Detecting Extranodal Extension in HPV-Related Oropharyngeal Cancer via Machine Learning.

Objective: To assess the performance of machine learning (ML) models trained on large publicly available databases in predicting extranodal extension (ENE) within a smaller single-institution cohort.

Methods: Patient data were derived from two datasets. The first, the NCDB containing 5551 patients, was used to train and validate the ML algorithms. NCDB data was split into training (80%) and testing (20%) sets. To ensure fair comparison of prediction performance, data were randomly divided. The second, the WVU cohort containing 62 patients, was used as an independent test set for validation and simulation of a real-world scenario. Six key ENE-related variables (predictors) were selected based on importance and data availability.

Results: The generalizability of the ML models was validated using real data from our institute with an average AUC ROC of 0.74. Of all the models, AdaBoost had reasonable performance: precision of 72%, NPV of 0.74%, specificity of 0.81, AUC ROC of 0.74. Similarly, the XGBOOST and ensemble models demonstrated the strongest combination of sensitivity and precision, critical for correctly predicting ENE, with AUC ROC ranging from 0.71 to 0.75. Most models demonstrated modest sensitivity in predicting ENE.

Conclusion: This study proved the feasibility of using ML algorithms to predict ENE by using oncologic variables from the NCDB as a training dataset. These trained models generalize well, such that they accurately predicted ENE on a real-world single institution dataset.