Risk factor assessments of temporomandibular disorders via machine learning

This study aimed to use artificial intelligence to determine whether biological and psychosocial factors, such as stress, socioeconomic status, and working conditions, were major risk factors for temporomandibular disorders (TMDs). Data were retrieved from the fourth Korea National Health and Nutritional Examination Survey (2009), with information concerning 4744 participants’ TMDs, demographic factors, socioeconomic status, working conditions, and health-related determinants. Based on variable importance observed from the random forest, the top 20 determinants of self-reported TMDs were body mass index (BMI), household income (monthly), sleep (daily), obesity (subjective), health (subjective), working conditions (control, hygiene, respect, risks, and workload), occupation, education, region (metropolitan), residence type (apartment), stress, smoking status, marital status, and sex. The top 20 determinants of temporomandibular disorders determined via a doctor’s diagnosis were BMI, age, household income (monthly), sleep (daily), obesity (subjective), working conditions (control, hygiene, risks, and workload), household income (subjective), subjective health, education, smoking status, residence type (apartment), region (metropolitan), sex, marital status, and allergic rhinitis. This study supports the hypothesis, highlighting the importance of obesity, general health, stress, socioeconomic status, and working conditions in the management of TMDs.


Methods
Study population and sampling. Data were obtained from the fourth Korea National Health and Nutritional Examination Survey, that is, KNHANES IV-3 2009 26 . KNHANES is a nationwide annual cross-sectional survey conducted by the Division of Chronic Disease Surveillance of the Korea Disease Control and Prevention Agency. The survey collects information from approximately 10,000 nationally-representative and noninstitutionalized civilians in Korea regarding their socioeconomic status, health-related behaviors, quality of life, healthcare utilization, anthropometric parameters, biochemical and clinical profiles for noncommunicable diseases, and dietary intake. The data are de-identified and publicly available upon request. The requirement for ethical approval from the institutional review board of the Asan Medical Center was waived (waiver number: 2020-0362). All methods were conducted in accordance with relevant guidelines and regulations. The final sample consisted of 4744 participants aged ≥ 19 years with information about TMDs. According to the KNHANES IV 2009 survey, participants who responded "yes" to the question "Have you had symptoms related to a TMD such as jaw pain, joint sound, and/or mouth opening limitations?" were considered as selfreported TMD (r-TMD) patients. Those who responded "yes" to the question "Have you been told by a doctor that you have a TMD?" were considered as TMD patients as diagnosed by a doctor (d-TMD). Those who responded "no" to these questions were categorized in a control group of those without TMDs. Information about both r-TMD and d-TMD was available only in the KNHANES IV-3 survey conducted in 2009.
In total, 37 independent variables were analyzed from the survey. These include: (1) demographic factors, (2) socioeconomic status, (3) mental stress/working environment, (4) biological variables, and (5) comorbidities. The full list of variables included is shown in Table 1.
Analysis. Six artificial intelligence approaches were used for identifying the factors associated with TMDs, and the accuracy of each model was compared: logistic regression, decision trees, naïve Bayes, random forest, support vector machines, and an artificial neural network. The following hyper parameters were used for these methods: GINI was considered as the impurity measure of the decision tree, 1000 was regarded as the number of decision trees in the random forest, radial basis functioning was used as the kernel of the support vector machine, and 10-10 was the sizes utilized for two hidden layers and quasi-Newton (lbfgs) as weight optimization in the artificial neural network (https:// scikit-learn. org/ stable/ index. html). The data on 4744 participants were divided into training and validation sets at a 75:25 ratio. The models were trained based on the training set with data from 3558 participants and then validated using the validation set using data from 1186 participants. This validation set was not involved in the training (or learning) of machine learning approaches but was designed to only validate (or evaluate) their performance. Accuracy-defined as the rate of correct predictions from the data from 1186 participants-was used as a criterion for validating the trained models. Variable importance-an accuracy (or mean-impurity) gap between a complete model and a model excluding a certain variable 27 -was analyzed from the random forest model to test the study hypothesis, which is to assess the impact of each variable in predicting the presence of TMDs. For example, let us assume that the variable importance of "household income" is 0.10; the accuracy of the random forest will decrease by 0.10 if household income is excluded from the model. In other words, the variable importance of a certain variable measures the degree of its contribution to the performance of the model. From the independent variables included, the top 20 variables (in order of importance) were considered as risk factors for TMDs. Python 3.52 (Centrum voor Wiskunde en Informatica, Amsterdam, Netherlands) was used for statistical analysis in September 2020. For the logistic regression, odds ratios were calculated, and a P-value of < 0.05 was considered as being statistically significant.

