Abstract T047

Artificial intelligence predicts EBV status and reveals prognostic morphology phenotypes in advanced-stage classic Hodgkin lymphoma

Background: Epstein–Barr virus (EBV) status and morphologic heterogeneity characterize the histopathologic phenotype of classic Hodgkin lymphoma (cHL). We asked whether artificial intelligence (AI) models trained on self-supervised histology foundation-models can (i) predict EBV status without testing for the virus and (ii) define clinically meaningful clusters from routine stains.

Methods: We analyzed 935 advanced stage cHL cases from the German Hodgkin Study Group HD21 trial. Embeddings were generated from H&E, CD30 and CD20 staining Whole-slide imaging (WSI) using the general pathology foundation model PathOrchestra. Unsupervised clustering with k-means was applied on mean patch embedding per case and tested against EBV status and interim PET positivity (PET-2) by Fishers exact test, and 3-year progression-free survival (PFS) differences using log-rank test. In addition, a supervised EBV prediction was trained using attention-based multiple instance learning per stain and extractor. All metrics are 5-fold cross-validated.

Results: Supervised prediction for detection of EBV status worked best using H&E (5-fold out-of-fold [OOF] held-out cross-validation prediction AUC for all staining combined: 0.74; CD20: 0.70; CD30: 0.71; H&E: 0.77). Unsupervised clustering revealed distinct phenotypes (k = 2 for H&E/CD30, k = 3 for CD20). The CD30 phenotype clusters were associated with EBV (19% Cluster 0 (C0) vs. 11% C1; p=0.01). CD20 phenotypes associated with interim PET (C0: 30%, C1: 25%, C2: 35%; p = 0.01). After the combination of similarly prognostic clusters a prognostic difference could be demonstrated (3-yr PFS C0&C2: 95.0% [95% CI: 93.0–96.4] vs. C1: 91.3% [95% CI: 87.2–94.1]; p=0.02).

Conclusion: AI analysis of conventional histology slides reveal clinical phenotypes and show feasibility of detecting molecular features of Hodgkin-Reed-Sternberg-cells like EBV without directly testing for this feature. These findings will be validated in an additional cohort that will be presented at the meeting.