We retrospectively collected and analyzed data from patients who underwent lung cancer surgery at the Affiliated Changshu Hospital of Soochow University and Nantong Eighth People’s Hospital from January 2020 to December 2023. Patients collected from January 2020 to December 2023 at the Affiliated Changshu Hospital of Soochow University were used as the training set, while patients collected from October 2022 to December 2023 at Nantong Eighth People’s Hospital were used as the testing set.

The diagnostic criteria for lung cancer were referenced from the 2021 Fifth Edition of the WHO Classification of Thoracic Tumors (8). Diagnosis of lung cancer required meeting the following criteria: (1) Confirmation of lung nodules by chest CT without any clinical or drug intervention; (2) Definitive pathological results confirming benign or malignant nodules after chest CT; (3) Age ≥ 18 years; (4) Preoperative testing for 7 lung cancer autoantibodies and tumor markers; (5) Absence of significant dysfunction in other major organs; (6) Absence of other primary malignant tumors; and (7) Lung nodule diameter ≤ 3 cm. Exclusion criteria included: absence of pathological examination despite confirmed lung nodules on chest CT; failure to undergo testing for the 7 lung cancer autoantibodies and tumor markers; clinical or drug intervention prior to blood sampling; presence of rheumatic immunological diseases; lung metastasis from other tumors; lung nodule diameter > 3 cm. This study was approved by the hospital ethics committee.

Demographic features, clinical information, and comorbidities were extracted from electronic medical records. Chest plain scans were performed using a 64-slice spiral CT scanner to obtain conventional CT imaging features, including air bronchogram sign, spiculated sign, lobulation sign, vascular penetration, pleural retraction, bronchial inflation sign, nodule diameter, and solid proportion. And the patient’s CT data were imported into the DeepRay medical image AI recognition system, which extracted quantitative features from medical images in high throughput and combined with convolutional neural networks to train deep learning on the data of the nodule’s size, density, and the proportion of solidity to get the improved CT indexes: the pulmonary nodule’s malignancy probability value and average CT value. Serum biomarkers primarily included 7 tumor-associated autoantibodies (TAABs) and commonly used tumor markers recommended by the American Clinical Biochemistry Committee and the European Tumor Marker Expert Group. TAABs detection involved extracting fasting peripheral venous blood (9–12) from patients preoperatively or before surgery. After centrifugation to separate serum, the levels of 7 lung cancer autoantibodies were measured using enzyme-linked immunosorbent assay (ELISA) (13), including tumor suppressor gene P53 (normal reference range: P53 < 13.09 U/mL), protein gene product PGP 9.5 (normal reference range: PGP9.5 < 11.1 U/mL), SRY-box containing gene 2 (normal reference range: SOX2 < 10.26 U/mL), G antigen 7 (GAGE7) (normal reference range: GAGE7 < 14.36 U/mL), RNA helicase autoantibody 4–5 (GBU4-5) (normal reference range: GBU4-5 < 6.99 U/mL), melanoma antigen A1 (MAGEA1) (normal reference range: MAGEA1 < 11.92 U/mL), and tumor-associated gene CAGE (normal reference range: CAGE <7.23 U/mL). TAABs detection results were considered positive if any of the indicators exceeded the normal reference range. Tumor markers were collected from blood tests and included primary lung cancer markers such as vascular endothelial growth factor (VEGF), carcinoembryonic antigen (CEA), neuron-specific enolase (NSE), cytokeratin fragment 19 (CYFRA21-1), pro-gastrin-releasing peptide (ProGRP), and squamous cell carcinoma antigen (SCC) (14).

Through the AI platform¹, the H2O package is installed in the R language to implement AutoML analysis. Autonomy and automation are achieved through three aspects: feature selection, model construction, and hyperparameter optimization. The integrated algorithms include Generalized Linear Models (GLM), Random Forests (RF), Gradient Boosting Machines (GBM), Deep Neural Networks (DL), and Naive Bayes (NB), among others. The training set is split into development and validation sets in a 6:4 ratio, and blind verification is conducted with the testing set to evaluate the average accuracy and stability of the models. A confusion matrix consisting of true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN) is established (15). Performance metrics including sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), positive likelihood ratio (LR+), negative likelihood ratio (LR-), accuracy, area under the receiver operating characteristic curve (AUC), and the F1-Measure are calculated. Formulas for calculation are as follows: Accuracy = (TP + TN)/(TP + FP + FN + TN); PPV = TP/(TP + NP); NPV = TN/(TN + FN); LR + =Sensitivity/(1−Specificity); LR− = (1−Sensitivity)/Specificity; F1-Measure = (2*precisionrecall)/(precision+recall). Through SHAP analysis (Shapley Additive Explanations), an additive explanatory model is constructed to determine significant factors influencing model predictions and their contributions to model performance.

