Periodontal health classification system capable of being used for self-screening, classification method and application

Abstract

The invention discloses a periodontal health classification system capable of being used for self-screening, the system comprises a classification model including logistic regression and random forest, and based on important features capable of distinguishing periodontal health conditions, the important features are combined and weighted to obtain periodontal health condition classification. The invention also discloses a classification method realized by using the classification system, and the classification method comprises the steps: collecting non-clinical parameter data related to periodontal health, and carrying out the preprocessing of the data; constructing a logistic regression model and a random forest model; and evaluating and classifying periodontal health conditions by using the constructed logistic regression model and random forest model.

Description

Technical Field

[0001] The present invention belongs to the technical field of data analysis and processing, and relates to a periodontal health classification system, classification method and application that can be used for self-screening. Background Art

[0002] Nonclinical methods for detecting and differentially diagnosing periodontal health status hold great promise for better prevention and management of periodontal disease (gingivitis and periodontitis), which remains a significant health burden and is the most common non-communicable disease in humans (Tonetti et al., 2017). Self-report measures proposed by the Centers for Disease Control and Prevention (CDC) and the American Academy of Periodontology (AAP) have shown promise in different populations (Carra et al., 2018; Deng et al., 2021b; Eke and Genco, 2007). Recent efforts have also focused on the use of biomarkers and their combinations (Arias-Bujanda et al., 2020; Deng et al., 2021a; Deng et al., 2022; Grant et al., 2022). Despite persistent limitations, recent advances have been substantial (Gürsoy and Kantarci, 2022). Combining information from multiple variables shows the potential to improve accuracy. For example, incorporating demographics into biomarker data has improved the performance of multivariate diagnostic models for periodontal health and disease. (Deng et al., 2021a, 2021b; Grant et al., 2022).

[0003] Previous work by this group has identified the different utility of specific questions in the Chinese version of the AAP / CDC questionnaire, gingival bleeding during toothbrushing, and levels of active matrix metalloproteinase-8 (aMMP-8) in distinguishing periodontal health, gingivitis, and periodontitis. (Deng et al., 2021a, 2021b, 2021c; Deng et al., 2022). The Chinese version of the CDC / AAP questionnaire showed good accuracy in screening for severe periodontitis (Deng et al., 2021b). However, the questionnaire was less helpful in detecting gingivitis and incipient periodontitis. Recently conducted studies have shown that gingival bleeding during toothbrushing (GBoB) and hemoglobin (Hb) concentration analysis in oral rinses can serve as sentinel markers of gingival inflammation and have great potential for distinguishing periodontal health and disease, especially in detecting gingival inflammation. (Deng et al., 2021c; Tonetti et al., 2020). Although GBoB plays a valuable role in periodontal diagnosis, its accuracy for periodontitis is low to moderate, probably because this feature is common to both gingivitis and periodontitis. Point-of-care diagnostics based on salivary biomarkers can immediately indicate possible disease states, allowing periodontal health monitoring outside the clinic (Aro et al., 2017). Matrix metalloproteinases (MMPs) are a family of proteases that regulate cell-matrix composition, and MMP-8 is the main type of collagenase involved in collagen degradation in periodontal disease (Birkedal-Hansen et al., 1993; Sorsa et al., 2016). Consumer versions of point-of-care tests (POCT) for activated matrix metalloproteinase-8 (aMMP-8) have shown potential to detect or rule out periodontitis (Deng et al., 2021a; Deng et al., 2022). However, the accuracy of aMMP-8 alone has shown limitations.

[0004] A fundamental limitation of this approach is the use of traditional diagnostic accuracy analyses developed for binary conditions. Applying these analyses to multiple diagnostic situations, such as differentiating between periodontal health, gingivitis, and periodontitis, requires an artificial dichotomy of the diagnostic problem: healthy and gingivitis versus periodontitis, or healthy versus gingivitis versus periodontitis. Identifying the spectrum of periodontal case diagnoses may be valuable for optimizing case management. This problem is not unique to periodontal health and is typical of multistage diseases, for which specific analytical approaches have been proposed (Lorena et al., 2009). Machine learning is a subset of artificial intelligence (AI) and computer science that aids in disease diagnosis and differentiation by constructing models in the form of input-output variable analyses, allowing for the selection of optimal feature subsets and automated classification (Sajda, 2006). Machine learning-based approaches have been used to identify diagnostic patterns and improve accuracy (Maity and Das, 2017). Systematic reviews have identified potential benefits for diabetes and its complications and cardiovascular disease (Baashar et al., 2022; Shin et al., 2022). Preliminary methodological reports show the potential of applying artificial intelligence to population screening for periodontal health (Bashir et al., 2022). Summary of the invention

[0005] In order to solve the deficiencies in the prior art, an object of the present invention is to provide a periodontal health classification system, classification method and application that can be used for self-screening.

[0006] The present invention provides a periodontal health classification system that can be used for self-screening, the system includes a classification model, based on important features that can distinguish periodontal health status, the important features are combined and weighted to obtain a periodontal health status classification; the periodontal health status includes periodontal health, gingivitis, and periodontitis; the periodontitis is divided into stage I periodontitis, stage II periodontitis, stage III periodontitis, and stage IV periodontitis. The system in the present invention provides a computer application that enables a user to obtain a result after inputting the value of the important feature defined in the subject, subject group, or population.

[0007] The system of the present invention includes a validated questionnaire, self-reported signs and symptoms, self-testing of biomarkers reflecting the health-disease spectrum and its outcomes, a digital platform (e.g., computer, tablet, smartphone) running a specific computer algorithm, a computer algorithm capable of distinguishing periodontal health conditions based on input data, a computer application allowing subjects to interface with the software, and a learning module that enables the algorithm to learn during use.

[0008] The classification model includes one or more logistic regression models and / or one or more random forest models; the logistic regression model is used to preliminarily evaluate periodontal health status, and evaluates "periodontal health", "gingivitis", "periodontitis", "stage I / II periodontitis" and "stage III / IV periodontitis" respectively, and screens predictors to determine predictors that are important for the prediction results; the random forest model is used to perform multiple classifications on subjects, subject groups and / or populations presenting different periodontal health disease spectra; the multiple classifications include three classifications of "periodontal health", "gingivitis" and "periodontitis", six classifications of "periodontal health", "gingivitis", "stage I periodontitis", "stage II periodontitis", "stage III periodontitis" and "stage IV periodontitis", and three classifications of "periodontal health", "gingivitis and stage I periodontitis" and "stage II-IV periodontitis".

[0009] Specifically, the logistic regression model is used to evaluate periodontal health status, including "periodontal health", "gingivitis", "periodontitis", "stage I / II periodontitis", "stage III / IV periodontitis", and screen to determine important predictive factors; the predictive factors are obtained through self-report measurements collected through questionnaires, disease symptoms and signs, and self-test results, including tooth mobility, tooth displacement (migration or drift), tooth elongation, tooth and gum appearance, presence of gingival bleeding, gingival brushing bleeding test, and the presence and / or quantitative assessment of health / disease biomarkers in body fluids or exhaled gas; the body fluids or gases include saliva, mouthwash, blood, serum, plasma and / or exhaled gas;

[0010] The random forest model is used to perform multi-classification of subjects, subject groups and / or populations presenting different periodontal health disease spectra; the periodontal health disease spectrum includes healthy periodontium without inflammation and periodontal tissue destruction, gingivitis with gingival inflammation but without periodontal tissue destruction characteristics and / or health / disease biomarkers, and periodontitis with periodontal tissue destruction characteristics and / or health / disease biomarkers.

[0011] During the specific implementation process, the data collected included responses to validated questions, such as the presence of tooth hypermobility (loose teeth), tooth drift and / or elongation, tooth / gum appearance, previous treatment history, presence of gingival bleeding during brushing, detection of specific biomarkers for distinguishing periodontal health, gingivitis and different stages of periodontitis, and subject demographics, such as age, gender, amount of smoking, and presence of comorbidities.

[0012] The data used in the logistic regression model include data obtained from questionnaires, active matrix metalloproteinase-8 lateral flow immunochromatographic rapid test aMMP-8POCT and gum brushing bleeding GBoB test; as well as demographic data, lifestyle factors, and personal medical history; the demographic data include age and gender; the lifestyle factors include smoking status; the personal medical history includes whether or not suffering from cardiovascular disease and / or diabetes; the gum brushing bleeding test includes self-reported gum bleeding during brushing, salivary hemoglobin concentration and content.

[0013] The logistic regression model was incorporated with predictors using backward stepwise selection, with the parameter associated with each predictor being subjected to a Wald test and removed if not significant at the 95% confidence level, the model was rerun and the process repeated until no more predictors were removed.

[0014] According to the predicted probability of the logistic regression model, the receiving operating characteristic (ROC) curve, the area under the ROC curve (AUROC), sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV) were estimated; the predicted probability threshold (thredshold) was selected as the point that maximizes the sum of the sensitivity and specificity of the entire ROC curve, which is the upper left point on the curve; when the predicted probability is higher than the predicted probability threshold in each logistic regression model, it belongs to the corresponding classification in each logistic regression model.

[0015] When the data used in the logistic regression model also include demographic and lifestyle behavior factors, the evaluation formula for each logistic regression model is as follows:

[0016] Fitted linear predictor of estimated logarithm of periodontal health = -1.750*Q2+0.844*Q3a+0.951*Q7-0.027*Hb(total)-1.471*aMMP-8POCT-0.050*age+2.112;

[0017] Fitted linear predictor of estimated log gingivitis = -0.992*Q3a+1.396*self-reported GBoB-0.160*age-14.719;

[0018] The fitted linear predictor of the estimated logarithm of periodontitis = 1.407*Q2+1.845*aMMP-8POCT+0.102*age-4.203;

[0019] The fitted linear predictor of the estimated logarithm of stage I / II periodontitis = -1.540*Q4+1.022*aMMP-8POCT+0.521*sex-0.963;

[0020] The fitted linear predictor of the logarithm of stage III / IV periodontitis was estimated = 1.729*Q2+1.756*Q6+0.115*age+2.453*smoking-8.024.

[0021] In the generation of the random forest model, each forest uses multiple estimators and sets a maximum depth value, and node segmentation is determined using the Gini coefficient; the random forest model is validated by running multiple independent trainings on the entire data set and comparing the results to ensure the internal consistency of the model.

