Development of a periodontitis risk assessment model for primary care providers in an interdisciplinary setting

Abstract

BACKGROUND:

Periodontitis (PD), a form of gum disease, is a major public health concern as it is globally prevalent and harms both individual quality of life and economic productivity. Global cost in lost productivity is estimated at US$54 billion annually. Moreover, current PD assessment applies only after the damage has already occurred.

OBJECTIVE:

This study proposes and tests a new PD risk assessment model applicable at point-of-care, using supervised machine learning methods.

METHODS:

We compare the performance of five algorithms using retrospective clinical data: Naïve Bayes (NB), Logistic Regression (LR), Support Vector Machine (SVM), Artificial Neural Network (ANN), and Decision Tree (DT).

RESULTS:

DT and ANN demonstrated higher accuracy in classifying the patients with high or low PD risk as compared to NB, LR and SVM. The resultant model with DT showed a sensitivity of 87.08% (95% CI 84.12% to 89.76%) and specificity of 93.5% (95% CI 91% to 95.49%).

CONCLUSIONS:

A predictive model with high sensitivity and specificity to stratify individuals into low and high PD risk tiers was developed. Validation in other populations will inform translational value of this approach and its potential applicability as clinical decision support tool.

Keywords

Data mining decision support systems clinical health information systems periodontitis electronic health records risk assessment tools

1. Introduction

Similar to other chronic diseases, periodontitis (PD) progression is insidious and chronic with subtle symptomology that becomes apparent after much damage has been done to the underlying bone [1, 2, 3]. As the severity of the PD increases, the underlying alveolar bone and associated periodontal ligaments are damaged rapidly [1]. The 2018 report from the Centers for Disease Control and Prevention (CDC) estimated that 42% of the U.S. population more than 30 years of age and 66% of the population more than 65 years of age have some form of PD [4]. Notably PD also represents a potentially modifi able risk factor if treated in early stages, and may reduce the onset and progression of some systemic diseases, including diabetes [5, 6, 7, 8, 9, 10, 11, 12]. A study reported that a periodontal intervention in individuals who were recently diagnosed with Type 2 diabetes (T2DM) reduced overall healthcare costs by $ 1,799 over two years [14]. Hence, prevention, early recognition, and intervention are projected to not only decrease the bone damage and tooth loss but also reduce exacerbation of some systemic diseases that are driven by inflammatory mechanisms that are shared across many common chronic diseases.

Due to the complex relationship between oral and systemic diseases, multiple factors need to be considered simultaneously to effectively evaluate oral or systemic health risk. The current evidence base suggests that modeling medical variables, such as blood glucose levels, duration of T2DM, and severity of T2DM, along with dental parameters, such as tooth loss, periodontal pocket depth, gingival bleeding, and radiographic bone loss, have utility in assessing PD risk in dental settings [15, 16, 17, 18, 19]. However, interactions among multiple candidate risk factors for PD converge to make risk assessment at point of care (POC) difficult for medical provider. Moreover, primary care providers (PCPs) currently lack training and availability of analytical tools in medical setting to inform medical decision-making and care delivery at POC [5]. Hence, effective approaches for mining, processing and evaluating PD risk based on historical clinical and demographic data captured during the patient visit in medical and dental records are needed to develop models that process these data and create algorithms that project relative risk for disease. Demonstration of generalizability of predictive accuracy would support potential for creation of clinical decision support tools for translation into clinical practice.

Machine learning (ML) approaches to PD risk evaluation have been previously applied to predict and diagnose PD and PD risk in dental settings with electronic capture of dental variables [17, 19]. However, to date, there are no models that assess PD risk in an interdisciplinary setting. Few studies have applied ML to model PD risk. Two historic studies have attempted PD prediction using ML algorithms and were undertaken in dental settings. A study by Shankarapillai et al. assessed PD risk by comparing the Levenberg Marquadt algorithm (LMA) and the Scaled Conjugate Gradient (SCG) algorithm [19]. The study utilized data of 230 patients and 16 variables including history of diabetes and hypertension as medical variables and scored the PD risk on scale of 1 to 5. The study showed that LMA (a variant of backpropagation with an ANN) outperformed the SCG algorithm. A second study by Ozden et al. attempted to diagnose periodontal diseases using SVM, DT and ANN algorithms predicting classification across six periodontal conditions [17]. The study utilized 100 training and 50 test sets of 150 patients with 11 dental features. The total accuracy predicted by SVM and DT was 98%, however the study did not model medical parameters.

