HPV Subanalysis
The vast majority of patients included in the study were with missing
HPV status (78%). An additional analysis was conducted excluding all
patients with missing HPV information to further explore the data. The
machine learning models were redeveloped and retested using this new
population. In all, 4,284 patients were included in the subset analysis.
Full demographic information is detailed in Supplemental Table 1. Of
note, primary treatment with surgical resection was slightly more common
than primary treatment with radiation (60% vs. 40%). In Supplemental
Table 2, the performance of the machine learning models is displayed.
The decision forest again yielded the strongest model with an AUC of
75% (95% CI, 72% to 79%), accuracy of 72%, precision of 68%, and
recall of 65%. The most important factor found by way of the PFI scores
was patient clinical T- classification. This was closely followed by the
same tumor descriptors including primary site of the cancer, clinical N-
classification, and grade. The full results of the PFI analysis are
displayed in Supplemental Table 3.