[Establishment of artificial neural network model for predicting lymph node metastasis in patients with stage Ⅱ-Ⅲ...

Objective: To establish a neural network model for predicting lymph node metastasis in patients with stage II-III gastric cancer. Methods: Case inclusion criteria: (1) gastric adenocarcinoma diagnosed by pathology as stage II-III (the 8th edition of AJCC staging); (2) no distant metastasis of liver, lung and abdominal cavity in preoperative chest film, abdominal ultrasound and upper abdominal CT; (3) undergoing R0 resection. Case exclusion criteria: (1) receiving preoperative neoadjuvant chemotherapy or radiotherapy; (2) incomplete clinical data; (3) gastric stump cancer.Clinicopathological data of 1231 patients with stage II-III gastric cancer who underwent radical surgery at the Fujian Medical University Union Hospital from January 2010 to August 2014 were retrospectively analyzed. A total of 1035 patients with lymph node metastasis were confirmed after operation, and 196 patients had no lymph node metastasis. According to the postoperative pathologic staging. 416 patients (33.8%) were stage Ⅱ and 815 patients (66.2%) were stage III. Patients were randomly divided into training group (861/1231, 69.9%) and validation group (370/1231, 30.1%) to establish an artificial neural network model (N+-ANN) for the prediction of lymph node metastasis. Firstly, the Logistic univariate analysis method was used to retrospectively analyze the case samples of the training group, screen the variables affecting lymph node metastasis, determine the variable items of the input point of the artificial neural network, and then the multi-layer perceptron (MLP) to train N+-ANN. The input layer of N+-ANN was composed of the variables screened by Logistic univariate analysis. Artificial intelligence analyzed the status of lymph node metastasis according to the input data and compared it with the real value. The accuracy of the model was evaluated by drawing the receiver operating characteristic (ROC) curve and obtaining the area under the curve (AUC). The ability of N+-ANN was evaluated by sensitivity, specificity, positive predictive values, negative predictive values, and AUC values. Results: There were no significant differences in baseline data between the training group and validation group (all P>0.05). Univariate analysis of the training group showed that preoperative platelet to lymphocyte ratio (PLR), preoperative systemic immune inflammation index (SII), tumor size, clinical N (cN) stage were closely related to postoperative lymph node metastasis. The N+-ANN was constructed based on the above variables as the input layer variables. In the training group, the accuracy of N+-ANN for predicting postoperative lymph node metastasis was 88.4% (761/861), the sensitivity was 98.9% (717/725), the specificity was 32.4% (44/136), the positive predictive value was 88.6% (717/809), the negative predictive value was 84.6% (44/52), and the AUC value was 0.748 (95%CI: 0.717-0.776). In the validation group, N+-ANN had a prediction accuracy of 88.4% (327/370) with a sensitivity of 99.7% (309/310), specificity of 30.0% (18/60), positive predictive value of 88.0% (309/351), negative predictive value of 94.7% (18/19), and an AUC of 0.717 (95%CI:0.668-0.763). According to the individualized lymph node metastasis probability output by N+-ANN, the cut-off values of 0-50%, >50%-75%, >75%-90% and >90%-100% were applied and patients were divided into N0 group, N1 group, N2 group and N3 group. The overall prediction accuracy of N+-ANN for pN staging in the training group and the validation group was 53.7% and 54.1% respectively, while the overall prediction accuracy of cN staging for pN staging in the training group and the validation group was 30.1% and 33.2% respectively, indicating that N+-ANN had a better prediction than cN stage. Conclusions: The N+-ANN constructed in this study can accurately predict postoperative lymph node metastasis in patients with stage Ⅱ-Ⅲ gastric cancer. The N+-ANN based on individualized lymph node metastasis probability has better accurate prediction for pN staging as compared to cN staging.