Nagarajan et al. further demonstrated capacity to apply bioinformatics classification algorithms including linear discriminant analysis (LDA), quadrant discriminant analysis (QDA), Naïve Bayes (NB) and Support Vector Machines (SVM) analysis to explore risk profiling of patients with gingivitis and PD based on salivary analysis of molecular marker expression across the two conditions [20]. In a separate study, these investigators further applied LDA, SVM and NB classification algorithms to predict individuals most at risk for disease progression among patients exhibiting a range of PD severity and treatment modalities based on levels of 27 biomarkers monitored in serum, saliva and gingival plaque. Thus, application of classifier algorithms demonstrates high utility in analysis of PD, a condition characterized by heterogeneous etiological origin and complex host response where prediction of outcome is singularly challenging [21]. The objective of this study was to create a PD risk assessment prototype through assessment of relative risk contributed by historical clinical, demographic, behavioral exposures, and other relevant variables captured in an integrated electronic medical and dental record (iEHR). Availability of such a tool would allow early detection, increased opportunities for improving oral health status and reduce incidence of underlying infectious and inflammatory processes contributing to other systemic diseases. A further goal of the study was to identify and optimize the most informative subset of clinically available attributes for PD risk prediction through the application of feature selection methods. This study is among the first studies to create a PD risk assessment model using the machine learning approach and combining assessment of both relevant clinical data and dental data captured in an iEHR, with applicability in a medical, as well as dental, setting.

2. Methods

2.1 Data retrieval

The Institutional Review Board of the healthcare organization was reviewed and approved the study. To achieve the objective of the study, retrospective data were extracted from the large healthcare organization’s enterprise data warehouse (EDW) for patients who visited dental and medical clinics between 2010 and 2016. Candidate factors previously associated with PD are catalogued in Table 1 and were specified for collection and analysis.

Table 1
List of all data features included in the prediction model development

Feature number	Feature	Categories
1	Age	21 to 89 years (numerical)
2	Gender	Male/Female
3	Medicaid status	Yes/No
4	Medicare status	Yes/No
5	Body mass index (BMI)	$<$ 18.5 $=$ Underweight 18.5–24.99 $=$ Normal 25.0–29.99 $=$ Overweight 30.0 and above $=$ Obese
6	Height	In centimeters (numerical)
7	Weight	In pounds (numerical)
8	Blood pressure (BP)-systolic	$<$ 120 mm of Hg 120–139 mm of Hg 140–159 mm of Hg $>$ 160 mm of Hg
9	BP-diastolic	$<$ 80 mm of Hg 80–89 mm of Hg 90–99 mm of Hg $>$ 100 mm of Hg
10	Tobacco use status	Current Former Never
11	High density lipids (HDL)	Less than 40 mg/dl $=$ Low 40 mg/dl to 59 mg/dl $=$ Healthy 60 mg/dl and above $=$ High
12	Low density lipids (LDL)	Less than 100 mg/dl $=$ Optimal 100 mg/dl to 129 mg/dl $=$ Near optimal 130 mg/dl to 159 mg/dl $=$ Borderline 160 mg/dl $<$ $=$ High
13	Triglycerides	Less than 150 mg/dl $=$ Healthy 150 mg/dl to 199 mg/dl $=$ Borderline 200 mg/dl to 499 mg/dl $=$ High 500 mg/dl and above $=$ Very High
14	Total cholesterol	Desirable Borderline High
15	Random blood glucose	Numeric value
16	Diabetes category	Type II diabetes Prediabetes No diabetes
17	Duration of diabetes	In years
18	Frequency of tooth brushing	Frequency of tooth brushing in a day
19	Frequency of tooth flossing	Frequency of tooth flossing in a day
20	Oral hygiene status	Poor/Fair/Good/Excellent
21	Presence of dental calculus	Yes/No
22	Total number of missing teeth	Numeric
23–78	Periodontal pocket depth (PPD)	Numeric for all teeth excluding third molars

