Non-invasive periodontal screening using self-reported-oral-health (SROH) questionnaire and salivary biomarkers: development and validation of machine learning models Authors: Jan Yang Ho, Ainol Haniza Kherul Anuwar, Lee Peng Karen-Ng, Abd Aziz Alias, Siti Lailatul Akmar Zainuddin, Mohammad Tariqur Rahman, Nor Shafina Mohamed Nazari, Nur Falihin Soehairq Kamal Nurzaman, Nor Adinar Baharuddin Published: 2026 ABSTRACT AbstractBackgroundAccurate and accessible screening tools for periodontitis are essential for early detection and disease prevention. This study evaluated a non-invasive diagnostic approach integrating sociodemographic data, self-reported oral health (SROH) questionnaires, and salivary biomarkers, using both conventional statistical and machine learning (ML) predictive models.MethodsSeventy-seven adults completed a validated SROH questionnaire and provided saliva samples for quantification of six biomarkers: interleukins (IL-1β, IL-6), tumour necrosis factor (TNF-α), matrix metalloproteinases (MMP-8, MMP-9), and metallothionein (MT). Participants were clinically classified as having (i) periodontal health, (ii) gingivitis, or (iii) periodontitis. Predictive models were developed using Logistic Regression (LR), Random Forest (RF), and Naive Bayes (NB) across three feature sets: (i) SROH, salivary biomarkers and sociodemographic, (ii) SROH and salivary biomarkers, (iii) SROH and sociodemographic and (v) SROH alone. Model performance was assessed using 10-fold cross-validation and standard evaluation metrics.ResultsThe RF model trained on SROH, and salivary biomarkers achieved the highest accuracy with area under the receiver operating characteristic curve (AUC = 0.856), with superior precision (70.13%), sensitivity (0.701) and lower error rates (RMSE = 0.371) compared with NB (AUC = 0.795) and LR (AUC = 0.724) models in detecting periodontitis.ConclusionsThis non-invasive, SROH and biomarker-integrated approach shows potential as a first-line screening and referral tool in primary care and population-based settings where comprehensive periodontal examination is not routinely available. Further validation in larger, more diverse populations is warranted to support clinical translation.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12903-026-08759-9.Keywords: Non-invasive, Periodontitis Screening, Machine Learning, Salivary Biomarkers, Self-Reported Oral Health Background Periodontal diseases are complex, chronic inflammatory diseases of the supporting periodontal tissues, characterised by bacterial dysbiosis and a dysregulated host immune-inflammatory response [1]. According to the Global Burden of Disease Study 2021, over one billion people worldwide were affected by severe periodontitis, with a global age-standardised prevalence of 12.5% [2]. If left untreated, periodontal diseases can lead to tooth loss, masticatory dysfunction, aesthetic impairment, and reduced oral health-related quality of life (OHRQoL) [3]. They also pose systemic risks, with strong epidemiological evidence linking periodontitis to diabetes mellitus (DM), atherosclerotic cardiovascular disease, adverse pregnancy outcomes, and other chronic conditions [4–11]. These associations highlight the need for timely and accurate screening and diagnosis of periodontal diseases. Despite advances in periodontology, periodontal screening and diagnosis still rely primarily on conventional clinical periodontal examinations and radiographic assessments. These methods are time-consuming, operator-dependent, and resource-intensive, limiting their feasibility for large-scale screening. Even simplified indices, like the Basic Periodontal Examination (BPE) and Community Periodontal Index of Treatment Needs (CPITN), require standardised probes, clinical settings, and trained personnel [12, 13]. Consequently, effective population-level screening approaches are needed. In line with this, researchers are investigating alternative, scalable tools for periodontitis detection, notably salivary biomarkers and self-reported oral health questionnaires. Self-reported oral health (SROH) tool has been proposed for surveillance and epidemiological studies [14–18]. The original eight-item SROH questionnaire developed by Eke and Genco [16] was validated against full-mouth periodontal examination in a nationally representative U.S. sample. Since then, adapted versions of the SROH have been evaluated in various populations, including Portugal [19], Spain [20, 21], Japan [22], New Zealand [23], Brazil [24], Hong Kong [25], and France [26]. Recent exploratory work by the authors’ research group has further examined its applicability in Malaysia, contributing