867P - Predicting HPV-association using regular H&E slides can identify subgroups of patients with favorable prognosis at a highly detailed level

Background

Oropharyngeal squamous cell cancer (OPSCC) related to Human Papilloma Virus (HPV) is a subgroup of head and neck cancer that can be identified by p16 immunohistology and HPV-DNA testing. Although the prognosis is generally favorable, there has been a lack of success in implementing therapy de-escalation, owing to the heterogeneity of the disease. This underscores the need for precise biomarkers to facilitate patient stratification.

Methods

Our retrospective, multi-institutional study enrolled 906 patients and utilized deep learning to develop an algorithm that accurately predict HPV-association and strongly correlates with prognosis, based on regular H&E slides.

Results

When comparing our algorithm with HPV status, it showed good overall performance (AUROC = 0.83; 95% CI=0.77-0.9). In a subset of the validation cohort (n=639) the implementation of a fixed threshold for filtering resulted in increased AUROC to 0.88, with n=258 cases meeting threshold criteria. The algorithm was compared to the gold standard of HPV-testing in terms of its prognostic relevance and produced better results than the HPV test, indicated by its higher likelihood-ratio test value (LR, 49.23, p<0.001), higher concordance index (0.71), and higher 5-year overall survival rate (OS, 96%, 95% CI=90-100%). In contrast, the HPV test had lower LR (39.72, p<0.001), lower concordance index (0.65), and lower OS (80%, 95% CI=71-90%). The multivariate analysis using three prognostic groups demonstrated good discrimination. The algorithm had a high hazard ratio (HR) of 0.15 (95% CI=0.05-0.44) for the high-risk group, a medium HR of 0.58 (95% CI=0.34-0.98) for the intermediate-risk group, and a significant p-value of 0.043 for the entire group of 211 patients. In comparison, HPV testing had an HR of 0.29 (95% CI=0.15-0.54) and a highly significant p-value of <0.001 for the same group of 211 patients.

Conclusions

Our algorithm can identify patients with OPSCC who have a favorable prognosis using standard hematoxylin and eosin (H&E) histologic slides. In multiple scenarios, our stratification method performs better than the gold standard (p16/HPV-DNA) and could potentially be used to select patients for therapy de-escalation strategies.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

The authors.

Funding

Has not received any funding.

Disclosure

All authors have declared no conflicts of interest.