2.2 Data preparation

A cohort of 11,048 (4,766 positive cases and 6,282 controls) patients with no missing data was randomly selected with enrollment limited to Non-Hispanic/Latino ethnicity and White/Caucasian race since other races and ethnicities were underrepresented. The goal of predicting PD was treated as a classification problem, with patients who have moderate or severe PD being classified as ‘high risk PD candidates’ (positive cases) and mild or no PD as ‘low risk PD candidates’ (controls). We used Centers for Disease Control and Prevention (CDC)-American Academy of Periodontology (AAP) definition to define the severity of PD [6]. Frequency distribution analysis was first performed to summarize the data in histogram formats illustrating relative distributions. For optimizing the accuracy of the model, outliers were removed from the dataset. Categorical discretization was done for some of the variables, including high density lipids (HDL), low density lipids (LDL), total cholesterol and, triglycerides, according to Adult Treatment Panel III classification [22]. Performance estimation was conducted using a stratified 10-fold cross validation approach. Table 1 summarizes the list of all data features include in the prediction model.

2.3 Machine learning algorithms

To test the generalizability of supervised ML in classifying patients for PD risk, the study focused on five well-known methods: Naïve Bayes (NB), Logistic Regression (LR), Artificial Neural Network (ANN), Support Vector Machine (SVM) and Decision Tree (DT). The study utilized the implementation of these ML algorithms available in the Waikato Environment for Knowledge Analysis (WEKA) open-source tool [23].

To identify a representative subset of attributes, a univariate filter, i.e., information gain with the ranker method was employed [24]. To validate the predictive performance of the resultant models, a subset of 10% (1,104/11,048) of the total data set was kept aside.

2.4 Performance measures

To assess the prediction model performance of different algorithms, the study compared ML algorithms using the following performance measures:

(1)
The area under the ROC curve (AUC) as defined by Hand and Till for binary classification [25].

$\displaystyle\text{AUC}=\frac{\frac{\left[{S0-n0\left({n0+1}\right)}\right]}{2% }}{n_{0^{n_{1}}}}$ (1)

where $n_{0}$ and $n_{1}$ are the numbers of ‘positive cases’ and ‘controls’, respectively, and $S_{0}=\sum r_{i}$ , where $r_{i}$ is the rank of the ith ‘positive cases’ in the ranked list.
(2)
Sensitivity, also termed recall, is the ratio of the number of correctly classified ‘positive cases’ instances to the total number of ‘controls’ instances.

$\displaystyle\text{Recall/Sensitivity (Se)}=\frac{\textit{TP}}{\textit{TP}+% \textit{FN}}$ (2)

where $\textit{TP}=$ true positive, $\textit{FN}=$ false negative.
(3)
Precision is the ratio of the number of correctly classified ‘positive cases’ instances to the total number of instances that are classified as ‘controls’.

$\displaystyle\text{Precision}=\frac{\textit{TP}}{\textit{TP}+\textit{FP}}$ (3)

where $\textit{FP}=$ false positive.
(4)
Specificity is the ratio of the number of correctly classified ‘positive cases’ instances to the total number of instances that are classified as ‘controls’.

$\displaystyle\text{Specificity}=\frac{\textit{TN}}{\textit{TN}+\textit{FP}}$ (4)

where $\textit{TN}=$ true negative.
(5)
Accuracy is the ratio of the number of correctly classified instances to the total number of instances.

$\displaystyle\text{Accuracy}=\frac{\textit{TN}+\textit{TP}}{\textit{TP}+% \textit{TN}+\textit{FP}+\textit{FN}}$ (5)
(6)
F-measure is the harmonic mean of precision and recall.

$\displaystyle\text{F-measure}=\frac{2{\ast}\text{Precision}\ast\text{Recall}}{% \text{Precision}+\text{Recall}}$ (6)
(7)
Matthew’s Correlation Coefficient (MCC) considers the accuracy and error rates of high PD risk and low PD risk and is calculated by the following equation:

$\displaystyle\text{MCC}=\frac{\textit{TP}\ast\textit{TN}-\textit{FP}\ast% \textit{FN}}{\sqrt{\left({\textit{TP}+\textit{FP}}\right)\left({\textit{TP}+% \textit{FP}}\right)\left({\textit{TN}+\textit{FP}}\right)\left({\textit{TN }+% \textit{FN}}\right)}}$ (7)