[0022] The random forest model was fitted using the Python Scikit-learn package, and the Yellowbrick package was used to calculate the micro- and macro-averaged multi-class AUROC values ​​and the AUROC for each class; the fitting results reported the mean and deviation of the model performance, including sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and accuracy;

[0023] Among them, sensitivity and specificity were defined as low (<60%), moderate (60% to 79%), or high (≥80%);

[0024] The accuracy results derived from the AUROC values ​​were interpreted as low (0.50 to 0.70), medium (0.71 to 0.90), and high (>0.90);

[0025] Accuracy was calculated according to the following definition: Accuracy = Sensitivity * Prevalence + Specificity * (1-Prevalence).

[0026] The classification system also includes a data input module and a data output module;

[0027] The data input module is used to input the acquired non-clinical parameter data into the logistic regression model and / or random forest model;

[0028] The data output module is used to output the classification results generated by the logistic regression model and / or the random forest model.

[0029] The non-clinical parameter data (nine predictors) used for the random forest model included: "Q1 self-reported presence or absence of periodontitis" (question 1), "Q2 self-rating of gingival / dental health" (question 2), "Q3a dental cleaning history" (question 3a), reported symptoms of "Q4 loose teeth" (question 4), aMMP-8POCT (when the predicted probability was higher than the predicted probability cutoff value in each logistic regression model, it belonged to the corresponding category in each logistic regression model), self-reported GBoB (gingival bleeding during brushing), hemoglobin Hb concentration (in saliva, toothpaste mixture or toothbrush pulp) and age.

[0030] The present invention also proposes a method for constructing the above classification system, which comprises the following steps:

[0031] Step 1: Collect clinical data on periodontal health status classification as training and validation sets for the model;

[0032] Step 2: Collect non-clinical parameter data related to the periodontal health disease spectrum and pre-process the data;

[0033] Step 3: Collect biological fluids (saliva, mouthwash, etc.) and / or gas samples (exhaled gas oralbreadth), process the samples and quantify the biomarkers therein;

[0034] Step 4: Build a logistic regression model and a random forest model;

[0035] Step 5: Train and optimize the constructed model to obtain the classification system that can accurately classify periodontal health conditions.

[0036] Specifically, in step 1, clinical data will be collected by trained and calibrated experts and the periodontal health status of each subject in a group of individuals will be established, and the data of these individuals will be used to define the training set for building the model. This data collection will be carried out according to the International Classification and Case Definition of Periodontal Diseases and Conditions (Caton et al., 2018) to establish a basic truth / gold standard to evaluate the accuracy of the classification system based on non-clinical parameters.

[0037] In step one, the clinical data collected for the classification of periodontal health status are confirmed through internationally recognized case definitions and used as the gold standard for classification; during the clinical data collection process, quality control is performed to ensure the reliability of the data and thus ensure the accuracy of the gold standard assessment; the quality control elements include training of clinical parameter collection personnel, calibration of high-level intraclass correlation coefficients of necessary clinical parameters, questionnaire validation, and biomarker test reliability validation evaluation.

[0038] In step 2, in the process of obtaining non-clinical parameter data, part or all of the non-clinical parameter data related to the spectrum of periodontal health diseases are obtained by the subjects answering questions in the questionnaire, and the answers are checked to ensure that the answers are within the expected reasonable range.

[0039] In step three, the biofluid / vapor used to measure the relevant biomarkers - the sample - is collected by the subject and pre-treated to remove impurities and make it suitable for measurement using a filtration system and a screw-lock syringe; the biomarkers in the sample are then tested using a rapid test method to see if they can be detected by a preset minimum detection limit.

[0040] In step 4, the results from non-clinical parameters and tests were evaluated to assess their statistical distribution and select appropriate analytical methods; multivariate models (e.g., logistic regression for binary prediction and random forest for multi-class prediction) were constructed to predict health status based on self-reported parameters, tests, and using clinically determined basic facts / gold standards as references. During the model construction process, the model construction met a series of goals: i) high significance of the overall model; ii) elimination of predictors that did not contribute significantly to the overall performance; iii) achieving a high level of diagnostic accuracy for binary and / or multi-class predictions by calculating positive predictive value, negative predictive value, specificity, sensitivity, overall accuracy, and area under the receiver operating characteristic curve AUROC and constructing a confusion matrix, and reducing clinically relevant errors and evaluating the errors of the model; the high level of diagnostic accuracy refers to accuracy>0.90.

[0041] In step five, during the training and optimization of the model, the model is trained and adjusted based on the results of the confusion matrix by providing specific weights in the analysis to reduce the probability of error. For example, declaring periodontally healthy or gingivitis as subjects affected by periodontitis (especially advanced periodontitis). This is achieved by comparing the confusion matrix after applying different types of data transformations, applying light weights or increasingly heavy weights to specific observations and trees in the random forest, or using the mean instead of the mode in the data distribution.

[0042] The specific steps of constructing the logistic regression model are:

[0043] Step 1: Input the training data into the logistic regression model to be optimized, check the distribution pattern of the data in the model, and select the optimization analysis method;

[0044] Step 2, selecting prediction conditions, wherein the prediction conditions include healthy and gingivitis and periodontitis, gingivitis and healthy and periodontitis, and periodontitis and healthy and gingivitis;

[0045] Step 3: Select the positive prediction probability cutoff value, or use the standard receiver operating curve to select the classification result corresponding to the logistic regression model;

[0046] Step 4: Place the predictors into the logistic regression model, run standard model diagnostic tests, and use backward stepwise selection to remove predictors that do not significantly affect the results to obtain the final logistic regression model.

[0047] Step 5: Calculate the diagnostic accuracy parameters of the final model (sensitivity, specificity, positive and negative predictive values, overall accuracy) to evaluate whether it meets the performance criteria (high accuracy) and determine the parameters of the multivariate prediction equation, and use the bootstrap method to enhance the stability of the model; the diagnostic accuracy parameters include sensitivity, specificity, positive and negative predictive values, and overall accuracy.

[0048] The construction of the random forest model includes the following steps:

[0049] Step a: input the training data into the random forest model to be optimized, check the distribution pattern of the data in the model, and select the optimization analysis method;

[0050] Step b, selecting a prediction type, the prediction type includes three categories: 1) periodontal health, gingivitis, periodontitis; 2) periodontal health, gingivitis, periodontitis stage I, periodontitis stage II, periodontitis stage III or periodontitis stage IV; 3) periodontal health, gingivitis / stage I periodontitis, stage II-IV periodontitis;

[0051] Step c. Random forest models were constructed using the Python (version 3.9.13, www.Python.org) Scikit-learn package (Pedregosa et al., 2011). Input data included the most important parameters determined in the binary (logistic) analysis and were based on theoretical assumptions about the importance of different features in the health and disease spectrum. Random forests were generated using 100 estimators (“trees”) per forest, with a maximum depth value of 6. Node splitting was determined using the Gini coefficient. These parameters were set empirically after several preliminary attempts to verify that they produced reliable results. The random forest model was validated by training the entire set of data 100 times independently and comparing the results to ensure the internal consistency of the model. Means and deviations of model performance, including sensitivity, specificity, PPV, NPV, and accuracy, were reported. Sensitivity and specificity values ​​were defined as low (<60%), moderate (60%-79%), or high (≥80%) (Nelson et al., 2001). The accuracy results obtained by AUROC values ​​were interpreted as low (0.50 to 0.70), medium (0.71 to 0.90), and high (>0.90) (Swets, 1988). The package “Yellowbrick” in Python was used to calculate the multi-class AUROC values ​​at the micro and macro average as well as the AUROC for each class.

[0052] Microaverage calculated the global average by adding the true positives and false positives of all categories. Macroaverage calculated the average of the metrics for all categories by taking the average of the curves for all categories. In addition to the micro- and macro-average curves for each category, a curve for each category was plotted to assess the trade-off between sensitivity and specificity for each category. Accuracy was calculated according to the following definition: Accuracy = Sensitivity * Prevalence + Specificity * (1-Prevalence). The MedCalc tool (Version 20.115, MedCalc Software bv, Ostend, Belgium; https: / / www.medcalc.org) was used.

[0053] In step c, during the training process of optimizing the random forest model, a feature importance map is constructed to establish the relative weight of each predictor included in the random forest model; and in multiple iterations, the stability of the model is tested by evaluating the impact of excluding one or more predictors from the model until the model is stable and includes all important predictors.

[0054] The present invention also proposes the application of the logistic regression model and / or random forest model constructed by the above method in the classification of periodontal health status.

[0055] The present invention also proposes a method for classifying periodontal health conditions, which comprises the following steps:

[0056] Step (i) collecting non-clinical parameter data related to periodontal health and preprocessing the data;

[0057] Step (ii), constructing logistic regression model and random forest model;

[0058] Step (iii) using the constructed logistic regression model and random forest model to evaluate and classify periodontal health status.

[0059] The non-clinical parameter data were derived from three consecutive screening tests and a full-mouth periodontal examination;

[0060] The screening test further includes a questionnaire, active matrix metalloproteinase-8 lateral flow immunochromatographic rapid test aMMP-8POCT, gingival bleeding by brushing GBoB test;

[0061] The questionnaire included “Q1 Self-assessment of gum disease”, “Q2 Evaluation of gum / tooth health”, “Q3a Self-reported experience of supragingival scaling”, “Q3b Deep teeth cleaning”, “Q4 Self-perceived tooth movement”, “Q5 Self-reported professionally diagnosed bone loss”, “Q6 Self-perceived changes in tooth appearance”, “Q7 Use of dental floss” and “Q8 Mouthwash”.

[0062] The present invention also proposes the application of the above classification system or the above classification method in the classification and prediction of periodontal health status.

[0063] The benefits of the present invention include: the method of the present invention is different from the prior art because it uses a combination of self-reported non-clinical parameters and self-performed diagnostic tests to assess biological parameters (collectively including three indicator tests), and by using a specific interface, the results can be obtained without the aid of trained professionals.

[0064] Furthermore, the method differs from the prior art in that it is the first method to classify subjects into relevant groups that differ in treatment need and / or level of care required by using a random forest model. The prior art only allows for binary classification, thus requiring different disease states to be grouped in the prediction (e.g. combining healthy and gingivitis to predict periodontitis or combining gingivitis and periodontitis to predict healthy). Since periodontal health is not a binary state (there are multiple conditions that need to be distinguished), the prior art has never achieved a high level of diagnostic accuracy and therefore no method has ever entered routine use. By applying the system of the present invention, users are able to self-determine periodontal health status as periodontal health (requiring ongoing preventive services), gingivitis (superficial disease requiring simple interventions and preventive services) and different stages (severity) of periodontitis requiring primary care (stages I and II) or specialist care (stages III and IV). The system provided by the present invention exhibits high accuracy and is therefore of great use to both users and care health professionals. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 is a flow chart of the present invention for inclusion of participants and study procedures.

