Classificação das desordens temporomandibulares com o uso do algoritmo k-nearest neighbors aplicado à dinâmica mandibular

Abstract

A number of different studies in scientific literature use machine learning and pattern recognition for detecting lesions, abnormalities, tumors, among others. However, there exists a lack of studies that have applied this method for recognizing quantitative patterns of movement of the mandible in TMD studies. As such, the specific objective of the present study was to apply four different algorithms for the classification of data arising from mandibular kinematics, in order to evaluate the parameters of sensitivity, specificity, accuracy and precision in classifying models. A convenience sample was used in this study. Forty participants that underwent clinical assessment along with the use of the Research Diagnostic Criteria for Temporomandibular Disorders were divided into three groups, those being: arthropathies (GART, 10 participants, 40% men), myopathies (GMIO, 10 participants, 30% men), and control group (GC, 20 participants asymptomatic, 25% men). The participants were instructed to perform functional movements of unassisted maximal mouth opening and closing, right and left laterality, protrusion with maximum extension, as well as with contact between teeth. Mandibular motion tracking was registered using an optoelectronic system composed of infrared cameras and reflective markers placed on specific points of the face. The movements were analyzed on the following reference axes (Cartesian system): X - mid - lateral, Y - vertical, Z – anteroposterior. Significant differences were found in GCxGART – unassisted maximal mouth opening and closing projected on the Y axis (AFY), GCxGMIO – unassisted maximal mouth opening and closing on the X axis (AFX), and in the measurements DLAX (lateral deviations with mouth open projected on the X axis), DLFX (lateral deviations with mouth closed projected on the X axis), and in the execution speeds for all movements for both comparisons between the groups. In regards to the comparison GARTxGMIO, a significant difference was found for the lateral deviations of protrusion projected on the X axis (DLPX) using the ‘Conover-Iman test of multiple comparisons using rank sums’ with Bonferroni correction (p<0.05). In conclusion, the movements for maximal mouth opening in individuals with TMD tend to present greater deviation when compared to asymptomatic individuals, in addition to the reduction in the speed for executing the movement. The classifying algorithms used were: k-Nearest Neighbors (k-NN), Random Forest, Naïve Bayes and Support Vector Machine, where the results obtained using the k-NN algorithm were the most satisfactory, thus concluding that this methodology is able to separate the groups with acceptable levels of sensitivity, specificity, precision and accuracy.