A paired t-test was performed to find if there were any statistically significant differences in the performance for the various algorithms.
3. Results

3.1 Patient characteristics for dataset excluding the evaluation dataset

The overall mean age of the patients was 47.36 years ( $+$ 16.60 years). Sixty five percent of the patients were female. Of all the patients, 7,315 were Medicaid patients and 3,578 were Medicare patients. Mean brushing frequency was 1.6 ( $+$ 0.60) per day. The distribution of the patients for presence or absence of diabetes was: 1887 ‘Type 2 Diabetes, 2834 ‘Pre-diabetes’, and 6327 ‘No diabetes’. A majority of patients (1369/1887; 72.5%) diagnosed with diabetes had a duration of diabetes $<$ 1 year. One hundred and thirty (3%) patients had documentation of poor oral hygiene while the rest were categorized as good, fair, and excellent. Of these patients, 2,994 (30%) were current smokers, 3,460 (34.8%) were former smokers, and 3,490 (35.2%) never smoked tobacco. The mean random blood glucose level for this cohort was 112 ( $+$ 46.36). More than half of the patients showed a normal blood pressure level. Approximately twice as many patients ( $n=$ 5,813) with BMI of more than 30 were classified as obese, whereas ( $n=$ 3,060) with BMI between 25 and 29.99 were classified as overweight. There was significant difference ( $p<$ 0.0001) between the control group and positive cases for most of the variables utilized in the study.

3.2 Feature selection

The application of information gain with the ranker method to the dataset of 9,944 patients generated a representative subset of 71 variables from 190 variables as shown in Fig. 1.

Figure 1.

The results of feature selection.

Machine learning: The results of the performance estimated through 10-fold cross validation are shown in Table 2. The ROC curves are shown in Fig. 2.

The results show that DT significantly outperformed the other algorithms in terms of sensitivity, specificity, AUC, F-measure, MCC, and precision ( $p<$ 0.001).

DT and ANN demonstrated higher accuracy in classifying the patients with high or low PD risk as compared to NB, LR, and SVM. The resultant model with the DT algorithm classified $n=$ 519 high PD risk [sensitivity: 87.08 % (95%CI 84.12% to 89.67%)], and $n=$ 475 with low PD risk [specificity: 93.50% (95%CI 91% to 95.49%)] among the test set of 1104 patients. By comparison, sensitivity and specificity of ANN were 84.60% [95%CI (83.12%–85.46%)] and 84.60% [95%CI (83.48%–85.82%)], respectively.

In this study, the specificity and precision for DT achieved 90.20% [95%CI (89.20–91.10)] and 90.20% (90.10–90.20)], respectively. The rates of precision in this study were slightly lower than those reported by Shankarapillai et al. [19], where the precision was 98% for DT in diagnosing PD. However, the rates of specificity in the current study were higher than those reported by Shankarapillai et al. (specificity $=$ 46%). Based on the scale suggested by Allaire et al. [26], the overall performance of DT was excellent, while performance of ANN, NB, LR, and SVM was ‘very good’, respectively.

Table 2

Results of classifiers after feature selection

ML algorithm	Accuracy (95%CI)	Precision (95%CI)	Recall (95%CI)	Specificity (95%CI)	F-measure (95%CI)	AUC (95% CI)	MCC (95%CI)
Naïve Bayes	0.801 (0.791–0.811)	0.808 (0.807–0.808)	0.804 (0.794–0.814)	0.794 (0.781–0.807)	0.800 (0.790–0.810)	0.856 (0.854–0.858)	0.609 (0.570–0.630)
Logistic regression	0.768 (0.778–0.795)	0.786 (0.785–0.786)	0.807 (0.799–0.804)	0.783 (0.772–0.794)	0.785 (0.77 0–0.790)	0.855 (0.854–0.856)	0.560 (0.530–0.590)
Support vector machine	0.790 (0.780–0.799)	0.790 (0.789–0.790)	0.812 (0.804–0.820)	0.809 (0.798–0.819)	0.787 (0.770–0.790)	0.808 (0.807–0.809)	0.566 (0.530–0.590)
Artificial neural network	0.841 (0.833–0.849)	0.841 (0.840–0.841)	0.841 (0.833–0.848)	0.843 (0.835–0.851)	0.841 (0.830–0.850)	0.895 (0.893–0.897)	0.674 (0.650–0.690)
Decision tree	0.901 (0.892–0.908)	0.902 (0.901–0.902)	0.902 (0.893–0.908)	0.902 (0.892–0.911)	0.902 (0.890–0.910)	0.901 (0.899–0.902)	0.799 (0.770–0.810)

Figure 2.

