揭示病理组学的潜力：胰腺癌的预后预测与机制研究

A machine learning-based pathomics model was investigated for its value and biological significance in predicting overall survival (OS) after surgery in pancreatic cancer patients. Data from 173 patients with pancreatic ductal adenocarcinoma (PDAC) who underwent surgery and continued follow-up in two centers were retrospectively analyzed. Pathomics parameters of both the tumor and peritumor were measured in all patients, and the optimal pathomics score (Pathscore) was calculated using five machine learning methods. The best Pathscore was then combined with multiple clinical parameters to analyze its incremental value and to construct a comprehensive nomogram. TCGA data, multiplex immunofluorescence, spatial analysis, and single-cell sequencing were used to explore the biological mechanisms of pathomics. In predicting OS, pathomics parameters from the tumor and peritumoral regions provided complementary prognostic information.

The LASSO-based combined model achieved the best predictive accuracy. Multivariate Cox regression analysis identified T-stage, N-stage, CA19-9, and Pathscore as independent predictors of OS in patients with PDAC. The integrated nomogram demonstrated superior and more stable predictive performance. Analysis of the TCGA dataset suggested that the pathomics model was associated with the immune status of pancreatic cancer, a finding supported by trends in the validation cohort.

Spatial analysis and single-cell analysis further revealed a strong association between the Pathscore and immune cell infiltration, in particular CD8+ T cells. Machine learning-based pathomics models can help to predict the immune status and OS of patients with PDAC. The integration of pathomics with clinical parameters provides a robust basis for immune evaluation, prognostic prediction, and therapeutic decision-making in PDAC. © 2026 The Pathological Society of Great Britain and Ireland.