Applying Deep Learning to Understand Predictors of Tooth Mobility among Urban Latinos

Abstract

We applied deep learning algorithms to build correlate models that predict tooth mobility in a convenience sample of urban Latinos. Our application of deep learning identified age, general health, soda consumption, flossing, financial stress, and years living in the US as the strongest correlates of self-reported tooth mobility among 78 variables entered. The application of deep learning was useful for gaining insights into the most important modifiable and non-modifiable factors predicting tooth mobility, and maybe useful for guiding targeted interventions in urban Latinos.

1. Introduction

Inflammation of the gums (also called ‘gingivitis’) causes oral health problems, with consequences of the supporting bone structure loss and edentulism. Partial and complete edentulism is also linked to increased risk of heart disease, brain health, and whole body infection.¹ Among ethnic minority groups in the US, Latinos have been reported to have the poorest oral health status² Besides non-modifiable factors (e.g., aging), there are multiple potentially innovative modifiable behavioral factors that may increase risk for tooth mobility such as stress or soda intake. Yet, the most important risk factors for gum inflammation and tooth loss among Latinos living in U.S. urban areas are poorly understood. As historical evidence reports “Modern human skeletons (e.g. teeth) have shifted towards lighter –more fragile when we adopted agriculture (behaviors) 12,000 years ago”, it hints to us that looking into aging and behavioral determinants together may have the potential to better understand the determinants of teeth loss. In this study, we applied data mining techniques to explore the extent to which aging, behavioral, and psychological risk factors influence urban Latinos’ self-reported gum health as groundwork for future targeted intervention aimed at preventing tooth mobility.

Deep learning³ has recently demonstrated the usefulness of big data predictive modeling in the area of image process or speech recognition. Yet deep learning methods have infrequently been applied in medicine compared to their use in industry perhaps in part due to their lack of transparency, leading to challenges in interpreting results.⁴ Yet, we propose that health researchers and professionals can benefit from using data mining to more efficiently identify important variables predicting adverse health outcomes, such as tooth mobility, compared with traditional statistical approaches.⁴

2. Methods

We applied a data mining process to a community-based dataset to build a model for predicting self-reported gum health after obtaining institutional review board approval. We used Weka 3.8 and Deep Neural Network-BigML software to conduct data mining analyses. We extracted 925 demographic and physio-psycho social variables from survey of 4,623 Latinos in the local REDCap database. First, 78 of total 925 variables were selected by dental and healthy behavior domain experts as potentially relevant to gum health based on the literature.⁵ During feature selection, we applied deep learning in BigML and feature subset selection algorithm in Weka, which evaluates the worth of a subset of attributes by considering the individual predictive ability of each feature along with the degree of redundancy between them. The outcome variable was the dichotomized self-reported of non-traumatic tooth mobility (‘experience of any teeth becoming loose on their own, without an injury’ versus ‘no experience of any teeth becoming loose on their own, without an injury’). BigML and Weka with default configuration were chosen to avoid algorithm dependency because selected features can vary by different tools. We applied 1) correlation based, 2) information gain based (entropy), 3) learner based techniques to create different views of the dataset and selected nine final variables based on the criteria of clinical meaningfulness based on an understanding of the literature on reasons for tooth mobility.⁵ Next, we organized the variables into three conceptual categories: demographic, psychological, and behavioral factors.

Next, we iteratively applied nine data mining algorithms including deep learning algorithms such as MultilayerPerceptron (Weka) and Deep Neural Networks (BigML) to build the prediction models for self-reported tooth mobility among urban Hispanic adults. For cross-validation (10-fold), the dataset was randomly divided into training and evaluating datasets for the model validation before applying the algorithm. We chose the final models based on the model predicative accuracy, interpretability, applicability and clinical meaningfulness, and the area under the receiver operating characteristic curve (AUC). Lastly, dental and behavioral science experts interpreted the models according to clinical meaningfulness.

