The following is a summary of “Spatially-aware clustering improves AJCC-8 risk stratification performance in oropharyngeal carcinomas,” published in the September 2023 issue of Oral Oncology by Canahuate et al.
The objective is to evaluate the efficacy of machine learning tools that integrate spatial information, such as disease location and patterns of lymph node metastasis, for predicting survival and toxicity in HPV+ oropharyngeal cancer (OPC). The retrospective collection of 675 HPV+ OPC patients treated with curative intent with IMRT at MD Anderson Cancer Center between 2005 and 2013 with IRB approval The identification of risk stratifications incorporating patient radiometric data and lymph node metastasis patterns via an anatomically adjacent representation with hierarchical clustering.
Using independent subsets for training and validation, these clusterings were combined into a 3-level patient stratification and incorporated alongside other known clinical features in a Cox model for predicting survival outcomes and logistic regression for toxicity. The identification of four distinct groups and their combination into a three-level stratification structure The addition of patient stratifications to predictive models for 5-year overall survival (OS), 5-year recurrence-free survival (RFS), and radiation-associated dysphagia (RAD) enhanced model performance as measured by the area under the curve (AUC) in a consistent manner.
Test set AUC enhancements were 9% for predicting OS, 18% for predicting RFS, and 7% for predicting RAD over models with clinical covariates. For models with both clinical and AJCC covariates, the AUC for OS, RFS, and RAD improved by 7%, 9%, and 2%, respectively. Incorporating data-driven patient stratifications significantly improves survival and toxicity outcome prognosis over clinical staging and clinical covariates alone. These stratifications generalize well across cohorts, and sufficient information exists to reproduce these clusters.