Feed

Background: Thymic epithelial tumors (TETs) are the most common primary neoplasms of the anterior mediastinum and present a dual classification challenge, namely anatomical staging according to the Masaoka-Koga system and histological risk stratification according to the World Health Organization (WHO) classification. Both tasks rely on expert pathological assessment and may be affected by interobserver variability. This study applied supervised machine learning (ML) to quantitative immunohistochemical (IHC) H-score profiles to predict Masaoka-Koga stage and WHO risk group in TETs. Methods: Logistic regression (LR) and XGBoost were applied to 19 biomarkers, including cellular localization, across two parallel analyses. Masaoka-Koga stage prediction was performed in 81 patients, including 59 early-stage and 22 advanced-stage cases, using the Synthetic Minority Oversampling Technique (SMOTE) across 100 train/test splits. WHO risk group prediction was performed in 89 patients, including 45 low-risk and 44 high-risk tumors, without oversampling. A cross-endpoint analysis applied the optimal Masaoka-Koga model to the WHO endpoint. Results: LR consistently outperformed XGBoost. The optimal Masaoka-Koga model combined Eph receptor A6 (EphA6) membranous, Yes-associated protein (YAP) nuclear, and histone deacetylase 4 (HDAC4) cytoplasmic H-scores, achieving an area under the curve (AUC) of 0.756. The optimal WHO model combined transcriptional coactivator with PDZ-binding motif (TAZ) cytoplasmic, EphA6 membranous, and YAP nuclear H-scores, achieving an AUC of 0.936. The Masaoka-Koga triad predicted WHO risk group with an AUC of 0.901. No tetrad improved trivariate performance. Conclusions: IHC H-score profiling combined with supervised ML identifies biologically interpretable candidate signatures for TET classification, although prospective external validation is required before clinical application.