Machine Learning to Predict Delays in Adjuvant Radiation following Surgery for Head and Neck Cancer

Objective: To apply a novel methodology with machine learning (ML) to a large national cancer registry to help identify patients who are high risk for delayed adjuvant radiation. Study Design: Observational cohort study. Setting: National Cancer Database (NCDB). Subjects and Methods: A total of 76,573 patients were identified from the NCDB who had invasive head and neck cancer and underwent surgery, followed by radiation. The model was constructed from 80% of the patient data and subsequently evaluated and scored with the remaining 20%. Permutation feature importance analysis was used to understand the weighted model construction. Results: A total of 76,573 patients met inclusion and exclusion criteria. Our ML model was able to predict whether patients would start adjuvant therapy beyond 50 days after surgery with an overall accuracy of 64.41% and a precision of 58.5%. The 2 most important variables used to build the model were treating facility and urban versus rural demographics. Conclusion: Statistics can provide inferences within an overall system, while ML is a novel methodology that can make predictions. We can identify patients who are “high risk” for delayed radiation using information from >75,000 patient experiences, which has the potential for a direct impact on clinical care. Our inability to achieve greater accuracy is due to limitations of the data captured by the NCDB, and we need to continue to identify new variables that are correlated with delayed radiation therapy. ML will prove to be a valuable clinical tool in years to come, but its utility is limited by available data.