[0066] FIG. 2 is a schematic diagram showing the accuracy of the present invention using a random forest classifier to make three types of predictions for periodontal health, gingivitis, and periodontitis. Figure 2A Periodontal health (1, green); gingivitis (2, pink); and periodontitis (3, red). Box plots of sensitivity (upper left), specificity (lower left), positive predictive value (upper middle), negative predictive value (lower middle), and accuracy (upper right) are shown. The lower right plot is a confusion matrix, which plots the frequency of predictions versus actual diagnoses. Values ​​are averaged over 100 runs. Figure 2B . Receiver operating curve plots of true positive rate versus false positive rate and AUROC for 3-class random forest diagnosis using the “yellowbrick” package in Python. Figure 2C . Schematic diagram of the relative impact of multiple factors in random forest classification. Values ​​are averaged over 100 runs.

[0067] FIG3 is a schematic diagram of the accuracy of the six-level prediction of periodontal health, gingivitis, and stage I-IV periodontitis using a random forest classifier according to the present invention. Figure 3AHealthy gums (1, green); gingivitis (2, pink); stage I periodontitis (3, orange), stage II periodontitis (4, light red), stage III periodontitis (5, dark red), and stage IV periodontitis (6, purple). Box plots of sensitivity (upper left), specificity (lower left), positive predictive value (upper middle), negative predictive value (lower middle), and accuracy (upper right) are shown. The lower right plot is a confusion matrix, which plots the frequency of predictions versus actual diagnoses. Values ​​are averaged over 100 runs. Figure 3B . Receiver operating curve plots of true positive rate versus false positive rate and AUROC for 6-class random forest diagnosis using the “yellowbrick” package in Python. Figure 3C . Schematic diagram of the relative impact of multiple factors in a 6-class random forest classification. Values ​​are averaged over 100 runs.

[0068] FIG. 4 shows the accuracy of the present invention using the random forest classifier to predict three categories: periodontal health, gingivitis + stage I periodontitis, and stage II-IV periodontitis. Figure 4A Healthy periodontium (1, green); gingivitis and stage 1 periodontitis (2, orange); and stage II to IV periodontitis (3, red). Box plots of sensitivity (upper left), specificity (lower left), positive predictive value (upper middle), negative predictive value (lower middle), and accuracy (upper right) are shown. The lower right plot is a confusion matrix, which plots the frequency of predictions versus actual diagnoses. Values ​​are averaged over 100 runs. Figure 4B . Receiver operating curve plots of true positive rate versus false positive rate and AUROC for 3-class random forest diagnosis using the “yellowbrick” package in Python. Figure 4C . Schematic diagram of the relative impact of multiple factors in a 3-class random forest classification. Values ​​are averaged over 100 runs. DETAILED DESCRIPTION

[0069] The present invention is further described in detail with reference to the following specific examples and drawings. The process, conditions, experimental methods, etc. for implementing the present invention, except for the contents specifically mentioned below, are all common knowledge and common common sense in the art and are not particularly limited by the present invention.

[0070] The burden of periodontal disease remains high worldwide. Classification of periodontal health status requires a lengthy and sometimes painful clinical examination by a specialist. The diagnostic procedure requires 30 to 60 minutes of clinical time and is therefore costly. Difficult classification and the inability to screen effectively at a population level postpone diagnosis and treatment to later stages of the disease when treatment requires higher resources. The present invention relates to the development and use of a highly accurate self-screening method that can be performed by subjects in a non-clinical setting and accurately triage populations according to the level of care required.

[0071] The present invention combines specific non-clinical features of the periodontal health-disease spectrum to improve the discrimination of periodontal health, gingivitis, periodontitis and their stages with a single algorithm. Specific objectives are: i) to preliminarily evaluate candidate diagnostic methods in multivariate logistic regression; ii) to develop and internally validate a random forest-based screening tool, a machine learning algorithm that correctly classifies subjects into 3 case definitions (periodontal health, gingivitis and periodontitis, or periodontal health, gingivitis and stage I periodontitis, stage II-IV periodontitis) or 6 case definitions (periodontal health, gingivitis, stage I, stage II, stage III and stage IV periodontitis).

[0072] 1.1 Study design and population

[0073] The data sample for this paper involved a convenience sample of consecutive patients who sought dental care at the Philip Dental Hospital (PPDH) between July 2019 and August 2020. The details of the design and methods have been described elsewhere (Deng et al., 2021b). All willing adults aged 18 years or older were eligible to participate; edentulous patients, pregnant women, subjects who had used antibiotics in the past three months, and those with a history of periodontal treatment (except supragingival cleaning) in the past 12 months were excluded. The study protocol has been registered with ClinicalTrials.gov (NCT03928080) and the University of Hong Kong Clinical Trials Unit (HKUCTR-2631). This study received ethical approval from the local Institutional Review Board of the University of Hong Kong / Hong Kong Hospital Authority Hong Kong West Cluster (reference: UW19-188). Written informed consent was obtained from all participants. This study was conducted in accordance with the Declaration of Helsinki and followed the Transparent Reporting of Multivariate Prediction Models for Individual Prognosis or Diagnosis (TRIPOD) guidelines (Collins et al., 2015).

[0074] 1.2 Sample size considerations

[0075] There are no similar screening and diagnostic studies that distinguish between periodontal health, gingivitis, and different stages of periodontitis. Therefore, the sample size calculation was based primarily on existing models for screening for periodontitis, with current methods for detecting gingival inflammation as a secondary consideration. The sample size was estimated based on the observed screening accuracy of the CDC / AAP questionnaire for periodontitis in the 2009-2010 NHANES survey (sensitivity, 59.3%; specificity, 57.0%) (Eke et al., 2013) and the expected sensitivity and specificity of 70.0%. As the most recent systematic review and meta-analysis showed an overall composite sensitivity and specificity of 70% for diagnosing periodontitis using MMP-8 alone (Arias-Bujanda et al., 2020), the minimum acceptable level of the composite model was expected to be 70%. At a 5% significance level and 80% power, 159 participants with periodontitis and 109 participants without periodontitis (periodontal health and gingivitis) were required. Furthermore, there have been no previous investigations into the screening accuracy of gingivitis. One study evaluated the utility of self-reported GBoB for detecting gingival inflammation, defined as BOP ≥ 10%, and determined its sensitivity to be 47%. (Romano et al., 2020). Setting the acceptable sensitivity of a gingivitis screening / diagnostic model at 70%, a minimum of 35 gingivitis cases would be required.

[0076] Considering the 5% missing data and the prevalence of periodontitis in Hong Kong ranging from 40% to 59% (Department of Health, Government of the Hong Kong Special Administrative Region, 2011), a total of 400 subjects were recruited. According to the rule that the ratio of the number of subjects to the number of prediction items should be at least 3 (Mundfrom et al., 2005), the sample size of this study was sufficient for internal validation: the ratio of the number of subjects with different definitions of periodontal disease to the number of prediction items ranged from 3.4 to 17.5.

[0077] 1.3 Predictor - Index Test

[0078] Study procedures included three consecutive screening tests (index tests) and a full-mouth periodontal examination (reference standard). Demographic characteristics, including age (years) and sex (male / female), lifestyle factors, including smoking status (current, never / former), and personal medical history (e.g., presence of cardiovascular disease and diabetes) were collected by trained nurses through face-to-face interviews. Screening tests included a self-administered questionnaire, followed by oral rinse aMMP-8POCT, and finally GBoB testing.

[0079] The development and validation of the Chinese version of the CDC / AAP questionnaire followed a standardized forward-backward translation procedure, and a cognitive assessment was performed before the questionnaire was administered to patients. (Deng et al., 2021b). The full set of questionnaires included self-assessment of “periodontal disease” (Q1), “evaluation of gum / tooth health” (Q2), self-reported experience of “supragingival scaling” (Q3a), “deep scaling” (Q3b), self-perceived tooth movement (Q4), self-reported professionally diagnosed bone loss (Q5), self-perceived changes in “tooth appearance” (Q6), use of dental floss (Q7) and mouthwash (Q8).

[0080] The POCT used a commercial aMMP-8 lateral flow immunoassay system ( PRO, Dentognostics GmbH, Jena, Germany) and its digital analysis device ( Dentognostics GmbH, Jena, Germany). The aMMP-8 test was performed on participants by trained and calibrated nurses according to the manufacturer's instructions: 1) pre-rinse with tap water for 30 seconds; 2) wait for 60 seconds; 3) rinse with 5 mL of purified water for 30 seconds; 4) pour the oral rinse solution and draw 3 mL into the syringe; 5) put the filter into the syringe and add 3-4 drops of the filtered sample to the test system. The quantitative determination of the aMMP-8 concentration was displayed on the screen within 5 minutes. The test result was considered positive if the aMMP-8 concentration was >10 ng / mL; concentrations below the detection level were considered to be 10 ng / mL in statistical analysis.

[0081] GBoB testing was performed by the same nurse in a separate room with a mirror and wash basin in a standard manner: 1) all participants were asked to carry out their daily routine without any distractions; 2) their saliva / toothpaste slurry (TPS) was collected and stored for further quantitative analysis of Hb at the end of each session; 3) participants were asked to self-assess the presence of blood in the TPS (GBoB). Quantitative analysis of Hb in TPS samples by ultraviolet-visible (UV-Vis) spectroscopy has been described in detail previously (Deng et al., 2021c).

[0082] 1.4 Periodontal examination and clinical diagnosis - reference standards

[0083] A trained and calibrated examiner (DK), blinded to the results of all index tests, performed a comprehensive periodontal examination, which included measurement of periodontal probing depth (PPD), bleeding on probing (BOP), clinical attachment level (CAL), furcation lesions (FI) (Hamp et al., 1975), tooth mobility (Miller classification, 1950), and number of missing teeth due to periodontitis at six sites using a standardized periodontal probe (UNC-15, Hu Friedy, Chicago).

