AI-enhanced diagnostic model for pulmonary nodule classification.

Background: The identification of benign and malignant pulmonary nodules (BPN and MPN) can significantly reduce mortality. However, a reliable and validated diagnostic model for clinical decision-making is still lacking. Methods: Enzyme-linked immunosorbent assay and electro chemiluminescent immunoassay were utilized to determine the serum concentrations of 7AABs (p53, GAGE7, PGP9.5, CAGE, MAGEA1, SOX2, GBU4-5), and 4TTMs (CYFR21, CEA, NSE and SCC) in 260 participants (72 BPNs and 188 early-stage MPNs), respectively. The malignancy probability was calculated using Artificial intelligence pulmonary nodule auxiliary diagnosis system, or Mayo model. Along with age, sex, smoking history and nodule size, 18 variables were enrolled for model development. Baseline comparison, univariate ROC analysis, variable correlation analysis, lasso regression, univariate and stepwise logistic regression, and decision curve analysis (DCA) was used to reduce and screen variables. A nomogram and DCA were built for model construction and clinical use. Training (60%) and validation (40%) cohorts were used to for model validation. Results: Age, CYFRA21_1, AI, PGP9.5, GAGE7, and GBU4_5 was screened out from 18 variables and utilized to establish the regression model for identifying BPN and early-stage MPN, as well as nomogram and DCA for clinical practical use. The AUC of the nomogram in the training and validation cohorts were 0.884 and 0.820, respectively. Moreover, the calibration curve showed high coherence between the predicted and actual probability. Conclusion: This diagnostic model and DCA could provide evidence for upgrading or maintaining the current clinical decision based on malignancy probability stratification. It enables low and moderate risk or ambiguous patients to benefit from more precise clinical decision stratification, more timely detection of malignant nodules, and early treatment.