Development of a patient reported outcomes based machine learning model to predict recurrences in head and neck
Christopher M K L Yao1, Katrina Hueniken2, Shao Hui Huang3
1Department of Otolaryngology - Head and Neck Surgery, Princess Margaret Cancer Centre, University Health Network, University of Toronto, Toronto, Ontario, Canada; Department of Mechanical and Industrial Engineering, University of Toronto, Toronto, Ontario, Canada.
Introduction:
Recurrence rates among Head and Neck Cancer (HNC) patients are high, with earlier detection associated with improved survival. Patient-reported outcomes (PROs) have increasingly been found to predict patient care needs. Here, we examine whether PROs specific to HNC patients or general can predict disease progression using Machine Learning (ML) algorithms.
Methods:
This was an analysis of 1,302 HNC patients,including patients who completed at least one MD Anderson Symptom Inventory (MDASI) or Edmonton Symptom Assessment Score (ESAS) questionnaire 3 months following curative intent treatment. ML models, including least absolute shrinkage and selection operator (LASSO) logistic regression and Random Forest (RF) were applied to baseline or longitudinal PRO changes to predict recurrences. Predictive performances were assessed via area under the receiver-operating curve, computed with 10-fold cross-validation. Relative variable importance were computed with average decrease in out-of-bag prediction accuracy of each tree.
Results:
Disease recurrence occurred in 9.5 % (n = 123) of HNC patients. Baseline post-treatment MDASI, RF models demonstrated an area under the curve (AUC) approximating 0.675, sensitivity of 0.83 and specificity of 0.58 with pain, speech, and dry mouth as key variables. When stratifying patients by HPV status, our non-HPV model based on pain, distress, and mood yielded an AUC of 0.71 at 3 months and 0.70 at 6 months.
Conclusion:
ML models using HNC specific PROs can identify patients at high risk for disease progression with moderate accuracy. Prospective studies with larger dataset and further analysis are needed to refine these models and evaluate their potential in guiding post-treatment surveillance.
