An interpretable machine learning model for predicting survival in pancreatic cancer via SHAP: a multicenter study

Existing pancreatic cancer prediction models still have significant limitations until now. This multicenter retrospective study aimed to identify clinical features and develop machine learning models for predicting overall survival (OS) in patients with pancreatic cancer.

Clinicopathological and survival data from patients with pancreatic cancer who underwent radical surgery between 2012 and 2023 were collected at two major pancreatic centers in China. A total of 704 patients from the National Cancer Center of China (NCC) formed the training and internal validation cohort, while 131 patients from Sun Yat-sen Memorial Hospital constituted the external validation cohort. Five predictive machine learning models were developed and validated, and the optimal predictive model was determined by comparing area under the receiver operating characteristic curve (AUC) values. The SHapley Additive exPlanation (SHAP) method was employed to provide interpretability for the machine learning model.

The median OS for 704 postoperative pancreatic cancer patients in NCC was 24 months (21-26 months), with 1-year, 3-year, and 5-year survival rates of 72.8%, 34.0%, and 22.1%, respectively. Perioperative chemotherapy was significantly associated with improved survival (P = 0.0026). Survival data for NCC were generally consistent with Japan and the United States. Among the five predictive models, the Random Survival Forest (RSF) model exhibited superior performance, achieving AUC values of 0.81, 0.76 and 0.78 in the training, internal and external validation sets. The most influential variables contributing to the model predictions were identified using the SHAP method, with those of particular importance including chemotherapy, CA19-9, abdominal pain, the number of lymph node resection and TNM stage.

The 5-year survival rate for postoperative pancreatic cancer patients is 22.1% in NCC, which is comparable with the United States and Japan. Based on multicenter clinical data, we developed and validated an interpretable survival prediction model, which can guide clinical management and personalized treatment for pancreatic cancer patients.