e18025 Background: The current approach to neck treatment in clinical T1-2 oral cancers is to offer elective nodal dissection to all patients, despite the fact that the majority of patients are pathologically node negative. This is due to the poor predictive ability of clinico-radiological assessment and subsequently poorer survival in those in whom neck dissection is omitted based on this. A robust prediction model for pathological nodal status may allow individualized decisions for neck dissection. Our aim was to develop a multiparameter prediction model to identify pathological node-negative status using machine learning. Methods: We identified 497 patients with cT1-2 oral cancer from a single institutional database from 2011-2018 who underwent primary resection and neck dissection. We compared the sensitivity, positive predictive value and accuracy of prediction of pathologically negative neck from clinico-radiological staging alone vs. a model created from multiple parameters including clinical features (clinico-radiological nodal status, ages, sex, subsite of primary lesion) and pathological features of the resected primary tumor (maximum dimension, depth of invasion, lymphovascular invasion, perineural invasion, grade and margins of resection). The multiparameter model was built from a training dataset of the first 400 patients using an ensemble of logistic regression, random forests and support vector machines. A cohort of 97 patients was used for independent validation. Results: In this cohort 232 (47%) were clinico-radiologically node negative, while 307(62%) were pathologically node negative. The sensitivity, positive predictive value and accuracy of the clinico-radiologically assigned nodal status was 56%, 74% and 61%, while that of the multiparameter machine learning model was 87%, 89% and 89% respectively. The area under curve (AUC) of the clinico-radiological prediction was 0.62 whereas that of the multiparameter predictive model was 0.91. In the validation dataset, 58/62 pathologically node negative patients were predicted correctly by the model. The accuracy of the model on the external validation dataset was 82%. Conclusions: The performance of the multiparameter predictive model was considerably superior to clinico-radiological neck staging for prediction of pathological node negative neck. This could be validated on an independent dataset. This could be considered for prospective clinical evaluation of individualized neck dissection.
Commentary on Elective versus therapeutic neck dissection in node-negative oral cancer
