Developing a Prediction Model for Periodontal Disease Using Ordinal Random Forest

Abstract

Listen

Translate article

Background: Periodontitis, a common inflammatory disease affecting oral health, is a major public health concern. Early detection of this disease is highly important in order to achieve less invasive and less costly treatment. The ordinal nature of different levels of periodontitis is usually not considered in many studies. It is often recognized and used as a nominal variable, in which case some information related to the ordinal nature is ignored. The aim of this study was to use continuation ratio logistic and ordinal random forest (RF) models to predict the three stages of periodontitis. Methods: Overall, this study evaluated 300 patients referred to the Periodontology Department of Hamadan University, western Iran, between September 2016 and June 2018. The performance of continuous ratio logistic and ordinal forest models was evaluated using the same set of training and test data to predict different types of periodontitis (gingivitis, localized, and generalized periodontitis) based on input variables. Accuracy, kappa, gamma, Somers’d, and precision were utilized for comparison. Results: The results confirmed the higher predictive ability of the ordinal RF model compared to the logistic continuation ratio model for all scoring indices (accuracy of 0.87 vs. 0.80). Alveolar bone loss, attachment loss, probing pocket depth, simplified oral hygiene index, and plaque index were identified as the most important variables. Conclusion: Due to the ordinal nature of different levels of periodontal disease, the use of accurate prediction models such as ordinal RF is suggested since they can take into account the ordinal nature of the response in predicting and evaluating the effect of important variables. Ordinal RF is a well-suited machine learning technique for developing accurate predictive models of periodontal disease risk.