Novel Machine Learning Model to Predict Interval of Oral Cancer Recurrence for Surveillance Stratification.

Objective(s): We aimed to develop a machine learning (ML) model to accurately predict the timing of oral squamous cell carcinoma (OSCC) recurrence across four 1-year intervals.

Methods: Patients with surgically treated OSCC between 2012-2018 were retrospectively identified from the Yale-New Haven Health system tumor registry. Patients with known recurrence or minimum follow-up of 24 months from surgery were included. Patients were classified into one of five levels: four 1-year intervals and one level for no recurrence (within 4 years of surgery). Three sets of data inputs (comprehensive, feature selection, nomogram) were combined with 4 ML architectures (logistic regression, decision tree (DT), support vector machine (SVM), artificial neural network classifiers) yielding 12 models in total. Models were primarily evaluated using mean absolute error (MAE), lower values indicating better prediction of 1-year interval recurrence. Secondary outcomes included accuracy, weighted precision, and weighted recall.

Results: 389 patients met inclusion criteria: 102 (26.2%) recurred within 48 months of surgery. Median follow-up time was 25 months (IQR: 15-37.5) for patients with recurrence and 44 months (IQR: 32-57) for patients without recurrence. MAE of 0.654% and 80.8% accuracy were achieved on a 15-variable feature selection input by 2 ML models: DT and SVM classifiers.

Conclusions: To our knowledge, this is the first study to leverage multiclass ML models to predict time to OSCC recurrence. We developed a model using feature selection data input that reliably predicted recurrence within 1-year intervals. Precise modeling of recurrence timing has the potential to personalize surveillance protocols in the future to enhance early detection and reduce extraneous healthcare costs.

Level Of Evidence: 3 Laryngoscope, 133:1652-1659, 2023.