2829 - Integrating Radiomic Features and Circulating Tumor HPV DNA in Predictive AI Models for Personalized Treatment in HPV-Positive Oropharyngeal Cancer
Presenter(s)
Z. S. Rentiya1, G. L. A. de Sousa2, K. Debamita3, C. McLaughlin4, and K. Wijesooriya5; 1Department of Radiation Oncology, University of Virginia, Charlottesville, VA, 2Department of Physics, University of Virginia, Charlottesville, VA, 3Division of Biostatistics, University of Virginia, Charlottesville, VA, 4Department of Radiation Oncology, University of Virginia Health, Charlottesville, VA, 5University of Virginia, Department of Radiation Oncology, Department of Physics, Charlottesville, VA
Purpose/Objective(s): The rising incidence of HPV-positive oropharyngeal squamous cell carcinoma (OPSCC) presents challenges in treatment-related toxicity and early response assessment. While circulating tumor HPV DNA (ctHPVDNA) has emerged as a promising biomarker for treatment response, its clinical utility is limited by cost, accessibility, and turnaround time. This study aims to develop artificial intelligence (AI)-based predictive models that integrate pre-treatment CT radiomic features and ctHPVDNA data to enhance early response prediction. We investigate whether pre-treatment CT-based radiomic features can predict pre-treatment ctHPVDNA status and treatment-related tumor response with high accuracy. Our hypothesis is that combining radiomic features with mid-treatment ctHPVDNA data improves the early identification of treatment responders and non-responders, facilitating timely intervention and personalized treatment strategies for HPV-positive OPSCC. By leveraging AI-driven approaches, this study seeks to enable cost-effective response prediction, particularly in resource-limited settings, and to inform personalized dose de-escalation strategies.
Materials/Methods: Radiomic features were extracted from pre-treatment CT scans of HPV-positive OPSCC patients using PyRadiomics. Statistical analyses, including Mann-Whitney U and t-tests, identified radiomic features significantly associated with pre-treatment ctHPVDNA binary status. Spearman correlation was used to evaluate associations between radiomic features and post-treatment tumor volume reduction. Principal component analysis (PCA) reduced the dimensionality of significant features to facilitate predictive modeling.
Results: Analysis of 64 CT scans of 32 patients identified 22 radiomic features significantly correlated with pre-treatment ctHPVDNA binary status and 16 features associated with tumor volume reduction post-treatment. PCA reduced these features into six principal components. The integration of radiomic features and ctHPVDNA demonstrated enhanced predictive performance, enabling early stratification of responders and non-responders during treatment.
Conclusion: Our preliminary findings demonstrate the feasibility of combining radiomic features with ctHPVDNA to develop AI-driven predictive models for HPV-positive OPSCC. This approach has the potential to improve early treatment response prediction, guide dose de-escalation strategies, and provide scalable, cost-effective solutions, particularly in resource-constrained settings.