Results
Descriptive statistics for participants' categorical and continuous variables are shown in Appendix Table 1 and  Table 2, respectively. Among the 4744 participants (2479 males and 2265 females, with a median age of 45 years), 101 (2.1%), and 68 (1.4%) had a r-TMD and a d-TMD, respectively. Forty-three individuals (0.9%) were included as both r-TMD and d-TMD. Of all respondents, 42% reported living in a metropolitan area, while 28.4% lived in a rural area (Appendix Table 1). Regarding the level of education, 66.8% had at least a high school education (Appendix Table 1). The median monthly household income, body mass index (BMI), and number of hours of sleep were $2500, 23.47, and 7 h, respectively ( Table 2). Table 3 shows the performance of the machine learning models that were tested. The greatest mean accuracy was observed via methods employing logistic regression, random forest, support vector machines, and artificial neural networks for both r-TMD and d-TMD; the greatest area under the receiver-operating-characteristic curve (AUC) was observed via the use of an artificial neural network and logistic regression for r-TMD and d-TMD, respectively ( Table 3).
The proportions of r-TMD and d-TMD are minimal; that is, 2.1% and 1.4%, respectively. This caused a classimbalance problem. The machine learning approach was trained to classify all observations as r-TMD "No" or d-TMD "No". This led to a high degree of accuracy but a low level of AUC. A possible solution is the use of undersampling, as reported in Table 3: 404, 202, and 909 observations were randomly sampled without replacement from 4643 observations with r-TMD "No" (the proportions of observations with r-TMD "Yes" became 20%, 30%, and 10%, respectively). These approaches are referred to as under-sampling 101:404, 101:202, and 101:909, respectively. Although the accuracy of logistic regression and the random forest methodology decreased to 0.86, their corresponding AUC values increased to 0.75 and 0.74 (respectively) in the case of under-sampling 101:404 (r-TMD). The AUC values did not improve in the other cases of under-sampling (see Table 3).
Based on variable importance determined from the random forest, the top 20 determinants of r-TMD, d-TMD, and r-TMD (under-sampling 101:404) are shown in Table 4  The odds ratio, as a result of the logistic regression, is shown in Appendix Table 2. A statistically significant association was observed for the following variables: age, marital status, health insurance type, working condition (respect), stress, suicidal ideation, and comorbidities of lumbago and depression.