to the growing evidence supporting its cross-cultural utility [27]. While the SROH offers a practical and low-cost alternative for large-scale screening, its predictive accuracy can be enhanced when combined with other indicators, such as salivary biomarkers. Salivary biomarkers provide real-time insight into host-mediated tissue destruction and can therefore improve the detection of active disease processes that subjective questionnaires alone may overlook. Cytokines such as interleukin (IL)-1β and IL-6 reflect key upstream inflammatory pathways [28], while matrix metalloproteinase (MMP)-8 and MMP-9 indicate collagen degradation and connective tissue breakdown [29]. Incorporating these objective biochemical signals enhances clinical relevance by potentially identifying disease earlier, supporting targeted referral, and informing risk-based prevention strategies. Furthermore, saliva-based assays are increasingly cost-efficient, rapid and chair-side compatible, representing a feasible future point-of-care approach in public health settings where full periodontal examination is impractical. Recent study suggested that combining self-reported data with sociodemographic variables and salivary biomarkers can improve diagnostic performance in predictive models for periodontitis [30]. However, evidence on such integrative, multivariable approaches remains limited. Most existing models rely on conventional logistic regression, which may not adequately capture complex, non-linear relationships among predictors. Artificial intelligence (AI), particularly Machine learning (ML), are increasingly integrated to medical and dental research to assist in diagnosis, as well as providing advanced tools for healthcare data analysis. This study systematically evaluates the predictive performance of logistic regression in comparison with two widely adopted machine learning algorithms: Random Forest (RF) and Naive Bayes (NB). Both RF and NB have demonstrated broad applicability across medical domains, including periodontal research [31, 32]. The RF algorithm is a supervised learning method well-suited for classification and regression tasks, employing an ensemble of decision trees constructed on random subsets of input data to enhance predictive accuracy and generalisability. Its robustness and capacity to capture complex feature interactions make it particularly appropriate for biomedical datasets. In contrast, the NB classifier is a simple, yet effective probabilistic model based on Bayes’ theorem, which assumes conditional independence among predictors to estimate class probabilities. Its ability to generate probabilistic outputs enhances model transparency and supports clinical decision. This study aimed to evaluate the performance of predictive models incorporating sociodemographic factors, SROH questionnaire responses, and salivary biomarkers for screening clinically confirmed periodontitis, using both conventional LR and ML approaches. By adopting this integrative modelling strategy, the study seeks to advance the development of efficient, non-invasive screening tools suitable for population-level surveillance and primary care application. In this context, screening refers to the identification of individuals at increased risk of irreversible periodontal tissue destruction who would benefit from comprehensive periodontal assessment. The proposed models are not intended to replace clinical diagnosis, but to support early triage and referral in settings where clinical and radiographic assessments are impractical. Machine learning approaches were selected to address the complex, non-linear relationship between subjective symptoms and biochemical markers that are unlikely to be adequately captured using conventional regression-based models. To the best of our knowledge, this is the first study in a Malaysian population to compare the diagnostic performance of ML models (specifically RF and NB) with LR, using a combined dataset of SROH questionnaire responses and salivary biomarkers for periodontitis screening. Methods This cross-sectional study was conducted in three sequential parts: (i) collection of self-reported oral health data using a validated SROH questionnaire; (ii) salivary biomarkers measurement and clinical periodontal examination; and (iii) development and evaluation of predictive models for periodontitis screening. A total of 77 participants were recruited through convenience sampling from the Postgraduate Periodontology Clinic, Universiti Malaya (January 2023 – May 2024). This sample size satisfied minimum requirements for both 8-item questionnaire validation (n = 40; 1 item to 5 participants ratio) and