[0084] Case definitions for periodontal health, gingivitis, and periodontitis were based on the 2017 World Workshop on the Classification of Periodontal and Peri-implant Diseases and Conditions (Chapple et al., 2018; Papapanou et al., 2018; Tonetti et al., 2018; Trombelli et al., 2018). In the manuscript, “periodontal disease” refers to plaque-induced gingivitis and periodontitis to avoid misunderstandings that include a wider range of gum diseases and other forms of periodontitis. Periodontal health cases were defined as the absence of gingival inflammation (BOP < 10%) and attachment loss attributed to periodontitis (Chapple et al., 2018). Gingivitis cases were defined as the presence of gingival inflammation (BOP ≥ 10%) and no detection of attachment loss due to periodontitis (Trombelli et al., 2018). A periodontitis case was defined as the detection of interdental attachment loss on at least two non-adjacent teeth. The stage of periodontitis is determined by severity and complexity (Tonetti et al., 2018).

[0085] 1.5 Quality Control

[0086] To ensure data quality, all investigators were trained and calibrated before the implementation of the study; all standard procedures were strictly implemented and the unknown interpretation of the study results was ensured. In particular, the self-administered questionnaire was pre-tested during the pilot phase and adjusted accordingly before validation. In the aMMP-8 test, subjects were required to avoid eating, drinking, brushing teeth, and using mouthwash for at least 30 minutes before the test. (Deng et al., 2021a). Before sample analysis, the quantitative method of Hb was validated in a pilot experiment and had good validity and repeatability. Clinical examiners were trained on standardized diagnostic criteria and calibrated with reference examiners, with good reliability. The intra-examiner reliability assessed by the intraclass correlation coefficient (ICC) of PPD was greater than 0.90, and the ICC of CAL was greater than 0.83. (Deng et al., 2021b).

[0087] 1.6 Data Analysis

[0088] The analytical approach consisted of two steps: i) preliminary evaluation using multivariable logistic regression and ii) evaluation of multi-category screening tools that included different features of the periodontal health-disease spectrum using forest plots. IBM SPSS Statistics, version 26 (IBM Corp., Armonk, NY, USA) was used for descriptive and logistic regression analyses. To explore the discriminatory performance of the composite index test for differentiating between various periodontal diseases, multivariable logistic regression models were applied to predict “periodontal health,” “gingivitis,” “periodontitis,” “stage I / II periodontitis,” and “stage III / IV periodontitis,” respectively, according to the distribution of the disease in the study population. The algorithm to detect each condition was developed using the best predictor from the logistic regression model. Internal validation was performed using the bootstrap method. One thousand bootstrap samples were generated with replacement. The final logistic regression model included significant predictors using backward stepwise selection. Wald tests were performed on the parameters associated with each predictor and it was removed if it was not significant at 95% confidence. The model was rerun and the process was repeated until no more predictors were removed. Model 1 was a combination of the questionnaire, aMMP-8POCT, and GBoB test. Model 2 was a combination of the composite index test with demographic and lifestyle behavior factors. Based on the predicted probability of the logistic regression model, the receiving operating characteristic (ROC) curve, the area under the ROC curve (AUROC), sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were estimated. The predicted probability cutoff was chosen as the point that maximized the sum of sensitivity and specificity of the entire ROC curve.

[0089] In constructing the multiclass model, the unique features describing the periodontal health-disease spectrum were used to train the random forest algorithm, and in case of multiple measurements, the most important predictors obtained from the logistic regression analysis were used as input data. The multiclass decision algorithm for distinguishing subjects between periodontal health, gingivitis, and different stages of periodontitis was applied to the selected variables to train and construct the final model. Random forests were generated using 100 estimators (“trees”) per forest with a maximum depth value of 6. Node splits were determined using the Gini coefficient. These parameters were set empirically after several preliminary attempts to verify that they could produce reliable results. The random forest model was validated by running 100 independent training runs on the entire dataset and comparing the results to ensure the internal consistency of the model. The random forest model was fitted using the Python (version 3.9.13, www.python.org) Scikit-learn package (Pedregosa et al., 2011). The mean and deviation of model performance, including sensitivity, specificity, PPV, NPV, and accuracy, are reported. Sensitivity and specificity values ​​were defined as low (<60%), moderate (60% to 79%), or high (≥80%). (Nelson et al., 2001). Accuracy results derived from AUROC values ​​were interpreted as low (0.50 to 0.70), moderate (0.71 to 0.90), and high (>0.90), respectively. (Swets, 1988). The “Yellowbrick” package in Python was used to calculate micro- and macro-averaged multi-class AUROC values ​​as well as per-class AUROC values.

[0090] The micro-average method calculates a global average by adding the true positives and false positives of all categories. The macro-average method calculates the average of the indicators for all categories by taking the average of the curves of all categories. In addition to the micro- and macro-average curves for each category, a curve for each category was also plotted to evaluate the trade-off between sensitivity and specificity for each category. Accuracy was calculated according to the following definition: Accuracy = Sensitivity * Prevalence + Specificity * (1-Prevalence). 95% confidence intervals were obtained using the MedCalc tool (Version 20.115, MedCalc Software bv, Ostend, Belgium; https: / / www.medcalc.org).

[0091] 2. Results

[0092] After evaluating the eligibility criteria from 501 consecutive patients seeking dental treatment in the receiving clinic, 408 eligible Chinese subjects were included in this study, with a participation rate of 81.4% ( Figure 1). The demographic and clinical characteristics of the study population have been reported and are also shown in the Appendix (Tables S1-S2). (Deng et al., 2021a, 2021b). In brief, 66 (16.2%) subjects were periodontally healthy cases and 62 (15.2%) had gingivitis. 280 (68.6%) subjects were diagnosed with periodontitis, including stage I (15.9%), stage II (15.9%), stage III (29.7%), and stage IV (7.1%) periodontitis. There were 189 (46.3%) males and 219 (53.7%) females, aged between 18 and 86 (41 ± 18) years. 32 (7.8%) subjects were current smokers, and 20 (4.9%) had diabetes mellitus (DM). All subjects completed the questionnaire and aMMP-8 and GBoB tests, except for one invalid Hb measurement sample. Therefore, the entire dataset of self-reported parameters and salivary biomarker profiles of 407 subjects was used to create logistic regression and RF models to discriminate between various periodontal diagnoses.

[0093] 2.1 Prediction Accuracy of Binary Classifier and Logistic Regression

[0094] The performance of self-reported nonclinical parameters, salivary biomarkers, and Models 1-2 in predicting periodontal health, gingivitis, and different stages of periodontitis are shown in Tables 1-5 below.

[0095] Table 1. Diagnostic accuracy of self-reported nonclinical parameters and salivary biomarkers alone and in combination for periodontal health in the study sample (N = 408) using conventional logistic regression analysis.

[0096]

[0097] OR, probability; 95% CI, 95% confidence interval; AUROC, area under the receiver characteristic curve; PPV, positive predictive value; NPV, negative predictive value. Model 1 represents the combination of CDC-AAP questionnaire, aMMP-8POCT, and gingival bleeding during toothbrushing. Model 2 is model 1 plus age and smoking status.

[0098] Table 2. Diagnostic accuracy of self-reported non-clinical parameters or salivary biomarkers alone and in combination for gingivitis in the study sample (N=408) using conventional logistic regression analysis.

[0099]

[0100] OR, probability; 95% CI, 95% confidence interval; AUROC, area under the receiver characteristic curve; PPV, positive predictive value; NPV, negative predictive value. Model 1 represents the combination of CDC-AAP questionnaire, aMMP-8POCT, and gingival bleeding during toothbrushing. Model 2 is model 1 plus age and smoking status.

[0101] Table 3. Diagnostic accuracy of self-reported non-clinical parameters or salivary biomarkers alone and in combination for periodontitis in the study sample (N=408) using conventional logistic regression analysis.

[0102]

[0103]

[0104] OR, probability; 95% CI, 95% confidence interval; AUROC, area under the receiver characteristic curve; PPV, positive predictive value; NPV, negative predictive value. Model 1 represents the combination of CDC-AAP questionnaire, aMMP-8POCT, and gingival bleeding during toothbrushing. Model 2 is model 1 plus age and smoking status.

[0105] Table 4. Diagnostic accuracy of self-reported non-clinical parameters or salivary biomarkers alone and in combination for stage I / II periodontitis in the study sample (N=408) using conventional logistic regression analysis.

[0106]

[0107] OR, probability; 95% CI, 95% confidence interval; AUROC, area under the receiver characteristic curve; PPV, positive predictive value; NPV, negative predictive value. Model 1 represents the combination of CDC-AAP questionnaire, aMMP-8POCT, and gingival bleeding during toothbrushing. Model 2 is model 1 plus age and smoking status.

[0108] NA: Not applicable because the predictor was not significant in the model.

[0109] Table 5. Diagnostic accuracy of self-reported non-clinical parameters or salivary biomarkers alone and in combination for stage III / IV periodontitis in the study sample (N=408) using conventional logistic regression analysis.

[0110]

[0111] OR, probability; 95% CI, 95% confidence interval; AUROC, area under the receiver characteristic curve; PPV, positive predictive value; NPV, negative predictive value. Model 1 represents the combination of CDC-AAP questionnaire, aMMP-8POCT, and gingival bleeding during toothbrushing. Model 2 is model 1 plus age and smoking status.

[0112] NA: Not applicable because the predictor was not significant in the model.

[0113] Results of conventional logistic regression analysis indicated that: i) the combination of self-reported nonclinical parameters and salivary biomarkers (model 1) achieved high accuracy in discriminating periodontal health and disease and moderate accuracy in detecting gingivitis; ii) the combined use of aMMP-8POCT and the CDC / AAP questionnaire slightly improved the screening accuracy for periodontitis compared with the questionnaire alone; iii) recognized risk factors / indicators (age and smoking) showed moderate / high accuracy in detecting periodontitis / stage III / IV periodontitis, respectively; iv) the predictors alone or in combination performed poorly in identifying stage I / II periodontitis; v) the CDC / AAP tool had moderate to high accuracy in screening for stage III / IV periodontitis and performed excellently when combined with age and smoking to strongly predict stage III / IV periodontitis.

[0114] According to Model 2 (Appendix, Tables S3–S7), the fitted linear predictor for the estimated logarithm of periodontal health was = −1.750*Q2+0.844*Q3a+0.951*Q7-0.027*Hb (total)-1.471*aMMP-8POCT-0.050*age+2.112; the fitted linear predictor for the estimated logarithm of gingivitis was = −0.992*Q3a+1.396*self-reported GBoB-0.160*age-14.719; the fitted linear predictor for the estimated logarithm of periodontitis .... The linear predictor was =1.407*Q2+1.845*aMMP-8POCT+0.102*age-4.203; the fitted linear predictor for estimating the logarithm of stage I / II periodontitis was =-1.540*Q4+1.022*aMMP-8POCT+0.521*sex-0.963; the fitted linear predictor for estimating the logarithm of stage III / IV periodontitis was =1.729*Q2+1.756*Q6+0.115*age+2.453*smoking-8.024.