Discussion
To our knowledge, this study is the first attempt made in applying artificial intelligence methodology in identifying the etiologic factors of TMDs as based upon large-scale, nationwide survey data of 4744 patients. A predictive model was developed using 37 independent variables regarding demographic factors, socioeconomic status, www.nature.com/scientificreports/ stress, working conditions, biological factors, and comorbidities, and which can be used as a decision-support system in the diagnosis of TMDs. Additionally, we analyzed the association between each etiologic factor and the presence of TMDs using the random forest variable importance measures. The random forest is a group of many decision trees with a majority vote concerning the dependent variable. For example, in a random forest with 1000 decision trees used in this study, 1000 training sets were sampled with replacements, 1000 decision trees were trained with the 1000 training sets, and the 1000 decision trees took a majority vote on the dependent variable. This explains the high reliability and immense popularity of the random forest method 28 . Moreover, the method is free from the assumption that all the other variables remained constant, which is the case for statistical models used in most previous studies on the subject. According to the literature review and the survey results, TMDs have significant associations with various psychological factors (such as anxiety, depression, and stress 29-31 ), working conditions (including employment, occupation, working schedule, and working hours 32 ), a high socioeconomic status (such as a high household income and a high degree of education 33 ), and other chronic diseases (such as osteoarthritis, sinusitis, allergic rhinitis, mental depression, and thyroid disorders 25 ). The findings of the literature are consistent with those identified via this study of the associations between TMDs and sleep, stress, employment/occupation, household income, and education. These variables were among the top 20 determinants of r-TMD as identified through our study. It needs to be noted, however, that the findings obtained via logistic regression are based on an unrealistic assumption of ceteris paribus ("the all the other variables remain constant"). For this reason, the results of logistic regression need to be considered as just supplementary information to the variable importance in terms of the random forest method.
Our results indicated that BMI is the most important determinant for the presence of TMDs; this was the case in both r-TMD and d-TMD. Self-perceived obesity was also ranked as fifth most important variable in both  Table 4. Random forest variable importance. *404 observations were randomly sampled without replacement from 4643 observations with TMD Self-Reported "No" (The proportion of observations with TMD Self-Reported "Yes" became 20%).  www.nature.com/scientificreports/ measures of TMDs. Obesity is a leading cause of disability and is associated with increased overall mortality 34 . It is recognized that adipose tissue plays a role in regulating inflammation in addition to storing energy 35 . Generally, increased adiposity is associated with the increased production of proinflammatory molecules, resulting in a proinflammatory state. Obesity has been designated as a risk factor for chronic musculoskeletal pain [36][37][38] . It is also strongly associated with osteoarthritis 39 and rheumatoid arthritis 40,41 . A recent study has confirmed the critical role played by adipokines (cytokines secreted by adipose tissues) in the pathophysiologic features of osteoarthritis, concluding that mechanical overload cannot completely explain the aggravation of knee osteoarthritis 39 . This, in turn, indicates the possible impact had by adipokines. We speculate that this systemic effect of increased inflammatory cytokines from the increased adiposity may be associated with the pathogenesis of TMDs. Another explanation could be that elevated BMI is associated with a low socioeconomic status 42 , which is another risk factor for TMDs. Similarly, it has been reported that arthritis and joint symptoms are highly prevalent among those with poor general health, a high BMI, and a low socioeconomic status [43][44][45] . In contrast, Busija et al. reported that the association between BMI and arthritis is strong, but relatively independent of one's age and socioeconomic status 46 .
Unlike arthritis, there are limited studies on the association between obesity and TMDs. Furthermore, for those studies that do exist, their results remain contentious. For example, a significant relationship between TMDs and obesity was observed via the use of univariate analysis. However, the association was not significant when the effects of sex, the presence of a migraine, nonspecific somatic symptoms, and obstructive sleep apnea syndrome were controlled for 47 . According to the research presented via the OPPERA studies-a large prospective and case-control investigation-BMI was not associated with TMDs 30 . Among adolescents, it has been concluded that obesity is not a risk factor for TMDs 48 .
This paper's efforts improve upon previous research concerning the importance of working conditions in managing TMDs. A recent study used the working schedule (shift vs. daytime) and working hours (< 40 vs 40-48, 49-60, > 60) as two aspects of working conditions concerning the prediction of TMDs 32 . This research has introduced five dimensions of working conditions (hygiene, risk, workload, control, and respect), with their importance rankings being higher than 20: control (7th), workload (9th), risk (10th), hygiene (11th) and respect (14th), indicating the most important determinants of r-TMD. These findings suggest that preventive measures concerning these working conditions should be considered for central health policies.
This study has some limitations. First, we made use of survey data; therefore, the presence of TMDs is based on the self-reporting of participants. However, both outcomes of r-TMD and d-TMD indicated similar risk factors. Further studies using diagnostic criteria such as DC/TMD 5 , which include detailed questionnaires on the signs and symptoms of TMD and have additional subclassifications of pain related TMDs and intra-articular disorders, may be warranted; the artificial intelligence model could be fine-tuned using such additional data. Another limitation is that this research applied a cross-sectional design; hence, we were only able to observe the associations between risk factors and the presence of TMDs, and their causal effects could not be identified. Moreover, the prevalence of TMDs was relatively low in our sample compared with its worldwide prevalence. The prevalence of TMDs varies greatly depending on the diagnostic criteria and the target population 6 . Also, Asians have a lower prevalence of TMDs than whites and African Americans 20 . Due to the low prevalence of participants with TMDs, there was a class imbalance; therefore, down-sampling was performed to avoid overfitting. This warrants further study with a larger number of participants with TMDs. Furthermore, this research did not consider potential relationships or mediating effects among the independent variables. A subgroup analysis across age and sex would offer more insights into the major determinants of TMDs.
This study identified the etiologic factors that may be associated with the disease; efforts to eliminate the identified factors may help improve the prognosis. A recent study demonstrated the automated detection of temporomandibular joint arthritis from cone-beam computed tomography images using deep learning techniques 49 . When the machine learning methods from our study are combined with the image analysis algorithms, a personalized real-time diagnosis based on imaging data and demographic and biological records data may be possible. Possible risk factors may be identified for the purpose of determining a prognosis. This line of research is expected to break new ground for cutting-edge precision medicine concerning the diagnosis, prognosis, and treatment of TMDs.

Conclusions
Artificial intelligence provides a decision-support system to predict TMDs and to analyze their determinants. Interventions regarding stress, socioeconomic status, and working conditions are needed for effective management of TMDs.

Data availability
The data are available from the Korea Disease Control and Prevention Agency database from the following webpage: https:// knhan es. kdca. go. kr/ knhan es/ sub03/ sub03_ 02_ 05. do. The data are available to anyone with the appropriate qualifications. www.nature.com/scientificreports/