ROC curves for the algorithms.

4. Discussion

This study demonstrated successful creation of algorithms with high capacity to classify PD risk. Using univariate filtering, the study selected the subset of features with non-zero information gain and recognized that random blood glucose, dental calculus, number of missing teeth, lipid panels (including triglyceride levels and HDL), diastolic blood pressure, BMI, oral hygiene status determined by the dental provider, frequency of tooth brushing, diabetes status, tobacco use status, age, gender, and PPD displayed highest performance.

Overall, the results of performance assessment of algorithms created to classify relative risk for PD showed that DT demonstrated highest performance for disease risk classification than other ML algorithms. Notably, the F-measure, which represents a harmonic mean of precision and recall, has a greater importance in evaluation as it evaluates the relationships between high PD risk instances within the data and those given by the classifiers. Despite the association of high type 1 error with cross-validated t-tests, evidence suggests that cross-validated t-tests are powerful in determining whether a learning algorithm outperforms another on a particular learning task [28]. To determine the real difference between algorithms (seen in type II errors), this study statistically analyzed the results of the algorithms by conducting a 10-fold cross validated t-test on F-measure. The results suggest that for such type of data, ANN or DT can be used for modelling. DT offers the additional advantage of being interpretable, however, allowing people to review the decision rules that have been learned, which may be important in gaining the trust of clinical providers.

Selection of data variables led to the novel observation that measurement of probing depths at interproximal tooth surfaces significantly outperformed measures taken at the buccal and lingual surfaces. The study posits that the apparent superiority of interproximal measurement may be attributable to the interproximal bone, which is more coronal in position than the labial or lingual/palatal bone. A slight deepening of the pocket in the interproximal areas could more easily impact the bone. This observation adds to the current standard of care with respect to commonly applied indices used in clinical dentistry (e.g., Russell’s periodontal index, Loe and Silness gingival index) and also are consistent with the definition proposed by AAPD for periodontal diseases, where interproximal surfaces are assessed to determine the severity of PD [29, 30]. The progression of PD involves furcation areas of the multi-rooted teeth in maxillary (upper jaw) and mandibular (lower jaw). Tooth surfaces, such as mesiolingual, mesiobuccal, distobuccal and distolingual of maxillary and mandibular molars, were identified as significant determinants during feature selection. For calculating clinical attachment loss, location of cemento-enamel junction (CEJ) is necessary. Interestingly, the sites identified by feature selection are also used as reference lines for determining relative clinical attachment (RCAL) loss when it is difficult to locate CEJ [31]. Moreover, these sites were also consistent with outcomes of a study that investigated the deepest crevice points in the mouth to provide practitioners with minimum number of sites to probe [32]. Clinical implications of the model.

In the present study, random blood glucose (RBG) level was identified as a significant factor contributing to PD risk. This observation further reinforces association between PD and T2DM and would have considerable relevance in a clinical setting.

Notably, Medicare and Medicaid status were incorporated into the model to explore the relationship between insurance status and PD risk and were retained as significant variables. These findings suggest that there may be a possible correlation between Medicare/Medicaid status and PD risk especially in aging populations and attests to the value of machine learning to discover such relational patterns.