[0115] 2.2 Multi-class classification and performance using random forests

[0116] Eleven parameters estimating different features of the periodontal health-disease spectrum were included in the RF: “gum disease” (Q1), “gingival / dental health assessment” (Q2), “tooth cleaning” (Q3a), “loose teeth” symptoms (Q4), “tooth appearance” (Q6), “flossing” (Q7), aMMP-8POCT, self-reported GBoB, Hb, smoking, and age. The robustness of the RF model was assessed by: i) analyzing the pattern of misclassification; ii) the method of leaving out or including. The method of leaving out or including refers to removing or (and) adding a predictor from the model one by one to assess whether the accuracy is improved. The misclassification pattern refers to the misclassification of different case definitions (e.g., underestimating stage III / IV periodontitis as periodontally healthy cases or overestimating periodontally healthy cases as periodontitis cases). Focusing on the more serious misclassification (declaring a subject with periodontitis as healthy, N = 17), the pattern of misclassification showed that all errors occurred in subjects who were younger (less than 35 years old), reported that their teeth looked good, used dental floss, did not bleed when brushing, and had a negative aMMP-8 test. Overall, except for smoking (which only accounted for 7.8% of the population) and "tooth appearance" (Q6), which did not significantly affect the estimates, these two factors were removed, and the exclusion analysis resulted in a decrease in precision. The best combination of predictors that distinguished periodontal health, gingivitis, and different stages of periodontitis was "periodontal disease" (Q1), "rating of gum / teeth health" (Q2), "teeth washing" (Q3a), symptoms of "loose teeth" (Q4), "flossing" (Q7), aMMP-8 POCT, self-reported GBoB, Hb, and age. Figures 2-4 and Tables S17-19 show the performance of three-category discrimination (periodontal health, gingivitis, periodontitis), six-category discrimination (periodontal health, gingivitis, periodontitis stages I-IV), and three-category discrimination combining gingivitis and stage I periodontitis (periodontal health, gingivitis + stage I periodontitis, stage II-IV periodontitis). Overall, the RF classifier provided high accuracy for multi-class discrimination compared with the results obtained by the logistic regression algorithm for binary classification. In 100 runs of the RF algorithm, the classifier had a sensitivity of 89.3% ± 2.2%, a specificity of 92.3% ± 0.4%, and an accuracy of 91.8% ± 0.5% for periodontal health; a sensitivity of 91.0% ± 1.1%, a specificity of 92.6% ± 0.6%, and an accuracy of 92.4% ± 0.6% for periodontal health; and a sensitivity of 91.0% ± 1.1%, a specificity of 92.6% ± 0.6%, and an accuracy of 92.4% ± 0.6% for periodontitis in the three-category classification (Figure 2). It is noteworthy that the sensitivity for healthy and stage I-II periodontitis was high, much better than the results of logistic regression analysis (Figure 3). As shown in Figure 3, the probability of misclassification of periodontal health, gingivitis, and different stages of periodontitis was low (< 6%). Among them, underestimation of periodontitis cases was more frequent than overestimation.

[0117] 3. Discussion

[0118] The present invention demonstrates that incorporating non-clinical parameters that reflect unique characteristics of periodontal health, gingivitis, and periodontitis in a machine learning algorithm allows the development of a multi-class prediction / diagnosis model of periodontal health with high accuracy. The random forest model had high accuracy (AUROC > 0.94) for three-class predictions (healthy, gingivitis, and periodontitis or healthy, gingivitis / stage I periodontitis, and stages II, III, and IV periodontitis) and six-class predictions (healthy, gingivitis, stages I, II, III, and IV periodontitis). The level of accuracy achieved by the RF multi-class classifier was significantly better than that of multivariate logistic regression analysis, especially for the more difficult diagnoses to distinguish: gingivitis, stage I, and stage II periodontitis. Using RF, the probability of misclassification was low ( Figure 2A , 3A and 4A). Further studies in independent populations are needed to validate the model and approach. These results are noteworthy because multi-category screening of periodontal health in nonclinical settings may provide specific utility and increase the efficiency of managing preventive and therapeutic services / care pathways. Ideally, screening results may be able to identify subjects who only require primary preventive services (Tonetti et al., 2015), those who need management of gingivitis before periodontitis develops, and those who should receive periodontitis treatment in the primary setting (stages I and II periodontitis) or in a specialist setting (stages III and IV periodontitis) (Tonetti et al., 2017).

[0119] Nine predictors, namely “periodontal disease” (Q1), “rating of gum / teeth health” (Q2), “teeth washing” (Q3a), symptoms of “loose teeth” (Q4), “flossing” (Q7), aMMP-8POCT, self-reported GBoB, Hb, and age, resulted in high accuracy. Random forest is a machine learning algorithm widely used for classification that generates multiple decision trees and uses randomly selected subsets of variables to reduce overfitting and bias (Breiman, 2001). The random forest model is stable, leaving out one parameter and using multiple seeds to build / train the model. Mechanistically, the predictors in the model provide information about different potential features of the periodontal disease spectrum (gingivitis and various stages of periodontitis), which increases biological plausibility. Interestingly, the analysis of the influencing factors ( Figure 2C , 3C and 4C) ranked slightly differently for the three test options. This appears to be mechanistically consistent with the current understanding of the different characteristics of periodontal health, gingivitis, and different stages of periodontitis.

[0120] Notably, the set of predictors used in the machine learning tool for multi-class identification performed better than the logistic regression model in bivariate analysis. Machine learning algorithms have been reported to have improved accuracy in multi-class identification of complex diseases such as diabetes or cardiovascular disease (Baashar et al., 2022; Shin et al., 2022). In the present study, machine learning algorithms improved the accuracy of detection of stage I and stage II periodontitis from an AUROC of 0.68 for logistic regression to an AUROC of 0.97 for predicting stage I periodontitis and 0.94 for predicting stage II periodontitis for the random forest classifier. Identification of incipient periodontitis offers the most significant challenge, even for experienced clinicians performing a complete periodontal examination. The application of artificial intelligence based on non-clinical parameters may be handy for this diagnostic problem. As expected, the prediction accuracy for more severe forms of the disease was high even when using only the CDC / AAP questionnaire. Judging from the 6 by 6 (3 by 3) confusion matrix of the random forest model, the value of adding additional parameters lies in the ability to better distinguish between more subtle / initial stages in the disease spectrum.

[0121] Logistic regression analysis revealed additional findings. Although the added value of aMMP-8 and GBoB for periodontitis screening was small, the combined predictor performed best in differentiating periodontal health from disease and detecting gingivitis compared with individual questionnaires or biomarker tests. Although requiring slightly more resources and medical costs, the combined predictor has substantial potential public health benefits: Nearly 80% of suspected periodontal disease cases, including gingivitis, stage I / II periodontitis, and stage III / IV periodontitis, could be identified for early referral, further professional diagnosis, and timely intervention. Detection of gingivitis and incipient periodontitis in particular could contribute to more favorable treatment outcomes and less medical costs, thereby mitigating socioeconomic consequences and promoting periodontal health and overall well-being. Consistent with the logistic regression analysis, the feature importance plots from the RF also reinforced the rationale for incorporating biomarker testing (such as Hb), consistent with current medical understanding. The different significance of GBoB in gingivitis screening and aMMP-8POCT in periodontitis highlight the key features of these different diseases: superficial inflammation reflected by bleeding during toothbrushing and periodontal destruction reflected by MMP activation. A diagnostic strategy combining them may show higher overall accuracy.

[0122] As mentioned previously, individual questions of the CDC / AAP questionnaire appear to be more suitable for confirming / excluding specific diagnoses (Deng et al., 2021b). Combining them into specific sets may help to achieve a multi-level classification of periodontal health status. In addition, models that incorporate recognized risk indicators (such as age or smoking status) are generally more accurate. In this regard, it must be emphasized that this part of the model is particularly prone to errors in populations with different characteristics. Extensive validation is needed to better estimate their importance and contribution to the model.

[0123] Artificial intelligence prediction based on nonclinical data is a promising approach to classify periodontal health status and has the potential to open new avenues for better management of periodontitis.

[0124] appendix

[0125] Table S1 Demographic and systematic characteristics of the study sample by periodontal case definition.

[0126]

[0127] Data are expressed as mean ± SD or n (%); H, periodontal health; Data are expressed as mean ± SD or n (%); H, periodontal health; G, gingivitis; I, stage I periodontitis; II, stage II periodontitis; III, stage III periodontitis; IV, stage IV periodontitis; P, periodontitis, CVD, cardiovascular disease; DM, diabetes mellitus; poorly controlled diabetes: HbA1c ≥ 7.0%; heavy smokers: ≥ 10 cigarettes / day; former smokers: quit smoking ≥ 3 years ago. Chi-square test or Fisher's exact test (for categorical data) and Kruskal-Wallis test (for continuous data) were used to evaluate differences between groups.

[0128] Table S2. Responses to the CDC-AAP questionnaire by periodontal case definition.

[0129]

[0130]

[0131] aThe reference category is "No"

[0132] bThe reference category was a combination of “excellent,” “very good,” and “good,” and responses of “fair” and “poor” were also combined.

[0133] c “Don’t know” and missing responses were excluded from further univariate analysis

[0134] Chi-square test (Q1, Q4, Q5, Q7, Q8) or Fisher’s exact test (Q 2, Q 3a, Q3b, Q6) was performed to compare the differences in responses to each question: if it was shown that there was a difference between the groups (P < 0.05), then pairwise comparisons between the groups were performed using the Bonferroni method to adjust the significance level for multiple comparisons

[0135] For each question, the percentages with different superscripts (e.g., A, B, and C) are significantly different.

[0136] Table S3. Logistic regression analysis of Model 1 and Model 2 for predicting periodontal health status based on the entire sample after internal validation.

[0137]

[0138] X, predictor variable retained in the final model; OR, odds ratio; 95% CI, 95% confidence interval; AUROC, area under the receiver operating characteristic curve; ***P < 0.001, **P < 0.01, *P < 0.05; The reference category for Q1, Q3a, Q3b, Q4, Q5, Q6, Q7, Q8 was a “no” response; the reference category for Q2 was a combination of “good, very good, and excellent” responses; the reference category for current smokers was never / former smokers; for age, the β estimates represent an increase of one year in age; for hemoglobin, the β estimates represent an increase of one unit (μg / ml)

[0139] Table S4. Logistic regression analysis of Model 1 and Model 2 for predicting gingivitis from the entire sample after internal validation.

