Establishment and validation of a mathematical diagnosis model to distinguish benign pulmonary nodules from early non-small cell lung cancer in Chinese people.

Background: In this study, we aimed to establish and validate a mathematical diagnosis model to distinguish benign pulmonary nodules (BPNs) from early non-small cell lung cancer (eNSCLC) based on clinical characteristics, radiomics features, and hematological biomarkers.

Methods: Medical records from 81 patients (27 BPNs, 54 eNSCLC) were used to establish a novel mathematical diagnosis model and an additional 61 patients (21 BPNs, 40 eNSCLC) were used to validate this new model. To establish a clinical diagnosis model, a least absolute shrinkage and selection operator (LASSO) regression was applied to select predictors for eNSCLC, then multivariate logistic regression analysis was performed to determine independent predictors of the probability of eNSCLC, and to establish a clinical diagnosis model. The diagnostic accuracy and discriminative ability of our model were compared with the PKUPH and Mayo models using the following 4 indices: area under the receiver-operating characteristics curve (ROC), net reclassification improvement index (NRI), integrated discrimination improvement index (IDI), and decision curve analysis (DCA).

Results: Multivariate logistic regression analysis identified age, border, and albumin (ALB) as independent diagnostic markers of eNSCLC. In the training cohort, the AUC of our model was 0.740, which was larger than the AUCs for the PKUPH model (0.717, P=0.755) and the Mayo model (0.652, P=0.275). Compared with the PKUPH and Mayo models, the NRI of our model increased by 3.7% (P=0.731) and 27.78% (P=0.008), respectively, while the IDI changed -4.77% (P=0.437) and 11.67% (P=0.015), respectively. Moreover, the DCA demonstrated that our model had a higher overall net benefit compared to previously published models. Importantly, similar findings were confirmed in the validation cohort.

Conclusions: Age, border, and serum ALB levels were independent diagnostic markers of eNSCLC. Thus, our model could more accurately distinguish BPNs from eNSCLC and outperformed previously published models.