salivary biomarkers analysis (n ≥ 51; effect size = 0.45, power = 80%, α = 0.05) [33–35]. Eligible participants were adults aged ≥ 18 years, systemically healthy, and with at least 20 permanent teeth. Exclusion criteria included a history of periodontal treatment within the past year, systemic illness, pregnancy, recent medication use, or tobacco use within the past five years. This study was approved by the Faculty of Dentistry Medical Ethics Committee, Faculty of Dentistry, Universiti Malaya (Approval No: DF RD2013/0064/2333/23105 (P). All procedures were performed in accordance with the International Conference of Harmonisation - Guidelines for Good Clinical Practice (ICH-GCP) and Helsinki Declaration. Written informed consent was obtained from all participants prior to participation in the study. Part I: questionnaire adaptation and validation Self-reported data were collected using a locally adapted and validated version of the SROH questionnaire originally developed by Eke and Genco [16]. The adapted tool comprised of nine items assessing key indicators of periodontal health, including self-perceived gum disease, bleeding gums, tooth mobility, tooth loss, bone loss, dental appearance, and oral hygiene behaviours such as the use of dental floss and mouthwash. Psychometric evaluation demonstrated strong measurement properties, including content validity (S-CVI/UA = 1), face validity (S-FVI/Ave = 0.840), internal consistency (Cronbach alpha = 0.670), test-retest reliability (ICC = 0.980), and concurrent validity (AUCC = 0.932). The validated SROH questionnaire was administered electronically via Google Form, accessible through a QR code. Part II: salivary biomarkers measurement and clinical periodontal examination Saliva sample collection Participants who completed the SROH questionnaire were also asked to provide saliva samples for protein biomarkers quantification. They were instructed to refrain from toothbrushing, food or fluid intake (except water), and gum-chewing for at least an hour prior to saliva collection. To minimise diurnal variation and pre-analytical influences from diet, oral hygiene, or physical activity, saliva collection was standardised to morning sessions between 9 to 11am. Unstimulated whole saliva was collected using the passive drool method. Participants allowed saliva to accumulate at the floor of the mouth and then tilted their heads downward so that saliva could be drawn off from the lower lip. A sterile funnel was attached to the opening of a 15 ml Falcon tube, into which approximately 5 ml of saliva was collected. Samples were centrifuged at 4,000 relative centrifugal force (RCF) for 20 min at 4 °C. The supernatant was aliquoted into 1 ml microcentrifuge tubes and stored in -80 °C until further analysis, for a maximum of six months after collection. Enzyme-linked immunosorbent assay (ELISA) A sandwich ELISA method was utilised to quantify salivary protein biomarkers (IL-6, IL-1β, TNF-α, MMP-8, MMP-9 and MT) in accordance with the manufacturer’s protocol. Detection limits for the Elabscience Human ELISA kits (Elabscience®, Texas, USA) were 1.56–100 pg/ml for IL-6, 7.81–500 pg/ml for IL-1β and TNF-α, and 0.16-10ng/ml for MMP-8 and MMP-9. The detection range for the Cusabio Human ELISA kit (Cusabio®, Wuhan, China) for MT was 6.25–400 pg/ml. Periodontal examination Clinical periodontal assessments were performed by a calibrated examiner (Jan Yang Ho). Full-mouth examinations included probing pocket depth (PPD), clinical attachment loss (CAL), bleeding on probing (BOP), and plaque presence, recorded at six sites per tooth for all present teeth [36–37]. Examiner calibration was conducted prior to data collection. Inter- and intra-examiner reliability for PPD and CAL measurements was assessed using the intraclass correlation coefficient (ICC), with a ± 1 mm tolerance [38]. Inter-examiner ICC values were 0.91 (95% CI: 0.88–0.94) for PPD and 0.80 (95% CI: 0.71–0.86) for CAL. Intra-examiner ICC values were 0.82 (95% CI: 0.72–0.89) for PPD and 0.91 (95% CI: 0.85 − 0.94) for CAL, indicating moderate to excellent agreement [39]. Participants were classified as having periodontal health, gingivitis, or periodontitis according to the 2017 World Workshop classification criteria [40–42]. Part III: statistical analysis Descriptive and inferential statistics Analyses were conducted using SPSS version 29. Data normality was assessed using the Kolmogorov–Smirnov test (n > 50). Continuous variables were presented as mean ± standard deviation (SD) or median (interquartile range, IQR), while categorical variables were presented as frequencies and percentages. Group comparisons were performed