Development and Evaluation of an Interdisciplinary Periodontal Risk Prediction Tool Using a Machine Learning Approach

Abstract

Periodontitis (PD) is a major public health concern which profoundly affects oral health and concomitantly, general health of the population worldwide. Evidence-based research continues to support association between PD and systemic diseases such as diabetes and hypertension, among others. Notably PD also represents a modifiable risk factor that may reduce the onset and progression of some systemic diseases, including diabetes. Due to lack of oral screening in medical settings, this population does not get flagged with the risk of developing PD. This study sought to develop a PD risk assessment model applicable at clinical point-of-care (POC) by comparing performance of five supervised machine learning (ML) algorithms: Naïve Bayes, Logistic Regression, Support Vector Machine, Artificial Neural Network and Decision Tree, for modeling risk by retrospectively interrogating clinical data collected across seven different models of care (MOC) within the interdisciplinary settings. Risk assessment modeling was accomplished using Waikato Environment for Knowledge Analysis (WEKA) open-sourced tool, which supported comparative assessment of the relative performance of the five ML algorithms when applied to risk prediction. To align with current conventions for clinical classification of disease severity, predicting PD risk was treated as a ‘classification problem’, where patients were sorted into two categories based on disease severity and ‘low risk PD’ was defined as no or mild gum disease (‘controls’) or ‘high risk PD’ defined as moderate to severe disease (‘cases’). To assess the predictive performance of models, the study compared performance of ML algorithms applying analysis of recall, specificity, area under the curve, precision, F-measure and Matthew’s correlation coefficient (MCC) and receiver operating characteristic (ROC) curve. A tenfold-cross validation was performed. External validation of the resultant models was achieved by creating validation data subsets applying random selection of approximately 10% of each class of data proportionately. Findings from this study have prognostic implications for assessing PD risk. Models evolved in the present study have translational value in that they can be incorporated into the Electronic Health Record (EHR) to support POC screening. Additionally, the study has defined relative performance of PD risk prediction models across various MOC environments. Moreover, these findings have established the power ML application can serve to create a decision support tool for dental providers in assessing PD status, severity and inform treatment decisions. Further, such risk scores could also inform medical providers regarding the need for patient referrals and management of comorbid conditions impacted by presence of oral disease such as PD. Finally, this study illustrates the benefit of the integrated medical and dental care delivery environment for detecting risk of periodontitis at a stage when implementation of proven interventions could delay and even prevent disease progression. Keywords: Periodontitis, Risk Assessment, Interprofessional Relations, Machine learning, Electronic Health Records, Decision Support Systems