Development of a machine learning based method for the diagnosis of Bruxism with photoplethysmography signal

Bruxism is a jaw muscle activity that can occur during sleep or wakefulness and is characterized by the habit of grinding or clenching the teeth and/or a thrusting movement that displaces the jaw forward or sideways. It is of significant clinical importance due to its potential to cause or exacerbate dental damage, temporomandibular joint disorders, muscle fatigue, orofacial pain, headaches, and sleep disruption. Left untreated, this can lead to more serious complications such as irreversible tooth wear, musculoskeletal strain, chronic pain, and psychological distress. Given its multifactorial nature and complex etiology, early and accurate diagnosis of sleep bruxism is essential for effective intervention and prevention of long-term health consequences. Polysomnography (PSG), which enables the definitive diagnosis of bruxism, is the gold standard in the diagnosis of many sleep disorders other than bruxism, too. In diagnosis with PSG, the individual must sleep throughout the night under expert control with a large number of electrodes connected to the body. Because the individual is not in a home environment and many electrodes are connected to the body, the sleep disorder may show different characteristics and cause errors in diagnosis. PSG is available in sleep laboratories. The number of sleep laboratories is limited, and bed capacities are low. Therefore, appointments are scheduled for later dates. In addition, since sleep laboratories are generally used to diagnose life-threatening conditions such as sleep apnea (SA), individuals with sleep bruxism cannot be referred to sleep laboratories for definitive diagnosis. Thus, individuals with bruxism receive a possible or probable diagnosis through clinical examination and patient history. Although bruxism is not a life-threatening sleep disorder such as SA, it can affect the quality of life and general health of the person. When left untreated, it can lead to increased physical damage in the long term and trigger other life-threatening disorders. Therefore, diagnosing bruxism early and starting treatment as soon as possible is important. Developing practical systems that enable the diagnosis of bruxism with high accuracy will make it possible to diagnose bruxism before dental or general health problems caused by bruxism occur, improve the quality of life of individuals, and make equal opportunity in health possible. Sleep disorders can affect each other. Bruxism that can be diagnosed early will prevent the emergence of other disorders, too. Considering all these, this thesis aims to develop an artificial intelligence (AI)-based model to diagnose bruxism accurately. For this purpose, photoplethysmogram (PPG), which can be easily obtained from the body without causing discomfort to the individual, was used. Considering that using more than one signal would increase classification success, the heart rate variability (HRV) signal obtained from PPG, which is practical to obtain, was preferred as the second signal. Thus, developing two-signal diagnostic models with a single signal obtained from the body was possible. The data in the study were obtained from the PSG records of individuals admitted to the sleep laboratory with suspected SA. In this context, the data of 17 individuals, 12 with SA and five without sleep apnea (Non-SA), were retrospectively analyzed. PSG recordings containing 33 channels of data are in 30-second epochs. An expert physician analyzed electromyogram (EMG), electrooculogram (EOG), and electroencephalogram (EEG) data in the recordings according to the American Academy of Sleep Medicine (AASM) criteria, and each epoch was labeled as "sleep" or "awake." After discarding waking episodes, only sleep episodes were evaluated. An expert doctor labeled the jaw EMG recordings in these episodes according to the AASM bruxism criteria, and bruxism data were obtained. Similarly, EEG arousals were labeled according to AASM criteria, and control data were obtained. The control data consists of micro-arousals, respiratory effort- related arousals (RERA), and periodic leg movement (PLM) arousals. PPG data at the time intervals of the labeled data were obtained; a PPG data set consisting of bruxism and controls was obtained. The data set consists of 623 bruxism and 875 control labeled PPG data. Signal processing steps were then performed on these data. First, artifacts were cleaned from the data using an appropriate filter. Then, HRV was obtained from the filtered PPG signal. We extracted 38 characteristics and 29 statistical features from PPG and 29 statistical features from HRV. The aim was to better characterize the signals by keeping the number of features high. With the Principal Component Analysis (PCA) method, the database with 96 features could be represented 100% with 47 features. Thus, dimension reduction was possible. Fisher scores, which express the association level of 47 features with the result label, were calculated using the Fisher feature selection algorithm. According to Fisher's scores, the features were ranked from highest to lowest. Starting with the feature with the highest Fisher score and increasing the number of features one by one, 47 different data sets were obtained. Classification models were created using four machine learning methods: Support vector machine (SVM), k-nearest neighbor (kNN), Ensemble, and Neural Network (NN). It was aimed to identify the model that offers high performance with the minimum number of features. While creating the models, 5-fold cross-validation and hyperparameter optimization were performed. Classification performances of the models were evaluated using accuracy, sensitivity, selectivity, precision, F-score, Kappa, and area under the ROC curve (AUC) performance criteria. The Leaving-One-Out (LOO) method was used to observe whether the performances obtained in the models were independent of the individuals in the data set used. Bruxism-Micro-arousal, bruxism-RERA, and bruxism-PLM arousal classifications were also designed to see how well the micro-arousal, RERA, and PLM arousal labels in the control data could be distinguished from the bruxism labels. Of the 17 individuals in the study, 12 had SA, and five were Non-SA. We performed bruxism-control classification separately for individuals with and Non-SA and observed whether the performance obtained was generalizable. All labeling in the study was performed by a specialist physician who had previously worked in a sleep laboratory and experienced sleep staging (sleep-awake) and bruxism-arousal labeling. Evaluation of the labelings by more than one expert may increase the reliability of the study. In addition, the performance of machine learning models can improve as the sample size increases. Larger datasets can increase the generalizability of the model. We tried to minimize the adverse effects of sample size by performing LOO and 5-fold cross-validation. According to the performance results obtained: The bruxism data were successfully discriminated from the control data both in the whole dataset (92,8% accuracy, 0,9727 AUC) and when the data of SA (92,3% accuracy, 0,9722 AUC) and Non-SA (92,9% accuracy, 0,9708 AUC) individuals were considered separately. The performance values obtained with SA and Non-SA datasets confirmed the performance of the classification obtained using the whole dataset. An accuracy of 92,8% and an AUC of 0,9727 before PCA (96 features) and an accuracy of 92,4% and an AUC of 0,9676 after PCA (47 features) were achieved. These results confirmed that the dataset obtained with PCA is 100% representative. When only one feature (number 86) was used, the SVM method achieved 89,1% accuracy and 0,9374 AUC. Among 47 features, the Ensemble method achieved the highest performance using 33 features with 92,7% accuracy and 0,9685 AUC value. Close to the highest performance, the Ensemble model with only 15 features achieved 92% accuracy and 0,9691 AUC value. In addition, the Ensemble model with fewer features (10 features) could perform high-performance classification with 91,1% accuracy and 0,96 AUC. Between the LOO method and the models where each individual is used as test data, the accuracy rates vary between 78,8% and 97,1%, and the AUC value varies between 0,88 and 1. The performance results confirm that the classification success is independent of the individuals. Bruxism could be classified as having 93% accuracy with micro-arousal, 89,3% accuracy with RERA, and 87% accuracy with PLM arousal. These findings show that bruxism can be successfully classified regardless of the different labels of the control data and that the bruxism-control classification is generalizable for each label. The results showed that AI-assisted models developed using PPG and HRV can be effective in early, reliable, and user-friendly diagnosis of bruxism with high classification performance. This approach will provide an important alternative to existing diagnostic systems and support equal opportunities in health. This study is also important as it is the first study in the literature to propose using PPG in diagnosing bruxism and is expected to shed light on future studies.