Compare the prognosis of pancreatic cancer patients with different treatment modalities and use machine learning methods to build predictive models

Lingling Fan¹Shujing Kong²Yizhou Deng³Yunfan Wang¹Xiancheng Yan¹Chao Jiang^1*Li Tao^2*Weimin Wang^1*

• ¹Department of Oncology, Yixing Hospital Affiliated to Medical College of Yangzhou University, Yixing 214200, Jiangsu, China, Yixing, China
• ²The Key Laboratory of Syndrome Differentiation and Treatment of Gastric Cancer, College of Medicine, Yangzhou University, Yangzhou 225009, Jiangsu, China, Yangzhou, China
• ³Department of Clinical Medicine, College of Medicine, Yangzhou University, Yangzhou 225009, Jiangsu, China, Yangzhou, China

Background:Pancreatic cancer (PC) is highly refractory to most treatments. Multimodal treatment, combining several types of therapies, is likely to benefit PC patients. However,it remains unclear which multimodal treatment is most effective and how to predict outcomes from different combinations. This study compared overall survival among PC patients receiving chemotherapy alone (C), immunotherapy combined with chemotherapy (CI), radiotherapy combined with chemotherapy (CR), and triple-combination therapy (CRI). A machine learning-based predictive model between monomodal and multimodal therapy was established using three years of clinical follow-up data.Method:We retrospectively analyzed 125 cases of PC patients treated at Yixing People’s Hospital from January 2014 to June 2024 (C, n=50; IC, n=38; RC, n=18; RIC, n=19). The group IC, RC and RIC were merged and defined as multiple modalities (MM) group (n=75), while the group C was defined as single modality (SM) treatment group (n=50). Subsequent sensitivity analysis and inverse probability weighting methods were used to achieve baseline balance between the two groups. Kaplan-Meier plots estimated the overall survival rate of each group and the survival rate of the SM group and the MM group.Cox proportional hazard models identified key prognostic factors, including cytokines and inflammation mediators. Four machine learning models, including logistic regression (LR), support vector machine (SVM), random forest (RF), and Extreme Gradient Boosting (XGBoost) were used to build predictive models. SHapley Additive exPlanations (SHAP) identified significant contributors to treatment outcomes.Results: Multimodal treatments significantly improved PC prognosis (P=0.0025).Univariate and multivariate Cox regression analysis showed that interleukin-2 (IL-2) was a protective factor, while neutrophil-to-lymphocyte ratio (NLR) was a risk factor. This study evaluated and compared the predictive performance of four machine learning models using the classifiers such as area under curve (AUC), accuracy and F1 score, etc. In the binary classification task, RF and XGBoost models both achieved good performance compared with the other two machine learning methods. In addition, SHAP analysis also proved that IL-6 contributed the most to the machine learning models.Conclusion:PC patients may benefit from more intensive multimodal therapies, which provides novel insights into predicting PC survival prognosis and highlights the potential of machine learning in biomarker identification and disease prognosis.