3. Results

Study participants (n=4,623), age 18–100 (mean 49.3 ± 16.8) years, were predominantly female (n=3,416, 74.0%), were immigrants (n=3,894, 84.2%), spoke Spanish (n=3,470, n=75.1%), were with limited health literacy (n=3437, 75.3%), and were Medicare/Medicaid beneficiaries (n=3,521, 77.5%). One out of ten participants (n=508, 11.0%) reported that they had experienced tooth mobility. Descriptive statistics for the study variables are summarized in Table 1. Among the factors selected by multiple algorithms, age, acculturation, general health status, soda intake, financial stress, depressive symptom, and flossing behavior were finally selected by dental and healthy behavior domain experts based on systematic review and meta-analysis literature (Figure 1). While physiological factors such as diastolic blood pressure (importance 6.3%) and Body Mass Index (importance 3.9%) also selected by deep learning algorithms, they were excluded in the final prediction model due to their relative lack of interpretability and applicability.

Table 1.

Descriptive Statistics for Study Variables (n=4,623)

Variables*	Mean (SD), N (%)
Psychological factor
Having Financial Stressa [Yes]	802 (17.4%)
Depressive symptomsb [0–27, 27:worst]	1.8 (SD 4.1)
Behavioral factor
Flossing [0–51 times last week]	3.5 (SD 3.3)
Soda intake [0–84 times/week]	2.8 (SD 5.1)
General health status
Excellent	1,069 (23.1%)
Very good	1,126 (24.4%)
Good	1,228 (26.6%)
Fair	1,007 (21.8%)
Poor	106 (2.3%)

^aPerceived Stress Scale (PSS) (Cohen, Kamarck, & Mermelstein, 1983)

^bModified Patient Heath Questionnaire-9 (modified PHQ-9 depression)

^*missing data ≤5% except for flossing behavior “don’t know” 27.6%, “refused” 2.3%

Figure 1.

Deep Neural Network Ranking of Factor Importance for Tooth Mobility

4. Discussion

Among the many risk factors for tooth loss, deep learning found that aging was a key determinant (variable importance: 44.2%, Figure 1) among Latinos living in the US. Consistent with a long history of archeological and epidemiological findings, aging is an consistent determinant of tooth mobility. This study found that Latinos under 55-year-old (figure 2) were likely to have no tooth mobility regardless of their general health conditions, whether reporting that they are sick or are healthy (n=2,724).

Figure 2.

A Prediction Model (J48) for Self-Reported Tooth mobility (model accuracy: 88.4%, Precision: 0.83, Recall: 0.88, F-measure: 0.84, ROC area: 0.72)

This study found soda intake behavior among age over 55 is a determinant of tooth mobility (figure 2). This finding is consistent with the literature, including the historical findings. Archeological evidence reports association between carbohydrate-heavy diet resulting from agriculture (behavior) and tooth decay.⁶

The novel finding reported here concerned financial stress. For those who were considerably younger, and who had relatively increased soda intake, the presence of financial stress increased the probably of tooth mobility (figure 2). Consider the paths among these predictors will be an important next step. Is it the patients who cope with financial stress by increasing soda intake who have the most tooth decay? If so, such a pattern would suggest coping enhancement and behavior substitution as treatments when financial stress occurs, rather than straight dietary advice. Targeting patients who are 45–50 may be more appropriate for such an intervention.

Deep learning has several practical and methodological limitations to researchers in academia or public/dental/clinical informatics. Deep learning requires substantial computing time and years of training for engineering. Although optimizing multiple parameters can be challenging, this study used the Deep Neural Network function in BigML which offered a user-friendly automatic optimization options (e.g., setting up the appropriate number of hidden layers) to help us discover the best parameterization during our network search. Most of all, the biggest limitation is the lack of transparency and interpretability of the prediction model. The combined use of traditional machine learning algorithms such as C4.5 and Deep Neural Network in BigML is suggested for similar studies as Deep Neural Network in BigML offers the automated parameters customized to a user’s dataset without engineering expertise.

5. Conclusion

Our deep learning algorithm revealed aging, general health, soda intake, flossing behavior and financial stress as the strongest risk predictors of self-reported tooth mobility among Latinos living in a US inner city. This new knowledge adds insights about dental symptom science and aging for future intervention development.