The study has some limitations. The study data was collected at a single healthcare system. Notably, the study population of this large healthcare organization’s service area extends across largely rural communities in northern, western, and central Wisconsin counties [33]. Residents of this area disproportionately exceed the State’s average population statistics for persons considered to be in lower socio-economic strata and also for individuals $>$ 65 years of age. Based on regional demographics, Medicare/Medicaid status and age were among variables test for retention in the model to explore the relationship between insurance status and PD risk. Interestingly, retention of this variable in the model suggest a possible correlation between Medicaid status and PD risk. This raises the potential for selection bias within the healthcare system. Lack of racial/ethnic diversity in our population necessitated delimiting the study dataset to White/Caucasian race and non-Hispanic/Latino ethnicity, therefore limiting the use of the algorithm in racial or ethnic minority populations without prior evaluation of these algorithms in a representative population cohort. Generalizability and extrapolation of the predictive model developed in this study to another health care center however, requires further testing and validation. Re-running the modelling effort using other data sets with other race and ethnicity would be needed. Some attributes, including waist circumference and clinical attachment loss were eliminated due to incomplete and missing data. The addition of these variables, genetic markers, or laboratory values for surrogate biological markers of systemic inflammation, such as high sensitivity C-reactive protein, may be necessary to further improve the PD risk prognostication in medical settings. Further, incomplete documentation resulting in missing structured data surrounding oral manifestations, such as bleeding gums on tooth brushing and swollen gums, might have adversely affected the classification results. A potentially useful approach could be to apply natural language processing techniques to unstructured resources, such as oral or dental complaints of patients or other text within clinical notes, to identify or test additional useful parameters to support the current PD risk model or capture data to supplement data incompletely captured in structured data fields.

Clinical relevance and future directions: potential translational value of the risk classification algorithms and their applicability to the future development of clinical decision support tools.

Practice patterns for dental and medical healthcare delivery, their respective individual reimbursement systems, and the current state of dental and medical academic practice continue to reflect the siloing of medical and dental healthcare delivery models [13, 33, 34]. The increasing scientific evidence supporting oral and systemic disease associations has cast primary care providers (PCPs) as proactive participants in establishing oral and systemic care for patients with chronic diseases, including PD [5]. Establishing interdisciplinary care for improving healthcare practice and expanding the access to preventive oral health care through PCPs has been proposed by the National Academy of Medicine and others [34]. Notably, evidence supports that patients visit their PCPs with higher periodicity compared to dental providers [35]. Based on the 2017 report of the National Center for Health Statistics, nearly 85% of the adult U.S population contacted a medical health care professional in the past year [36]. Proportionately it was also noted that about 64% of adult U.S. population had a dental visit in 2016 [37]. This paradigm shift in delivering integrated care has necessitated the incorporation of oral-health screenings at point-of-care in a medical setting [5]. Using a risk based approach, PCPs could assess the risk of developing a future disease based on evidence gathered from current and documented retrospective clinical data of patients [9, 16]. According to the Agency for Healthcare Research and Quality report, clinical decision support tools (CDST) can potentially lower costs, improve efficiency and reduce patient inconvenience [38]. The growing recognition of machine learning (ML) approaches using data in clinical settings is considered as a significant opportunity, not only to improve patient care but also to support quality of care initiatives [39, 40]. A better appreciation of the systemic effects and well-known periodontal risk factors along with behavioral factors has shifted the focus positing that collectively, the sum of risk contributed by a combination of individual factors provides better predictive power than with any single risk factor.

Findings from this study suggest that ML methods would be effective when applied to improving patient care through early detection of PD or to new preventive approaches to PD by assisting healthcare professionals to evaluate patients’ PD risk based on a combination of historical data and current status. As many factors might potentially affect individual variability in developing PD risk of a patient, this study considered a wide variety of predictive factors, including oral factors, such as dental calculus and number of teeth present, which are also evaluable in a medical setting. The study analyzed multiple widely available ML approaches to identify those with highest potential for translation into clinical care to assist healthcare providers in making effective and knowledge-driven decisions. Future steps include incorporation of such models into the iEHR and validating model performance in a clinical setting.

5. Conclusions

To the best of our knowledge, this study is the first to apply ML algorithms in the context of PD risk assessment in an interdisciplinary setting by comparing five predictive ML algorithms: NB, LR, SVM, ANN, and DT and evaluated their sensitivity and specificity for stratifying PD risk. Evaluation of performance of these algorithms in other populations is essential to demonstrate their generalizability and relevance and determine their potential translational value and utility as clinical decision support tools in the medical setting.

Footnotes

Acknowledgments

The authors would like to acknowledge Marshfield Clinic Research Institute and University of Wisconsin-Milwaukee. The authors would furthermore like to thank Dr. Ingrid Glurich for reviewing the initial version of the manuscript. Funding was provided by Marshfield Clinic Research Institute.

Conflict of interest

The authors declare no real or potential conflicts of interest.

References

Cekici

Kantarci

Hasturk

, et al. Inflammatory and immune pathways in the pathogenesis of periodontal disease. Periodontol 2000 2014; 64: 57-80.

