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

Area of Science:

• Oncology
• Machine Learning
• Biostatistics

Background:

• Pancreatic cancer survival prediction models have limitations.
• Accurate prediction is crucial for patient management and treatment strategies.

Purpose of the Study:

• To identify clinical features for predicting overall survival (OS) in pancreatic cancer patients.
• To develop and validate machine learning models for pancreatic cancer OS prediction.

Main Methods:

• Retrospective analysis of 704 patients (training/internal validation) and 131 patients (external validation).
• Development and comparison of five machine learning models, including Random Survival Forest (RSF).
• Utilized SHapley Additive exPlanation (SHAP) for model interpretability.

Main Results:

• The RSF model demonstrated superior performance with AUCs of 0.81 (training), 0.76 (internal), and 0.78 (external validation).
• Key predictors for OS included chemotherapy, CA19-9, abdominal pain, lymph node resection count, and TNM stage.
• The 5-year survival rate was 22.1%, comparable to international data.

Conclusions:

• An interpretable RSF model was developed and validated using multicenter data.
• This model can aid in clinical decision-making and personalized treatment for pancreatic cancer.
• Identified key prognostic factors offer insights into pancreatic cancer progression.