[0140]

[0141] X, predictor variable retained in the final model; OR, odds ratio; 95% CI, 95% confidence interval; AUROC, area under the receiver operating characteristic curve; ***P < 0.001, **P < 0.01, *P < 0.05; The reference category for Q1, Q3a, Q3b, Q4, Q5, Q6, Q7, Q8 was a “no” response; the reference category for Q2 was a combination of “good, very good, and excellent” responses; the reference category for current smokers was never / former smokers; for age, the β estimates represent an increase of one year in age; for hemoglobin, the β estimates represent an increase of one unit (μg / ml)

[0142] Table S5. Logistic regression analysis of Model 1 and Model 2 for predicting periodontitis from the entire sample after internal validation.

[0143]

[0144] X, retained in the final model; OR, odds ratio; 95% CI, 95% confidence interval; AUROC, area under the receiver operating characteristic curve; ***P < 0.001, **P < 0.01, *P < 0.05; The reference category for Q1, Q3a, Q3b, Q4, Q5, Q6, Q7, Q8 was a “no” response; the reference category for Q2 was a combination of “good, very good, and excellent” responses; the reference category for current smokers was never / former smokers; for age, the β estimates represented an increase of one year in age; for hemoglobin, the β estimates represented an increase of one unit (μg / ml)

[0145] Table S6. Logistic regression analysis of Model 1 and Model 2 for the prediction of stage I / II periodontitis from the entire sample after internal validation.

[0146]

[0147] X, retained in the final model; OR, odds ratio; 95% CI, 95% confidence interval; AUROC, area under the receiver operating characteristic curve; ***P < 0.001, **P < 0.01, *P < 0.05; The reference category for Q1, Q3a, Q3b, Q4, Q5, Q6, Q7, Q8 was a “no” response; the reference category for Q2 was a combination of “good, very good, and excellent” responses; the reference category for current smokers was never / former smokers; for age, the β estimates represented an increase of one year in age; for hemoglobin, the β estimates represented an increase of one unit (μg / ml)

[0148] Table S7. Logistic regression analysis of Model 1 and Model 2 predicting stage III / IV periodontitis in the entire sample after internal validation.

[0149]

[0150] X, predictor variable retained in the final model; OR, odds ratio; 95% CI, 95% confidence interval; AUROC, area under the receiver operating characteristic curve; ***P < 0.001, **P < 0.01, *P < 0.05; The reference category for Q1, Q3a, Q3b, Q4, Q5, Q6, Q7, Q8 was a “no” response; the reference category for Q2 was a combination of “good, very good, and excellent” responses; the reference category for current smokers was never / former smokers; for age, the β estimates represent an increase of one year in age; for hemoglobin, the β estimates represent an increase of one unit (μg / ml)

[0151] Table S8. Bootstrap 95% confidence intervals for beta coefficients of predictor variables in the logistic regression model for periodontal health: bootstrap validation.

[0152]

[0153] X, predictor variables for modeling; To prevent overfitting of the prediction model, only candidate predictor variables that showed univariate association with periodontal disease in the questionnaire were selected for predictive modeling, with a cutoff p value of 0.20; 95% CI, confidence interval of 95%; the reference category for Q1, Q3a, Q3b, Q4, Q5, Q6, Q7, Q8 was a “no” answer; the reference category for Q2 was a combination of “good, very good, and excellent” answers; the reference category for current smokers was never / former smokers; for age, the beta estimate represented an increase of one year in age; for hemoglobin, the beta estimate represented an increase of one unit (μg / ml) Table S 9. Bootstrap 95% confidence intervals for beta coefficients of predictor variables in the logistic regression model for gingivitis: Bootstrap validation.

[0154]

[0155] X, predictor variable for modeling; To prevent overfitting of the prediction model, only candidate predictor variables that showed univariate association with periodontal disease in the questionnaire were selected for predictive modeling, with a cutoff p value of 0.20; 95% CI, confidence interval of 95%; The reference category for Q1, Q3a, Q3b, Q4, Q5, Q6, Q7, Q8 was a “no” answer; The reference category for Q2 was a combination of “good, very good, and excellent” answers; The reference category for current smokers was never / former smokers; For age, the beta estimate represented an increase of one year in age; For hemoglobin, the beta estimate represented an increase of one unit (μg / ml) Table S10. Bootstrap 95% confidence intervals for beta coefficients of predictor variables in the logistic regression model for periodontitis: Bootstrap validation.

[0156]

[0157] X, predictor variable for modeling; To prevent overfitting of the prediction model, only candidate predictor variables that showed univariate association with periodontal disease in the questionnaire were selected for predictive modeling, with a cutoff p value of 0.20; 95% CI, confidence interval of 95%; the reference category for Q1, Q3a, Q3b, Q4, Q5, Q6, Q7, Q8 was a “no” answer; the reference category for Q2 was a combination of “good, very good, and excellent” answers; the reference category for current smokers was never / former smokers; for age, the beta estimate represented an increase of one year in age; for hemoglobin, the beta estimate represented an increase of one unit (μg / ml) Table S11. Bootstrap 95% confidence intervals for beta coefficients of predictor variables in the logistic regression model for stage I / II periodontitis: bootstrap validation.

[0158]

[0159] X, predictor variables for modeling; To prevent overfitting of the prediction model, only candidate predictor variables that showed univariate association with periodontal disease in the questionnaire were selected for prediction modeling, with a cutoff p value of 0.20; 95% CI, confidence interval of 95%; The reference category for Q1, Q3a, Q3b, Q4, Q5, Q6, Q7, Q8 was a “no” answer; The reference category for Q2 was a combination of “good, very good, and excellent” answers; The reference category for current smokers was never / former smokers; For age, the β estimate represented an increase of one year in age; For hemoglobin, the β estimate represented an increase of one unit (μg / ml)

[0160] Table S12. Bootstrap 95% confidence intervals for beta coefficients of predictor variables in the logistic regression model for stage III / IV periodontitis: bootstrap validation.

[0161]

[0162] X, modeled predictor variable; 95% CI, 95% confidence interval; reference category for Q1, Q3a, Q3b, Q4, Q5, Q6, Q7, Q8 was a “no” response; reference category for Q2 was a combination of “good, very good, and excellent” responses; reference category for current smokers was never / former smokers; for age, β estimates represented an increase of one year in age; for hemoglobin, β estimates represented an increase of one unit (μg / ml)

[0163] Table S13. Conventional logistic regression analysis of Model 1 and Model 2 for predicting periodontal health status from the entire sample.

[0164]

[0165] X, retained in the final model; OR, odds ratio; 95% CI, 95% confidence interval; AUROC, area under the receiver operating characteristic curve; ***P < 0.001, **P < 0.01, *P < 0.05; The reference category for Q1, Q3a, Q3b, Q4, Q5, Q6, Q7, Q8 was a “no” response; the reference category for Q2 was a combination of “good, very good, and excellent” responses; the reference category for current smokers was never / former smokers; for age, the β estimates represented an increase of one year in age; for hemoglobin, the β estimates represented an increase of one unit (μg / ml)

[0166] Table S14.1. Conventional logistic regression analysis of Model 1 and Model 2 for the prediction of gingivitis in the entire population.

[0167]

[0168] X, retained in the final model; OR, odds ratio; 95% CI, 95% confidence interval; AUROC, area under the receiver operating characteristic curve; ***P < 0.001, **P < 0.01, *P < 0.05; The reference category for Q1, Q3a, Q3b, Q4, Q5, Q6, Q7, Q8 was a “no” response; the reference category for Q2 was a combination of “good, very good, and excellent” responses; the reference category for current smokers was never / former smokers; for age, the β estimates represented an increase of one year in age; for hemoglobin, the β estimates represented an increase of one unit (μg / ml)

[0169] Table S14.2. Conventional logistic regression analysis of Model 1 and Model 2 for predicting periodontitis from the entire sample.

[0170]

[0171] X, retained in the final model; OR, odds ratio; 95% CI, 95% confidence interval; AUROC, area under the receiver operating characteristic curve; ***P < 0.001, **P < 0.01, *P < 0.05; The reference category for Q1, Q3a, Q3b, Q4, Q5, Q6, Q7, Q8 was a “no” response; the reference category for Q2 was a combination of “good, very good, and excellent” responses; the reference category for current smokers was never / former smokers; for age, the β estimates represented an increase of one year in age; for hemoglobin, the β estimates represented an increase of one unit (μg / ml)

[0172] Table S15. Conventional logistic regression analysis of Model 1 and Model 2 for predicting stage I / II periodontitis from the entire sample.

[0173]

[0174] X, retained in the final model; OR, odds ratio; 95% CI, 95% confidence interval; AUROC, area under the receiver operating characteristic curve; ***P < 0.001, **P < 0.01, *P < 0.05; The reference category for Q1, Q3a, Q3b, Q4, Q5, Q6, Q7, Q8 was a “no” response; the reference category for Q2 was a combination of “good, very good, and excellent” responses; the reference category for current smokers was never / former smokers; for age, the β estimates represented an increase of one year in age; for hemoglobin, the β estimates represented an increase of one unit (μg / ml)

[0175] Table S16. Conventional logistic regression analysis of Model 1 and Model 2 predicting stage III / IV periodontitis in the entire sample.

[0176]

[0177] X, retained in the final model; OR, odds ratio; 95% CI, 95% confidence interval; AUROC, area under the receiver operating characteristic curve; ***P < 0.001, **P < 0.01, *P < 0.05; The reference category for Q1, Q3a, Q3b, Q4, Q5, Q6, Q7, Q8 was a “no” response; the reference category for Q2 was a combination of “good, very good, and excellent” responses; the reference category for current smokers was never / former smokers; for age, the β estimates represented an increase of one year in age; for hemoglobin, the β estimates represented an increase of one unit (μg / ml)

[0178] Table S17. Accuracy of three-level classification using random forest classifier for predicting periodontal health, gingivitis, and periodontitis.

[0179]

[0180] Data are expressed as mean ± SD

[0181] Table S18. Accuracy of six-class prediction of periodontal health, gingivitis, and stage I-IV periodontitis using random forest classifier.

[0182]

[0183] Data are expressed as mean ± SD

[0184] Table S19. Accuracy of three-level prediction using random forest classifier for periodontal health, gingivitis + stage I periodontitis, and stage II-IV periodontitis.