Loesche

. Microbiology of Dental Decay and Periodontal Disease. University of Texas Medical Branch at Galvestonhttp//www.ncbi.nlm.nih.gov/pubmed/21413316 (1996, accessed 1 July 2018).

Khan

Kong

Meiller

, et al. Periodontal Diseases: Bug Induced, Host Promoted. PLOS Pathog 2015; 11: e1004952.

Eke

Thornton-Evans

Wei

, et al. Periodontitis in US Adults: National Health and Nutrition Examination Survey 2009-2014. J Am Dent Assoc 2018; 149: 576-588, e6.

Shimpi

Schroeder

Kilsdonk

, et al. Medical Providers’ Oral Health Knowledgeability, Attitudes, and Practice Behaviors: An Opportunity for Interprofessional Collaboration. J Evid Based Dent Pract 2016; 16: 19-29.

Eke

Page

Wei

, et al. Update of the Case Definitions for Population-Based Surveillance of Periodontitis. J Periodontol 2012; 83: 1449-1454.

Van Dyke

Sheilesh

. Risk factors for periodontitis. J Int Acad Periodontol 2005; 7: 3-7.

Glurich

Shimpi

Acharya

. Interdisciplinary Care Model: Chronic Kidney Disease and Oral Health. In: Acharya

Powell

Torres-Urquidy

, et al. Integration of Medical and Dental Care and Patient Data. Springer Nature Switzerland AG: Springer International Publishing, pp. 87-107.

Shimpi

Ashton

Sorenson

, et al. Interdisciplinary Care Model: Diabetes and Oral Health. In: Acharya

Powell

Torres-Urquidy

, et al. Integration of Medical and Dental Care and Patient Data. Cham: Springer Nature Switzerland, pp. 47-61.

10.

Glurich

Shimpi

Scannapieco

, et al. Interdisciplinary Care Model: Pneumonia and Oral Health. In: Acharya

Powell

Torres-Urquidy

, et al. Integration of Medical and Dental Care and Patient Data. Springer Nature Switzerland AG: Springer International Publishing, pp. 123-139.

11.

Shimpi

Pathak

Acharya

. Interdisciplinary Care Model: Metabolic Syndrome and Oral Health. In: Acharya

Powell

Torres-Urquidy

, et al. Integration of Medical and Dental Care and Patient Data. Cham: Springer Nature Switzerland, pp. 141-154.

12.

Shimpi

Dart

Umukoro

, et al. Interdisciplinary Care Model: Cardiovascular Diseases and Oral Health. In: Acharya

Powell

Torres-Urquidy

, et al. Integration of Medical and Dental Care and Patient Data. Cham: Springer Nature Switzerland, pp. 71-85.

13.

Nasseh

Vujicic

Glick

. The relationship between periodontal interventions and healthcare costs and utilization: Evidence from an Integrated Dental. Medical, and Pharmacy Commercial Claims Database. Health Econ. 2017; 26(4): 519-527.

14.

Eke

Dye

Wei

, et al. Prevalence of Periodontitis in Adults in the United States: 2009 and 2010. J Dent Res 2012; 91: 914-920.

15.

Goetzel

Staley

Ogden

Stange

Fox

Spangler

Tabrizi

, et al. A Framework for Patient-Centered Health Risk Assessments Providing Health Promotion and Disease Prevention Ser vices to Medicare Beneficiaries. Centers Dis Control Prev Atlanta, GA.

16.

Page

Martin

Loeb

. The Oral Health Information Suite (OHIS): its use in the management of periodontal disease. J Dent Educ 2005; 69: 509-20.

17.

Ozden

Ozgonenel

Ozden

, et al. Diagnosis of periodontal diseases using different classification algorithms: a preliminary study. Niger J Clin Pr 2015; 18: 416-421.

18.

Page

Krall

Martin

, et al. Validity and accuracy of a risk calculator in predicting periodontal disease. J Am Dent Assoc 2002; 133: 569-76.

19.

Shankarapillai

Mathur

Nair

, et al. Periodontitis Risk Assessment using two artificial Neural Networks-A Pilot Study. Int J Dent Clin 2010; 2: 36-40.

20.

Nagarajan

Miller

Dawson

, et al. Patient-Specific Variations in Biomarkers across Gingivitis and Periodontitis. PLoS One 2015; 10: e0136792.

