Objective To investigate the application value of a combined model based on clinicopathological features and contrast‑enhanced computed tomography (CT) deep learning model for predicting lymph node metastasis in early gastric cancer.

Methods The retrospective cohort study was conducted. The clinicopathological data of 1 156 patients with early gastric cancer who were admitted to Renji Hospital, Shanghai Jiao Tong University School of Medicine, Punan Hospital of Pudong New Area, Shanghai, and Huadong Hospital Affiliated to Fudan University from January 2010 to January 2020 were collected. There were 741 males and 415 females, aged (62±12) years. The 1 050 patients from Renji Hospital, Shanghai Jiao Tong University School of Medicine and Punan Hospital of Pudong New Area, Shanghai were randomly divided into a training set (735 cases) and an internal validation set (315 cases) at a ratio of 7∶3 using the random number table method. The 106 patients from Huadong Hospital Affiliated to Fudan University were assigned as the external valida-tion set. The training set was used to construct the prediction model, and the validation set was used to validate the predictive model. Observation indicators: (1) analysis of influencing factors for lymph node metastasis in early gastric cancer; (2) construction and efficacy evaluation of the predictive model for lymph node metastasis in early gastric cancer. Comparison of measurement data with normal distribution between groups was conducted using the independent sample t test. Comparison of measurement data with skewed distribution between groups was conducted using the Mann‑Whitney U test. Comparison of count data between groups was conducted using the chi‑square test or Fisher exact probability. Key factors were screened out based on the Akaike information criterion and incorporated into Logistic regression analysis. Receiver operating characteristic (ROC) curves were plotted, and the area under the curve (AUC) was used to evaluate the predictive efficacy of the model. DeLong test was used to compare the AUC of different models. The Kolmogorov‑Smirnov (KS) curve was used to assess the model′s ability to distinguish between positive and negative samples.

Results (1) Analysis of influencing factors for lymph node metastasis in early gastric cancer: among 735 patients in the training set, 627 cases had no lymph node metastasis (T1N0 stage) and 108 cases had lymph node metastasis (T1N+ stage). Results of Logistic regression analysis showed that sex, neutrophil count, carcinoembryonic antigen (CEA), tumor differentiation degree, carbohydrate antigen 19‑9 (CA19‑9), and tumor diameter were independent influencing factors for lymph node metastasis in early gastric cancer odds ratio (OR)=1.65, 1.10, 1.09, 2.17, 0.97, 0.76, 95% confidence interval (CI) as 1.02-2.67, 1.01-1.22, 1.03-1.18, 1.24-3.78, 0.95-0.98, 0.63-0.91, P<0.05. (2) Construction and efficacy evaluation of the predictive model for lymph node metastasis in early gastric cancer: a clinical model was constructed based on results of Logistic regression analysis, a deep learning model was constructed using VGG16 as the backbone network, and a combined model was further constructed by integrating the clinical model with deep learning model. In the training set, ROC analysis showed that the AUC of the combined model for predicting lymph node metastasis in early gastric cancer was 0.998 (95%CI as 0.995-1.000), the AUC of the clinical model was 0.775 (95%CI as 0.727-0.823), and the AUC of the deep learning model 0.997 (95%CI as 0.994-1.000). In the internal validation set, the AUC of the combined model was 0.994 (95%CI as 0.985-1.000), the AUC of the clinical model was 0.760 (95%CI as 0.683-0.838), and the AUC of the deep learning model was 0.997 (95%CI as 0.990-1.000). Results of DeLong test showed that there was no significant difference between the combined model and the deep learning model (P>0.05), while there were significant differences between the combined model and the clinical model, between the deep learning model and the clinical model (P<0.05). KS curve analysis in the internal validation set showed that the optimal cutoff value of the combined model was 8.211, with a corresponding KS value of 0.992. A confusion matrix was constructed based on the optimal cutoff value. In the internal validation set, the combined model correctly identified 273 cases and misclassified 1 cases of T1N0 stage patients, correctly identified 40 cases and missed 1 cases of T1N+stage patients. The sensitivity, specificity, accuracy, positive predictive value, and negative predictive value of the combined model were 97.6%, 99.6%, 99.4%, 97.6% and 99.6%, respectively. In the external validation set, the AUC of the combined model was 0.744 (95%CI as 0.612-0.875), the AUC of the clinical model was 0.684 (95%CI as 0.533-0.836), and the AUC of the deep learning model was 0.661 (95%CI as 0.530-0.792). In the external validation set, the combined model correctly identified 76 cases and misclassified 11 cases of T1N0 stage patients, correctly identified 9 cases and missed 10 cases of T1N+ stage patients. The sensitivity, specificity, accuracy, positive predictive value, and negative predictive value of the combined model were 47.4%, 87.4%, 80.2%, 45.0% and 88.4%, respectively. Comparison of the combined model scores between T1N0 stage and T1N+ stage patients showed a significant difference (t=3.88, P<0.05).

Conclusion The combined model based on clinicopathological characteristics and contrast‑enhanced CT deep learning model has good predictive value for lymph node metastasis in early gastric cancer.