• AI-Driven EMG Signal Profiling of Facial Muscles for Smart Detection of Dental Malocclusion During Full Occlusion
  • setareh tabasi,1 nasim kharazminezhad,2,* saniya tabasi,3
    1. Department of Biomedical Engineering, South Tehran Branch, Islamic Azad University, Tehran,Iran.
    2. Department of Biomedical Engineering, South Tehran Branch, Islamic Azad University, Tehran,Iran.


  • Introduction: Occlusion refers to the alignment of the upper and lower teeth during contact. When abnormalities—either congenital or acquired—affect this alignment, the condition is termed malocclusion. Malocclusion ranks among the most prevalent musculoskeletal disorders [1–3]. In cases where it impairs daily function, it is categorized as a skeletal occlusion disorder and requires timely therapeutic intervention [4–6]. While mild forms can be managed through orthodontic treatments [5–7], more advanced cases may necessitate orthognathic surgery [8–9]. Severe malocclusions often demand comprehensive diagnostic imaging, including CBCT or CT scans [5]. Nonetheless, analyzing EMG signals of facial muscles can serve as a promising non-invasive tool for early detection. Building upon prior research, this study introduces a novel method for classifying occlusal deviations by integrating signal processing techniques with machine learning to enhance previous methodologies while mitigating their limitations [10–11]. Malocclusion is typically categorized into three classes. Class I denotes normal alignment with proper contact between upper and lower teeth. In Class II, the mandibular arch is positioned posterior to the maxillary arch, whereas in Class III, it is positioned anteriorly [12]. Among the primary muscles involved in jaw movement are the masseter and temporalis muscles [13 ]. Power spectral density (PSD) analysis of EMG signals from these muscles is one of the established techniques for malocclusion assessment [14], as individuals exert different levels of muscular effort based on occlusal class. Earlier studies were often constrained by limited instrumentation, resulting in suboptimal signal analysis. Recent advances in high-speed signal processing have paved the way for more accurate and efficient assessment methods. In this study, we explore the correlation between occlusal class and muscle activity during maximal intercuspation. Specifically, we extract six key features—such as frequency shifts, fatigue duration, and fatigue coefficients—for the masseter and temporalis muscles, and evaluate their performance in classification tasks.
  • Methods: To conduct this study, surface electromyographic (EMG) signals were recorded from 32 male subjects aged between 22 and 27 years, who were referred to the dental clinic of the Biomedical Engineering Department at the Research Institute of Clinical Medicine, Baqiyatallah University, Tehran, Iran. The EMG signals were acquired from the temporalis and masseter muscles. 2.1. Signal Acquisition EMG recordings were performed using an MP150 system equipped with two dedicated EMG modules. Ag/AgCl electrodes (medium-sized, Skintact brand, model E503s) were used for surface signal acquisition. Recordings were made for 5 minutes at a sampling frequency of 1000 Hz, generating two time-series signals (one for each muscle) with 300,000 data points each. The signals were saved in .mat format to facilitate analysis using MATLAB software. Further statistical evaluations were performed using SPSS. To ensure consistency, only male participants with no noticeable facial skin laxity—especially around the perioral region—were included. Prior to signal acquisition, subjects underwent dental classification by a licensed dentist through manual examination of occlusion type. The skin at the electrode placement sites was cleansed with alcohol. A total of five surface electrodes were placed on each subject’s face: VIN+ and VIN– on the forehead and others on the temporalis and masseter regions. Participants were instructed to perform a maximum intercuspation bite during the recording. Based on the clinical assessment, 17 individuals were categorized as Class I (normal occlusion), 6 as Class II, and 9 as Class III (see Fig. 1).
  • Results: Among the extracted features, wavelet detail coefficients (Wd) were analyzed, including the following metrics: MedWdT1, MeanWdT4, MeanWdTF3, MedWdM2, MedWdMF2, MeanWdMF1. These coefficients were selected based on their discriminative potential, as detailed in Appendix 10. A total of 34 features were initially extracted for the purpose of classification. Due to the limitations of using high-dimensional feature sets in neural network models, Principal Component Analysis (PCA) was applied to reduce the number of variables. PCA was implemented using the Singular Value Decomposition (SVD) technique. The procedure is outlined in equations (1) to (4), beginning with mean-centering the data: • Compute the mean of each feature vector – (Equation 1) (Subsequent steps and equations would follow here as referenced in the original manuscript.) PCA and Statistical Validation Following initial preprocessing, the dataset underwent mean normalization, and the covariance matrix was computed. The next steps involved identifying the eigenvalues (Equation 3), diagonalizing the matrix, and applying Singular Value Decomposition (SVD) to reduce the dimensionality of the dataset. The feature space was thus transformed from n dimensions to a reduced set of k principal components. From the reduced space, two key components were selected and further examined for statistical significance. Ultimately, eight core features were identified and retained for use in the classification model. To validate the significance of these features, one-way ANOVA was performed, followed by post hoc tests including Tukey’s HSD and LSD (Least Significant Difference) analysis. All features showed statistically significant differences (p < 0.05) across the occlusion classes. Pairwise comparisons using independent two-sample t-tests also confirmed these findings, demonstrating consistent results across all examined features (see Table 1).
  • Conclusion: Through the analysis of 34 potential variables across a sample of 32 subjects, we successfully identified eight principal features that demonstrated strong discriminatory power for classifying malocclusion types. These features were validated through comprehensive statistical analysis, confirming their reliability and effectiveness. We propose this classification framework as a practical, low-cost, and non-invasive diagnostic tool that can be integrated into dental practice. It offers clinicians a safe and efficient method for early detection and categorization of occlusal abnormalities without subjecting patients to discomfort or exposure to radiation.
  • Keywords: Electromyography, Muscle Activity, Malocclusion, Maximum Intercuspation, Dental Classification