21.

Nagarajan

Miller

Dawson

, et al. Biologic modelling of periodontal disease progression. J Clin Periodontol 2019; 46: 160-169.

22.

High

Cholesterol

. ATP III At-A-Glance: Quick Desk Reference. Hypertension 2009; 1-6.

23.

Witten

Frank

Hall

. Data Mining: Practical Machine Learning Tools and Techniques, Third Edition. 2011. Epub ahead of print 2011. DOI: 10.1002/1521-3773(20010316)40:6.

24.

Mitchell

, T. Machine learning. MIT Press, 1997, 414.

25.

Hand

, Measuring classifier performance: a coherent alternative to the area under the ROC curve. Mach Learn 2009; 77: 103-123.

26.

Fawcett

. An introduction to ROC analysis . 2006. Epub ahead of print 2006. DOI: 10.1016/j.patrec.2005.10.010.

27.

Drummond

Holte

Chawla

Sheng

Fang

, et al. Exploiting the cost (in) sensitivity of decision tree splitting criteria. Int Conf Mach Learn. 2003; 66(1): 239-246.

28.

Dubey

Zhou

Wang

, et al. Analysis of sampling techniques for imbalanced data: An n= 648 ADNI study. Neuroimage 2014; 87: 220-41.

29.

Russell

. A System of Classification and Scoring for Prevalence Surveys of Periodontal Disease. J Dent Res 1956; 35: 350-359.

30.

Beltrán-Aguilar

Eke

Thornton-Evans

, et al. Recording and surveillance systems for periodontal diseases. Periodontol 2000 2012; 60: 40-53.

31.

Newman

Takei

Klokkevold

, et al. Carranza’s Clinical Periodontology 11th Ed. 2012 . Epub ahead of print 2012. DOI: 10.1007/s13398-014-0173-7.2.

32.

Pritchard

Laws

. Gingival crevice depth. I. Predictability of probing deepest points. Aust Dent J 1984; 29: 404-410.

33.

Acharya

. Marshfield Clinic Health System: Integrated Care Case Study. J Calif Dent Assoc 2016; 44: 177-81.

34.

Atchison

Rozier

Weintraub

. Integration of Oral Health and Primary Care: Communication, Coordination and Referral. 2018. https://nam.edu/integration-of-oral-health-and-primary-care-communication-coordination-and-referral/. Accessed June 28, 2019.

35.

Schiller

Lucas

Ward

Peregoy

. Summary health statistics for U.S. Adults: National Health Interview Survey, 2012. Natl Cent Heal Stat. 2012; 10(252): 1-171.

36.

Centers for Disease Control and Prevention. Ambulatory Care Use and Physician office visits. 2017. https://www.cdc.gov/nchs/fastats/physician-visits.htm. Accessed June 28, 2019.

37.

Centers for Disease Control and Prevention. Oral and Dental Health. 2017. https://www.cdc.gov/nchs/fastats/dental.htm. Accessed June 28, 2019.

38.

U.S. Department of Health & Human Services. Agency for Healthcare Research and Quality. Topic: Clinical Decision Support (CDS). 2019. https://www.ahrq.gov/topics/clinical-decision-support-cds.html. Accessed June 28, 2019.

39.

Islam

Hasan

Wang

Germack

Noor-E-Alm

. A Systematic Review on Healthcare Analytics: Application and Theoretical Perspective of Data Mining. Healthcare (Basel). 2018; 6(2): 54.

40.

Jiang

Hui

Dong

, et al. Artificial intelligence in healthcare: past, present and future. Stroke Vasc Neurol. 2017; 2(4): 230-243.

Development of a periodontitis risk assessment model for primary care providers in an interdisciplinary setting

Abstract

BACKGROUND:

OBJECTIVE:

METHODS:

RESULTS:

CONCLUSIONS:

Keywords

1. Introduction

2. Methods

2.1 Data retrieval

Table 1 List of all data features included in the prediction model development

2.3 Machine learning algorithms

2.4 Performance measures

3.1 Patient characteristics for dataset excluding the evaluation dataset

3.2 Feature selection

5. Conclusions

Footnotes

Acknowledgments

Conflict of interest

References

Table 1
List of all data features included in the prediction model development