Abstract
AbstractType 2 Diabetes Mellitus(T2DM) is a debilitating condition with a number of complications including those of the oral cavity which can further deteriorate patient’s general and oral health related quality of life (OHRQoL). Machine Learning (ML) can help assign an individual’s propensity to develop poor OHRQoL, given a set of variables, and at the same time identify the most important features contributing to this outcome. Previously inferential statistical methods have attempted to explain this, albeit with limited success. The aim of this cross sectional study is to determine the impact on OHRQoL in T2DM patients, and identify features most likely to be associated with this outcome and to compare ML and DL analytical methods with inferential statistics. Twelve-hundred T2DM patients were subjected to OHRQoL and demographic data questionnaires and WHO Oral Health Assessment form. K-means Clustering was performed to label individuals as having or not having an impact on OHRQoL. Class imbalance was addressed by undersampling of the majority class using informed subset selection. Further, using the collected data as input features we developed ML algorithms (Naive Bayes(NB), Random Forest(RF), Logistic Regression(LR), Kernel Support Vector Machine(SVM) and Artificial Neural Network(ANN)), to accurately classify individuals with or with-out poor oral health related quality of life (OHRQoL) and utilized SHapley Additive exPlanations (SHAP) analysis for feature importance. The best performing model was SVM (AUC=0.983; Sensitivity=1) for classifying the patients into into poor OHRQoL. SHAP values were highest for Age, Prosthetic Need, Tobacco use and years since onset of diabetes. Features closely related to diabetes, that is, periodontal pockets and loss of attachment were not identified as relevant by inferential statistics, but were deemed as important features associated with poor OHRQoL by SHAP analysis.
Publisher
Cold Spring Harbor Laboratory