Machine learning models for analysis of biomarkers of chronic periodontitis

Abstract

Periodontitis is a multifactorial inflammatory chronic disease initiated by dysbiosis of the commensal microbiota of the oral cavity. With the advent of the multi-ohm approach, which creates datasets with multiple features, machine learning algorithms have become an important technique in translational research. The most effective methods for identifying highly specific interactions of biomarkers with clinical parameters and their implementation in interpretable accurate clinical diagnoses are machine learning algorithms. The aim of the study was to evaluate laboratory diagnostic indicators that are significant for predicting the severity of periodontitis using machine learning models. We examined 133 patients aged 22 to 73 years with a diagnosis of chronic periodontitis, as well as 53 people without periodontal pathology. After the examination and assessment of the condition of the periodontium of patients, a biological sample of the periodontal fluid was taken. Real-time PCR was used to evaluate periodontopathogenic microflora and mRNA of pro-inflammatory cytokines. Cells of innate and acquired immunity were evaluated in mixed saliva. Machine learning was performed using logistic regression algorithms, naive Bayes classifier, ‘the Random Forest’ with 25 % training. When trained to 25 %, the naive Bayes classifier showed an accuracy of 23 %, logistic regression – 29 %. ‘The Random Forest’ showed 100 % accuracy and selected the following biomarkers associated with periodontitis severity: Porphyromonas endodontalis; CD 3+, CD 14+, CD 19+5–B 27+ cells; mRNA IL-1β, IL-10, IL-18, GATA3, TNFa, TLR 4. In chronic periodontitis, there is a relationship between local indicators of the immune-inflammatory process, such as mRNA of pro-inflammatory cytokines, cells of the immune system, and the severity of pathology. In the analysis of a multifactorial disease such as chronic periodontitis, a machine learning model optimized for a large heterogeneous data set with a large patient sample should be used.