For continuous data, the Shapiro–Wilk test and homogeneity of variance test (Homogeneity of variance test) were first performed. For normally distributed and homoscedastic continuous data, independent samples t-tests were employed, and results were presented as mean ± standard deviation. For non-normally distributed and heteroscedastic continuous data, the Wilcoxon rank-sum test was used, and results were presented as median (M25, M75). Categorical data were expressed as frequencies and percentages, and inter-group differences were assessed using the chi-square test or Fisher’s exact test. To prevent multicollinearity among variables, feature selection was conducted using the Least Absolute Shrinkage and Selection Operator (LASSO) regression model. Based on the selected variables, a binary logistic regression model was fitted. The predictive performance of the obtained model was evaluated using the area under the receiver operating characteristic curve (AUC), calibration curve, and decision curve analysis (DCA), and a Nomogram was constructed. The statistical significance level was set at p < 0.05. All statistical analyses were performed using R 4.3.3 software.

A total of 803 lung cancer patients were included in this study, with 376 cases (47.0%) exhibiting infiltrative lesions. The study protocol is detailed in Figure 1. Among them, 560 patients from the Affiliated Changshu Hospital of Soochow University (Hospital 1) were included in the training set. Nantong Eighth People’s Hospital (Hospital 2) contributed 243 patients as the testing set. In the training set, 64.3% (360/560) were male and 35.7% (200/560) were female, with a median age of 55 years. In the testing set, females were more common in the infiltrative group, and the age range of 40–60 years was the peak incidence, consistent with previous reports (16). There were no statistically significant differences between the two groups in terms of age, CY211, NSE, and Leafing (p > 0.05). Details are shown in Table 1.

Figure 7 shows that nodule diameter size is the most important feature, followed by average CT value, solid proportion, NSE, VEGF, CYFRA21-1, SCC, malignancy probability, CEA, and proGRP. Additionally, nodule diameter size, average CT value, malignancy probability, solid proportion, and VEGF were identified as important feature variables shared by both the GBM and logistic regression models.

Figure 8, the SHAP summary plot, displays the impact of all features on the predictive performance of the GBM model in the testing set. The x-axis represents the SHAP values, indicating the contribution of features to the overall prediction. A SHAP value greater than 0 indicates a positive contribution, meaning that as the variable’s value approaches 1, the likelihood of infiltration in patients increases. For example, on the SHAP plot corresponding to nodule diameter, red points are mainly located to the right of the zero axis, while blue points are more on the left, suggesting that as the nodule diameter increases, the likelihood of infiltrative lesions in lung nodules also increases.

As shown in Figure 9, partial dependence plots illustrate the impact of individual features on the final discrimination of the GBM model and their distribution in the dataset. Nodule diameter size, malignancy probability, and VEGF are positively correlated with the likelihood of infiltrative lesions. Nodule diameter is mainly distributed below 15 mm, but for lung cancer patients falling between 15 and 18 mm, there is a higher likelihood of infiltrative lesions, necessitating regular follow-up. As the average CT value gradually increases, it tends to indicate non-invasive lung cancer, particularly in patients with values above −200, essentially ruling out the possibility of infiltrative lung cancer.

The SHAP explanation illustrates the feature contributions for specific instances. As depicted in Figure 10, for instance 72, with a nodule diameter of 22 mm, average CT value of -525HU, and malignancy probability of 86%, these factors significantly contribute to the model’s final determination of infiltrative lung cancer. Conversely, in instance 98, although the nodule diameter is below 15 mm, predictions of infiltrative lung cancer are made based on factors such as average CT value, NSE value, and malignancy probability.