Novel machine-learning model for predicting lymph node metastasis in resectable pancreatic ductal adenocarcinoma

Abstract

Aim: Lymph node metastasis is an adverse prognostic factor in pancreatic ductal adenocarcinoma. However, it remains a challenge to predict lymph node metastasis using preoperative imaging alone. We used machine learning (combining preoperative imaging findings, tumor markers, and clinical information) to create a novel prediction model for lymph node metastasis in resectable pancreatic ductal adenocarcinoma.

Methods: The data of patients with resectable pancreatic ductal adenocarcinoma who underwent surgery between September 1991 and October 2022 were retrospectively examined. Machine-learning software (Statistical Package for the Social Sciences Modeler) was used to create a prediction model, and parameter tuning was performed to improve the model's accuracy. We also analyzed the contribution of each feature to prediction using individual conditional expectation and partial dependence plots.

Results: Of the 331 cases included in the study, 241 comprised the training cohort and 90 comprised the test cohort. After parameter tuning, the areas under the receiver operating characteristic curves for the training and test cohorts were 0.780 and 0.795, respectively. Individual conditional expectation and partial dependence plots showed that larger tumor size and carbohydrate antigen 19-9 and Duke pancreatic monoclonal antigen type 2 levels were associated with positive lymph node metastasis prediction in this model; neoadjuvant treatment was associated with negative lymph node metastasis prediction.

Conclusion: Machine learning may contribute to the creation of an effective predictive model of lymph node metastasis in pancreatic ductal adenocarcinoma. Prediction models using machine learning may contribute to the development of new treatment strategies in resectable pancreatic ductal adenocarcinoma.

Keywords: adenocarcinoma; early diagnosis; lymphatic metastasis; machine learning; pancreatic neoplasms.

FIGURE 1

Flowchart of machine learning used in this study. The training and test cohorts were used to create a prediction model with the XGB algorithm. After maximizing the prediction performance by hyperparameter tuning, the prediction model was interpreted using the ICE plot and PDP. ICE, individual conditional expectation; PDP, partial dependence plot; XGB, Extreme Gradient Boosting.

FIGURE 2

Patient selection process.

FIGURE 3

ROC curves of the prediction model in the training and test cohorts. AUC, area under ROC curves; ROC, receiver operating characteristic.

FIGURE 4

ICE plot and PDP of the prediction model. The average of each ICE plot (blue line in the graph) is the PDP (orange line). The ICE plot and PDP were used to evaluate the relationships between the model and each feature. The vertical axis shows the prediction of the model, with positive and negative values on the vertical axis corresponding to the positive and negative prediction of LNM, respectively. CA19‐9, carbohydrate antigen 19–9; CE, individual conditional expectation; CEA, carcinoembryonic antigen; DUPAN‐2, Duke pancreatic monoclonal antigen type 2; PDP, partial dependence plot.