An explainable prediction model for lymphovascular invasion and its prognostic value in resected pancreatic ductal Adenocarcinoma:A two-center study.

Abstract

Background: Lymphovascular invasion (LVI) is an independent risk factor for poor prognosis in pancreatic ductal adenocarcinoma (PDAC). This study aimed to develop and validate an explainable machine learning (ML)-based prediction model for LVI and assessed its prognostic value in patients with PDAC.

Methods: In this two-center retrospective study, a total of 262 patients (141 in the training cohort, 61 in the internal validation cohort, and 60 in the external validation cohort) with PDAC who underwent CECT examination were included. Preoperative indicators, including clinical characteristics, imaging findings and laboratory parameters, were utilized to construct prediction models with 10 ML algorithms. The Shapley Additive explanation method was further applied to explain the feature importance. Lastly, the association of the model-based risk stratification with disease-free survival (DFS) and overall survival (OS) was examined via Cox regression analysis.

Results: The light gradient boosting machine (LightGBM) model demonstrated the best discriminative ability among the 10 ML models in the internal validation cohort. After feature removal according to feature importance ranking, a final explainable LightGBM model was derived with 10 features. The final model could accurately predict LVI in the internal (AUC = 0.814) and external (AUC = 0.795) validation groups. The model-based LVI risk stratification was an independent predictor of both DFS (all P < 0.001) and OS (all P < 0.001), demonstrating good prognostic performance across all subgroups.

Conclusion: The explainable LightGBM model is an effective, non-invasive, and visualizable computer-aided tool for predicting the LVI status in patients with PDAC.