Clinical Heterogeneity Across Oropharyngeal Cancer Datasets and Implications for Outcome Prediction

Multi-institutional oropharyngeal cancer datasets are increasingly used for research, yet systematic evaluation of data completeness and its clinical impact remains limited. Unrecognized institutional heterogeneity in variable availability, coding conventions, and patient characteristics can introduce substantial bias into combined analyses.

To systematically characterize data availability patterns and clinical heterogeneity across 2 major oropharyngeal cancer datasets, evaluate their impact on survival, and develop evidence-based recommendations for multi-institutional research design.

A cross-sectional analysis and retrospective cohort study was conducted of the RADCURE dataset (2005-2017) from Princess Margaret Cancer Centre (Toronto, Canada) and the HNSCC-Oropharyngeal dataset (2005-2012) from MD Anderson Cancer Center (Houston, Texas). Patients with complete data on patient and disease characteristics were included. Data were analyzed from 2023 to 2026.

Institutional dataset, human papillomavirus (HPV) status, tumor (T) stage, node (N) stage, harmonized treatment modality, smoking status, and age at diagnosis.

Data completeness rates across clinical and end point variables; overall survival assessed using Kaplan-Meier analysis with Peto-Peto weighted log-rank tests, restricted mean survival time, and multivariable Cox proportional hazards regression stratified by HPV status; machine learning discrimination and calibration for 1-, 5-, and 10-year survival prediction.

The final analytic cohort comprised 1972 participants (median [IQR] age, 59.5 [53.4-67.0] years; 347 female [17.6%] and 1625 male [82.4%] individuals), 1482 from the RADCURE and 490 from the HNSCC-Oropharyngeal datasets. Of the total, 1263 had HPV-positive disease (64.0%). Substantial institutional differences were observed across demographic characteristics, HPV distribution, and treatment modality. After comprehensive adjustment, the hazard ratio for treating institutions was 1.19 (95% CI, 0.93-1.52), not excluding a clinically meaningful effect. HPV status violated the proportional hazards assumption, and its prognostic effect was quantified using restricted mean survival time, demonstrating a clinically meaningful survival advantage for patients with HPV-positive disease in both datasets. Machine learning ensemble models achieved moderate to good discrimination (1-year area under the curve [AUC], 0.820; 5-year AUC, 0.777; 10-year AUC 0.786) with institution consistently ranking as a low-importance predictor.

In this cross-sectional analysis and cohort study, multi-institutional oropharyngeal cancer datasets exhibited substantial heterogeneity requiring systematic harmonization prior to combined analyses. Systematic harmonization and rigorous assumption testing should be prerequisites for multi-institutional oropharyngeal cancer analyses; after appropriate adjustment, treating institution did not independently predict survival.