@article{10.37349/emed.2026.1001426,
abstract = {Aim: To investigate interobserver variability in programmed cell death ligand 1 (PD-L1) combined positive score (CPS) assessment in head and neck squamous cell carcinoma (HNSCC) and to develop an artificial intelligence (AI)-based model for predicting PD-L1 expression and patient prognosis from hematoxylin and eosin (H&E)-stained slides. Methods: Fifty HNSCC specimens were independently evaluated for PD-L1 by pathologists with different experience levels. Agreement was assessed using Fleiss’ and Cohen’s κ. Whole-slide images were processed into tiles for deep learning using DenseNet121. Tile-level features were integrated via two machine learning pipelines to construct whole-slide prediction models. Multiple algorithms were tested, with performance evaluated in validation and testing cohorts. Prognostic value was analyzed using AI-derived risk stratification. Results: Interobserver agreement was low (Fleiss’ κ = 0.34), indicating substantial variability in CPS assessment. DenseNet121 achieved moderate predictive performance (AUC 0.641 in validation, 0.616 in testing). AI models significantly improved prediction accuracy, with logistic regression demonstrating the best performance (AUC 0.900 in validation, 0.851 in testing). AI-derived prediction scores effectively stratified overall survival, with multiple models showing significant prognostic discrimination (P < 0.05). Conclusions: AI models integrating deep learning and machine learning can accurately predict PD-L1 expression and stratify prognosis in HNSCC, outperforming tile-level deep learning alone.},
author = {Cui, Yingying and Ding, Chuanyang and Li, Long and Cai, Xinjia},
doi = {10.37349/emed.2026.1001426},
journal = {Exploration of Medicine},
elocation-id = {1001426},
title = {A computational pathology-based AI framework for predicting PD-L1 expression and prognosis in HNSCC},
url = {https://www.explorationpub.com/Journals/em/Article/1001426},
volume = {07},
year = {2026}
}