Adapting an Evidence-based Diagnostic Model for Predicting Recurrence Risk Factors of Oral Cancer

Although the relationship between prognosis and oral cancer has been extensively investigated, its impact on recurrence and surgical margin has not been well studied. Clinical evaluation of a positive surgical margin in recurrent oral cancer is often challenging. The aim of this study was to propose an evidence-based diagnostic model using machine learning techniques for the prediction of risk factors of recurrent oral cancer. In addition, the performance of each technique was evaluated using accuracy, sensitivity, specificity, Fallout, F1 score, and Matthews correlation coefficient (MCC). An oral cancer dataset was provided by cancer registries of three hospitals in Taiwan. Of the 1,428 patients included in the current study, each patient in the dataset had 20 predictor variables. The results indicated that the KSTAR technique showed the best performance compared with other techniques. The GainRaito (RT) method was used in the screening to exclude five insignificant variables. The KSTAR technique also showed larger values for accuracy (77.04%), recall (77.98%), specificity (75.48%), Fallout (36.62%), F1 score (81.17%), and MCC (50.54%). Furthermore, the important risk factors for predicting recurrence in relation to the surgical margin in oral cancer were pathologic stage, behavior code, and lifestyle factors (smoking and betel nut chewing). Application of this proposed diagnostic model may facilitate targeted intervention to reduce the incidence of recurrence; however, our results suggest that adaptive machine learning techniques require incorporation of significant variables for optimal prediction