[0185]

[0186] Data are expressed as mean ± SD

[0187] References

[0188] Arias-Bujanda,N.,Regueira-Iglesias,A.,Balsa-Castro,C.,et al.(2020).Accuracy of single molecular biomarkers in saliva for the diagnosis ofperiodontitis:A systematic review and meta-analysis.J Clin Periodontol,47(1),2-18.doi:10.1111 / jcpe.13202

[0189] Aro,K.,Wei,F.,Wong,D.T.,et al.(2017).Saliva Liquid Biopsy for Point-of-Care Applications.Front Public Health,5,77.doi:10.3389 / fpubh.2017.00077

[0190] Baashar,Y.,Alkawsi,G.,Alhussian,H.,et al.(2022).Effectiveness ofArtificial Intelligence Models for Cardiovascular Disease Prediction:NetworkMeta-Analysis.Comput Intell Neurosci,2022,5849995.doi:10.1155 / 2022 / 5849995

[0191] Bashir,N.Z.,Rahman,Z.,&Chen,S.L.(2022).Systematic comparison ofmachine learning algorithms to develop and validate predictive models forperiodontitis.JClin Periodontol.doi:10.1111 / jcpe.13692

[0192] Birkedal-Hansen,H.,Moore,W.G.,Bodden,M.K.,et al.(1993).Matrixmetalloproteinases:a review.Crit Rev Oral Biol Med,4(2),197-250.doi:10.1177 / 10454411930040020401

[0193] Breiman,L.(2001).Random Forests.Machine Learning,45(1),5-32.doi:10.1023 / A:1010933404324

[0194] Carra,M.C.,Gueguen,A.,Thomas,F.,et al.(2018).Self-report assessmentof severe periodontitis:Periodontal screening score development.J ClinPeriodontol,45(7),818-831.doi:10.1111 / jcpe.12899

[0195] Chapple,I.L.C.,Mealey,B.L.,Van Dyke,T.E.,et al.(2018).Periodontalhealth and gingival diseases and conditions on an intact and areduced periodontium:Consensus report of workgroup 1 of the 2017 WorldWorkshop on the Classification ofPeriodontal and Peri-Implant Diseases andConditions.J Clin Periodontol,45 Suppl20,S68-s77.doi:10.1111 / jcpe.12940

[0196] Collins,G.S.,Reitsma,J.B.,Altman,D.G.,et al.(2015).Transparentreporting of a multivariable prediction model for individualprognosis or diagnosis(TRIPOD):the TRIPOD Statement.BMC Medicine,13(1),1.doi:10.1186 / s12916-014-0241-z

[0197] Deng,K.,Pelekos,G.,Jin,L.,et al.(2021a).Diagnostic accuracy ofapoint-of-care aMMP-8 test in the discrimination of periodontal health anddisease.J Clin Periodontol,48(8),1051-1065.doi:10.1111 / jcpe.13485

[0198] Deng,K.,Pelekos,G.,Jin,L.,et al.(2021b).Diagnostic accuracy ofself-reported measures of periodontal disease:A clinical validation study usingthe2017 case definitions.J Clin Periodontol,48(8),1037-1050.doi:10.1111 / jcpe.13484

[0199] Deng,K.,Pelekos,G.,Jin,L.,et al.(2021c).Gingival bleeding onbrushingas a sentinel sign of gingival inflammation:A diagnostic accuracytrial for thediscrimination of periodontal health and disease.J ClinPeriodontol,48(12),1537-1548.doi:10.1111 / jcpe.13545

[0200] Deng,K.,Wei,S.,Xu,M.,et al.(2022).Diagnostic accuracy of activematrixmetalloproteinase-8 point-of-care test for the discrimination ofperiodontal healthstatus:Comparison of saliva and oral rinse samples.JPeriodontal Res.doi:10.1111 / jre.12999

[0201] Eke,P.I.,Dye,B.A.,Wei,L.,et al.(2013).Self-reported measuresforsurveillance of periodontitis.J Dent Res,92(11),1041-1047.doi:10.1177 / 0022034513505621

[0202] Eke,P.I.,&Genco,R.J.(2007).CDC Periodontal Disease SurveillanceProject:Background,Objectives,and Progress Report.Journal of Periodontology,78(7s),1366-1371.doi:10.1902 / jop.2007.070134

[0203] Grant,M.M.,Taylor,J.J.,Jaedicke,K.,et al.(2022).Discovery,validation,and diagnostic ability of multiple protein-based biomarkers in saliva andgingivalcrevicular fluid to distinguish between health and periodontaldiseases.J ClinPeriodontol,49(7),622-632.doi:10.1111 / jcpe.13630

[0204] Gürsoy,U.K.,&Kantarci,A.(2022).Molecular biomarker researchinperiodontology:A roadmap for translation of science to clinical assayvalidation.J Clin Periodontol,49(6),556-561.doi:10.1111 / jcpe.13617

[0205] Hamp,S.E.,Nyman,S.,&Lindhe,J.(1975).Periodontal treatment ofmultirootedteeth.Results after 5 years.J Clin Periodontol,2(3),126-135.doi:10.1111 / j.1600-051x.1975.tb01734.x

[0206] Lorena,A.C.,de Carvalho,A.C.P.L.F.,&Gama,J.M.P.(2009).A reviewon thecombination of binary classifiers in multiclassproblems.ArtificialIntelligence Review,30(1),19.doi:10.1007 / s10462-009-9114-9

[0207] Maity,N.G.,&Das,S.(2017,4-11 March 2017).Machine learning forimproveddiagnosis and prognosis in healthcare.Paper presented at the 2017IEEE AerospaceConference.

[0208] Mundfrom,D.J.,Shaw,D.G.,&Ke,T.L.(2005).Minimum Sample SizeRecommendations for Conducting Factor Analyses.International Journal ofTesting,5(2),159-168.doi:10.1207 / s15327574ijt0502_4

[0209] Nelson,D.E.,Holtzman,D.,Bolen,J.,et al.(2001).Reliability andvalidity of measures from the Behavioral Risk Factor Surveillance System(BRFSS).Soz Praventivmed,46Suppl 1,S3-42.

[0210] Papapanou,P.N.,Sanz,M.,Buduneli,N.,et al.(2018).Periodontitis:Consensus report of workgroup 2of the 2017World Workshop on theClassification of Periodontal and Peri-Implant Diseases and Conditions.JPeriodontol,89Suppl 1,S173-s182.doi:10.1002 / jper.17-0721

[0211] Romano,F.,Perotto,S.,Bianco,L.,et al.(2020).Self-Perception ofPeriodontal Health and Associated Factors:A Cross-Sectional Population-BasedStudy.Int J Environ Res Public Health,17(8).doi:10.3390 / ijerph17082758

[0212] Sajda,P.(2006).Machine learning for detection and diagnosis ofdisease.Annu Rev Biomed Eng,8,537-565.doi:10.1146 / annurev.bioeng.8.061505.095802

[0213] Shin,H.,Schneeweiss,S.,Glynn,R.J.,et al.(2022).CardiovascularOutcomes in Patients Initiating First-Line Treatment of Type 2Diabetes WithSodium-Glucose Cotransporter-2Inhibitors Versus Metformin:A Cohort Study.AnnIntern Med,175(7),927-937.doi:10.7326 / m21-4012

[0214] Sorsa,T.,Gursoy,U.K.,Nwhator,S.,et al.(2016).Analysis of matrixmetalloproteinases,especially MMP-8,in gingival creviclular fluid,mouthrinseand saliva for monitoring periodontal diseases.Periodontol 2000,70(1),142-163.doi:10.1111 / prd.12101

[0215] Swets,J.A.(1988).Measuring the accuracy of diagnosticsystems.Science,240(4857),1285-1293.doi:10.1126 / science.3287615

[0216] Tonetti,M.S.,Chapple,I.L.,Jepsen,S.,et al.(2015).Primary andsecondary prevention of periodontal and peri-implant diseases:Introductionto,and objectives of the 11th European Workshop on Periodontology consensusconference.J Clin Periodontol,42Suppl 16,S1-4.doi:10.1111 / jcpe.12382

[0217] Tonetti,M.S.,Deng,K.,Christiansen,A.,et al.(2020).Self-reportedbleeding on brushing as a predictor of bleeding on probing:Early observationsfrom the deployment of an internet of things network of intelligent power-driven toothbrushes in a supportive periodontal care population.J ClinPeriodontol,47(10),1219-1226.doi:10.1111 / jcpe.13351

[0218] Tonetti,M.S.,Greenwell,H.,&Kornman,K.S.(2018).Staging and grading ofperiodontitis:Framework and proposal of a new classification and casedefinition.J Clin Periodontol,45Suppl 20,S149-S161.doi:10.1111 / jcpe.12945

[0219] Tonetti,M.S.,Jepsen,S.,Jin,L.,et al.(2017).Impact of the globalburden of periodontal diseases on health,nutrition and wellbeing of mankind:Acall for global action.J Clin Periodontol,44(5),456-462.doi:10.1111 / jcpe.12732

[0220] Trombelli,L.,Farina,R.,Silva,CO,et al.(2018).Plaque-inducedgingivitis:Case definition and diagnostic considerations.J Clin Periodontol,45Suppl 20,S44-S67.doi:10.1111 / jcpe.12939

[0221] The protection content of the present invention is not limited to the above embodiments. Without departing from the spirit and scope of the present invention, changes and advantages that can be thought of by those skilled in the art are included in the present invention and are protected by the attached claims.

Claims

1. A periodontal health classification system that can be used for self-screening, characterized in that: The periodontal health classification system includes a classification model, which is based on important features that can distinguish periodontal health status, combines and weights the important features to obtain a periodontal health status classification; the periodontal health status includes periodontal health, gingivitis, and periodontitis; the periodontitis is divided into stage I periodontitis, stage II periodontitis, stage III periodontitis, and stage IV periodontitis.

2. The periodontal health classification system according to claim 1, characterized in that: The classification model includes one or more logistic regression models and / or one or more random forest models; The logistic regression model was used to preliminarily evaluate periodontal health status, and evaluated "periodontal health", "gingivitis", "periodontitis", "stage I / II periodontitis" and "stage III / IV periodontitis" respectively, and screened the predictors to determine the predictors that were important for the prediction results; The random forest model is used to perform multi-classification on subjects, subject groups and / or populations presenting different periodontal health disease spectra; the multi-classification includes three categories of "periodontal health", "gingivitis" and "periodontitis", six categories of "periodontal health", "gingivitis", "stage I periodontitis", "stage II periodontitis", "stage III periodontitis" and "stage IV periodontitis", and three categories of "periodontal health", "gingivitis and stage I periodontitis" and "stages II-IV periodontitis".

