Machine Learning Strategies to Distinguish Oral Cancer from Periodontitis Using Salivary Metabolites

Development of non-invasive diagnostic tests which can immediately yield guidance in clinics, reducing the number of samples sent for laboratory testing, has the potential to revolutionize medical diagnostics improving patient outcomes, doctor’s workloads, and healthcare costs. It is difficult to imagine a testing method less invasive than measurement of salivary metabolites by swab or spit. Coupling data from this incredibly convenient measurement to an automated decision-making engine can provide clinicians with immediate feedback on the status of their patients. The aim of this research is to lay the foundation for a system, which when employed in dental practices can help to stratify between patients with periodontitis and oral cancers and illuminate metabolic networks important to both diseases. We built machine learning models trained on QSAR descriptors of metabolites whose concentrations changed drastically between the two diseases and used these same metabolites to illuminate networks important in the development of each disease. The Neural Network developed in TensorFlow performed best, achieving 81.29% classification accuracy between metabolites of periodontitis and oral cancers, respectively. We compared effects of two attribute selection methods, ranking by Correlation and Information Gain coefficients, on the accuracy of the models and employed principal component analysis to the data for dimensionality reduction before training. Models trained on attributes ranked by Information Gain coefficients regularly outperformed those ranked by Correlation coefficients across machine learning methods while relying on fewer principal components.