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

Area of Science:

• Oncology
• Machine Learning
• Data Science

Background:

• Extranodal extension (ENE) is a critical prognostic factor in various cancers.
• Accurate prediction of ENE is essential for treatment planning and patient management.
• Machine learning (ML) offers potential for improving ENE prediction accuracy.

Purpose of the Study:

• To evaluate the performance of ML models trained on large public databases for predicting ENE.
• To assess the generalizability of these ML models on a smaller, independent single-institution cohort.
• To determine the feasibility of using ML for ENE prediction in clinical practice.

Main Methods:

• ML algorithms were trained and validated using data from the National Cancer Data Repository (NCDB) (n=5551).
• Six key ENE-related variables were utilized as predictors.
• Model performance was assessed on an independent cohort from WVU (n=62) to simulate real-world validation.

Main Results:

• ML models achieved an average Area Under the Receiver Operating Characteristic Curve (AUC ROC) of 0.74 for ENE prediction.
• AdaBoost, XGBoost, and ensemble models showed promising performance, with AUC ROCs ranging from 0.71 to 0.75.
• Models demonstrated reasonable precision and specificity, though sensitivity for ENE prediction was modest.

Conclusions:

• ML algorithms trained on large oncologic datasets (NCDB) can effectively predict ENE.
• The developed ML models exhibit good generalizability and accurately predict ENE in a real-world clinical setting.
• This study confirms the feasibility of employing ML for ENE prediction, supporting its potential clinical utility.