3. The periodontal health classification system according to claim 2, characterized in that: The predictors before screening are obtained through self-reported measurements collected through questionnaires, disease symptoms and signs, and self-test results, including tooth mobility, tooth displacement, tooth elongation, tooth and gum appearance, presence of gingival bleeding, gum brushing bleeding test, presence and / or quantitative assessment of health / disease biomarkers in body fluids or exhaled breath; the tooth displacement includes tooth migration or drift; The periodontal health disease spectrum includes periodontal health, gingivitis, and periodontitis; the periodontal health refers to the absence of inflammation and periodontal tissue destruction, the gingivitis refers to the presence of gingival inflammation but no periodontal tissue destruction characteristics and / or health / disease biomarkers, and the periodontitis refers to the presence of periodontal tissue destruction characteristics and / or health / disease biomarkers.

4. The periodontal health classification system according to claim 2, characterized in that: The data used in the logistic regression model include data obtained from questionnaires, active matrix metalloproteinase-8 lateral flow immunochromatographic rapid test and gingival brushing bleeding test, as well as demographic data, lifestyle factors, and personal medical history; the gingival brushing bleeding test includes self-reported gingival bleeding during brushing, salivary hemoglobin concentration and content; the demographic data include age and gender; the lifestyle factors include smoking status; and the personal medical history includes whether one suffers from cardiovascular disease and / or diabetes.

5. The periodontal health classification system according to claim 2, characterized in that: The logistic regression model incorporates predictors by backward stepwise selection, and the parameters associated with each predictor are subjected to Wald test. If they are not significant at 95% confidence level, they are deleted, the model is rerun and the process is repeated until no more predictors are deleted, and the screened predictors that are important for predicting the results are obtained; According to the predicted probability of the logistic regression model, the receiving operating characteristic ROC curve, the area under the ROC curve, the sensitivity, the specificity, the positive predictive value PPV and the negative predictive value NPV are estimated; the predicted probability critical value refers to the point that maximizes the sum of the sensitivity and specificity of the entire ROC curve; when the predicted probability is higher than the predicted probability critical value in each logistic regression model, it belongs to the corresponding classification in each logistic regression model; and / or, The evaluation formula of the logistic regression model is as follows: Fitted linear predictor of estimated logarithm of periodontal health = -1.750*Q2+0.844*Q3a+0.951*Q7-0.027*total Hb-1.471*aMMP-8POCT-0.050*age+2.112; Fitted linear predictor of estimated log gingivitis = -0.992*Q3a+1.396*self-reported GBoB-0.160*age-14.719; Fitted linear predictor of estimated logarithm of periodontitis = 1.407*Q2+1.845*aMMP-8POCT+0.102*age-4.203; Estimated fitted linear predictor of the logarithm of stage I / II periodontitis = -1.540*Q4+1.022*aMMP-8POCT+0.521*sex-0.963; and / or, The fitted linear predictor of the logarithm of stage III / IV periodontitis was estimated = 1.729*Q2+1.756*Q6+0.115*age+2.453*smoking-8.

024.

6. The periodontal health classification system according to claim 2, wherein: In the generation of the random forest model, each forest uses multiple estimators and sets a maximum depth value, and node segmentation is determined using the Gini coefficient; the random forest model is validated by running multiple independent trainings on the entire data set and comparing the results to ensure the internal consistency of the model; and / or, The random forest model was fitted using Python's Scikit-learn package, and the Yellowbrick package was used to calculate the micro- and macro-averaged multi-class AUROC values ​​and the AUROC for each class; the fitting results reported the mean and deviation of the model performance, including sensitivity, specificity, positive predictive value PPV, negative predictive value NPV, and accuracy; Among them, sensitivity and specificity are defined as low, medium or high; low refers to sensitivity or specificity <60%, medium refers to sensitivity or specificity 60-79%, and high refers to sensitivity or specificity ≥80%; The accuracy results derived from the AUROC values ​​were interpreted as low, medium, and high, respectively; low refers to an accuracy of 0.50-0.70, medium refers to an accuracy of 0.71-0.90, and high refers to an accuracy >0.90; Accuracy was calculated according to the following definition: Accuracy = Sensitivity * Prevalence + Specificity * (1-Prevalence).

7. The periodontal health classification system according to claim 1, wherein: The classification system also includes a data input module and a data output module; The data input module is used to input the acquired non-clinical parameter data into the logistic regression model and / or random forest model; The data output module is used to output the classification results generated by the logistic regression model and / or the random forest model.

8. The periodontal health classification system according to claim 7, characterized in that: The non-clinical parameter data used for the random forest model include: "Q1 self-reported presence or absence of periodontitis", "Q2 self-rating of gum / tooth health", "Q3a tooth cleaning history", reported symptoms of "Q4 loose teeth", active matrix metalloproteinase-8 lateral flow immunochromatographic rapid test, self-reported gingival bleeding during brushing, hemoglobin Hb concentration and age; the detection sources of the hemoglobin concentration include saliva, toothpaste mixture, and tooth brushing pulp.

9. A method for constructing a periodontal health classification system according to any one of claims 1 to 8, characterized in that: The construction method comprises the following steps: Step 1: Collect clinical data on periodontal health status classification as training and validation sets for the model; Step 2: Collect non-clinical parameter data related to the periodontal health disease spectrum and pre-process the data; Step 3: Collect biological fluid and / or exhaled gas samples, process the samples and quantify the biomarkers therein; Step 4: Build a logistic regression model and a random forest model; Step 5: Train and optimize the constructed model to obtain the classification system that can accurately classify periodontal health conditions.

10. The construction method according to claim 9, characterized in that: In the step 1, the clinical data collected for the classification of periodontal health status are confirmed by internationally recognized case definitions and used as the classification gold standard; during the clinical data collection process, quality control is performed during the data collection and processing process to ensure data reliability; the quality control includes training of clinical parameter collection personnel, calibration of high-level intraclass correlation coefficients of necessary clinical parameters, questionnaire validation, and biomarker test reliability validation evaluation; In the step 2, in the process of obtaining the non-clinical parameter data, when the subject answers the questions in the questionnaire, the answers are checked to ensure that they are within the expected reasonable range; In the step three, a rapid test method is used to test whether the biomarkers in the sample can be detected by a preset minimum detection limit; In the step 4, during the model building process, the model building meets a series of goals: i) high significance of the overall model; ii) Eliminate predictors that do not contribute significantly to the overall performance; iii) Achieve a high level of diagnostic accuracy for binary and / or multi-class predictions and reduce clinically relevant errors and evaluate the errors of the model by calculating positive predictive value, negative predictive value, specificity, sensitivity, overall accuracy and area under the receiver operating characteristic curve (AUROC) and constructing a confusion matrix; The high level of diagnostic accuracy refers to an accuracy > 0.90; and / or, In step five, during the training and optimization process of the model, by providing specific weights in the analysis, the model is trained and adjusted based on the results of the confusion matrix to reduce the probability of error.

11. The construction method according to claim 9, characterized in that: The specific steps of constructing the logistic regression model are: Step 1: Input the training data into the logistic regression model to be optimized, check the distribution pattern of the data in the model, and select the optimization analysis method; Step 2, selecting prediction conditions, wherein the prediction conditions include healthy and gingivitis and periodontitis, gingivitis and healthy and periodontitis, and periodontitis and healthy and gingivitis; Step 3: Select the prediction probability cutoff value, or use the standard receiver operating curve to select the classification result corresponding to the logistic regression model; Step 4: Place the predictors into the logistic regression model, run standard model diagnostic tests, and use backward stepwise selection to remove predictors that do not significantly affect the results to obtain the final logistic regression model. Step 5: Calculate the diagnostic accuracy parameters of the final model to evaluate whether it meets the performance criteria and determine the parameters of the multivariate prediction equation, and use the bootstrap method to enhance the stability of the model; the diagnostic accuracy parameters include sensitivity, specificity, positive and negative predictive values, and overall accuracy.

12. The construction method according to claim 9, characterized in that: The construction of the random forest model includes the following steps: Step a: input the training data into the random forest model to be optimized, check the distribution pattern of the data in the model, and select the optimization analysis method; Step b, selecting a prediction type, the prediction type including three categories: 1) periodontal health, gingivitis, and periodontitis; 2) periodontal health, gingivitis, periodontitis stage I, periodontitis stage II, periodontitis stage III, or periodontitis stage IV; 3) periodontal health, gingivitis / stage I periodontitis, stage II-IV periodontitis; Step c: Use the important predictive factors obtained from the logistic regression analysis as input data to train and optimize the random forest model; use multiple estimators for each forest, set the maximum depth value, and run multiple independent trainings on the entire dataset and compare the results to ensure the internal consistency of the model.

13. The construction method according to claim 9, characterized in that: In step c, during the training process of optimizing the random forest model, constructing a feature importance map to establish the relative weight of each predictor included in the random forest model; The stability of the model was tested by assessing the impact of excluding one or more predictors from the model in multiple iterations until the model was stable and included all significant predictors.

14. A method for classifying periodontal health status, characterized in that: The classification method comprises the following steps: Step (i) collecting non-clinical parameter data related to periodontal health and preprocessing the data; Step (ii), constructing logistic regression model and random forest model; Step (iii) using the constructed logistic regression model and random forest model to evaluate and classify periodontal health status.

15. The classification method according to claim 14, characterized in that: The non-clinical parameter data sources include three consecutive screening tests and a full-mouth periodontal examination; The screening test further includes a questionnaire, a rapid flow immunochromatographic test for active matrix metalloproteinase-8, and a gingival brushing bleeding test; The contents of the questionnaire included "Q1 Self-assessment of gum disease", "Q2 Evaluation of gum / tooth health", "Q3a Self-reported experience of supragingival scaling", "Q3b Deep teeth cleaning", "Q4 Self-perceived tooth movement", "Q5 Self-reported professionally diagnosed bone loss", "Q6 Self-perceived changes in tooth appearance", "Q7 Use of dental floss" and "Q8 Mouthwash".

16. An application, characterized in that: The application includes the application of the classification system as described in any one of claims 1-8, or the classification method as described in any one of claims 14 or claim 15 in the classification of periodontal health status, and the application in the prediction of periodontal health status; including the application of the logistic regression model and / or random forest model constructed by the method as described in any one of claims 9-13 in the classification of periodontal health status.

Images