Work

AB024. Mixed histology is a prognostic factor for recurrence in completely resected thymic epithelial tumours: a multicentre retrospective study

Year: 2026
Type: conference-abstract
Abstract: BACKGROUND: The recurrence of thymic epithelial tumours (TETs) following complete resection remains unpredictable. While the Masaoka-Koga staging system and WHO subtype classification are useful for clinical decision-making, no externally validated prognostic model exists to inform risk-adapted follow-up. Tumours with composite histological patterns, such as B2+B3 or B3+TC, are usually classified by their dominant subtype, potentially underestimating their aggressiveness. This study aimed to determine whether a machine learning (ML) model using these routine pathological variables could predict recurrence in completely resected TETs and whether including mixed histology would improve risk stratification. METHODS: Clinical data were collected from four centres (EMC, UMC+, TCGA and HCL). Patients with R1/R2 resection or stage IV disease were excluded. An internal development cohort (EMC (n=89), UMC+ (n=102), TCGA (n=116) and a geographically independent external validation cohort (HCL, n=95) were defined. A binary variable was introduced to identify tumours presenting two or more WHO subtypes. Six ML algorithms were evaluated and explainability was ensured through SHAP-based feature attribution. RESULTS: Patients with mixed histology had a recurrence rate of 18.5% (5/27) compared to 5.1% (14/277) in single-subtype cases (χ²=5.49, P=0.02). Among the six algorithms, Random Forest achieved the most consistent performance, with a balanced accuracy (BAcc) of 0.73 and Area Under the Curve (AUC) of 0.73 on the internal development cohort and a BAcc of 0.73, AUC of 0.72 and sensitivity of 0.80 on the external validation cohort. Ablation analysis confirmed the independent contribution of the mixed subtype indicator, as its removal reduced specificity by 0.03 and 0.08 in the internal and external cohorts respectively. SHAP analysis revealed that Masaoka-Koga stage IIb and subtype B2 had the highest feature importance, partly reflecting their prevalence in the dataset. Importantly, the overall pattern was consistent with established clinical knowledge, that aggressive subtypes and higher stages driving recurrence and indolent subtypes acting as protective, suggesting the model learned authentic clinicopathological relationships. CONCLUSIONS: We developed an interpretable, externally validated model built on routine pathology data achieved clinically meaningful recurrence prediction in completely resected TETs. Our results support formal recognition of mixed histological subtypes as an independent risk factor and their inclusion in pathology may improve postoperative risk stratification.
Source: Mediastinum
Language: English
Field: Medicine
Open Access status: diamond
OpenAlex ID: W7225083173
Cites:
Cited by:
Related to: 10