part-solid nodules was 0.858 (95% CI 0.771?0.946), and for pure ground-glass nodules was 0.978 (95% CI 0.923?1.000). The AUC of MPI-RF was 0.840 (95% CI 0.716?0.963) for nodules within 8 mm diameter and 0.891 (95% CI 0.849?0.933) for nodules larger than 8 mm diameter. Compared with the existing models, the diagnostic performance of MPI-RF was better than that of Mayo Clinic model, VA model and Brock model (all P=0.000). Compared with Lung-RADS, MPI-RF had better diagnostic performance in the total samples and different types of nodules (all P=0.000). Conclusion ·MPI-RF is a model for the differential diagnosis of benign and malignant pulmonary nodules with excellent performance, and has potential clinical application value.

Key words: machine learning, lung adenocarcinoma, metabolomics, multi-modal model, pulmonary nodule