using ANOVA with Tukey’s post-hoc test or the Kruskal-Wallis test with Dunn’s post-hoc analysis. Correlations were examined using Pearson’s or Spearman’s methods and associations between categorical variables were tested using the Chi-square or Fisher’s Exact tests. Prediction of periodontitis using machine learning (ML) algorithms For the development and evaluation of predictive models for periodontitis screening, both conventional and ML approaches were employed. Logistic regression, a traditional statistical method, was compared with two ML algorithms: RF and NB. Model development and optimisation were performed using the Waikato Environmental for Knowledge Analysis (WEKA) platform. The objective was to explore and enhance predictive performance varying data inputs. Feature sets Three distinct feature sets were used to train and test the models: i.SROH, salivary biomarkers and sociodemographic. SROH, salivary biomarkers and sociodemographic. ii.SROH and sociodemographic. SROH and sociodemographic. iii.SROH and salivary biomarkers. SROH and salivary biomarkers. iv.SROH only. SROH only. Periodontal clinical measures were excluded from model inputs because their assessment requires invasive clinical examination. These measures were used solely to establish periodontal disease status and were not included as predictor variables. Accordingly, all combined models were intentionally designed to rely exclusively on non‑invasive inputs, including SROH measures and biomarkers, to support feasibility in screening‑based applications. SROH items were recoded as follows: self‑reported gum disease (yes vs. no/don’t know), self‑rated teeth and gum health (poor/fair vs. good/very good/excellent), bleeding gums (sometimes/fairly often/very often vs. never/hardly ever), history of gum treatment (yes vs. no/don’t know), loose teeth (yes vs. no/don’t know), self‑reported bone loss around teeth (yes vs. no/don’t know), concern about teeth appearance (yes vs. no/don’t know), floss use (never vs. 1–7 days per week), and mouthwash use (never vs. 1–7 days per week). Model training and evaluation The primary analysis focused on binary classification (periodontitis vs. non-periodontitis) to reflect the intended screening application, while multiclass classification (health, gingivitis, periodontitis) was explored as a secondary analysis to assess disease differentiation. Model performance was evaluated using an internal validation framework consistent with the exploratory aims of this study. Data pre-processing was performed within each training fold and included handling missing values, normalisation of numerical variables, and encoding of categorical variables. Feature selection was conducted exclusively within the training portion of each fold as part of the model‑building pipeline to ensure that no information from the held‑out data was used during feature selection or model fitting. The most frequently selected features across cross‑validation folds are reported. Predictive models were trained and evaluated using a 10‑fold cross‑validation approach to maximise efficient use of the available data and to reduce overfitting in the context of a modest sample size (n = 77). This procedure was applied consistently across all models, including conventional logistic regression and machine‑learning algorithms (RF and NB). In each fold, models were trained on nine partitions of the data and evaluated on the remaining held‑out partition. Model performance was assessed using multiple metrics, including accuracy, precision, recall, F1‑score, and the area under the receiver operating characteristic curve (AUC). Sensitivity and specificity were additionally reported to assess potential utility for screening applications. All reported metrics represent internally cross‑validated performance estimates derived from the 10‑fold validation procedure. Results Characteristics of the study population The sociodemographic and clinical characteristics of the study participants are presented in Table 1. The mean age of the periodontitis group was significantly higher than that of the periodontal health and gingivitis groups. A higher proportion of females were observed in the periodontal health and gingivitis groups, whereas the periodontitis group showed an almost equal gender distribution. Differences in ethnicity and education level across the three groups were not statistically significant (P > 0.05). Clinically, participants on the periodontitis group had fewer teeth, deeper PPD, greater CAL, and higher number of bleeding sites compared with both the periodontal health and gingivitis groups (P < 0.05). Table 1Sociodemographic profile and clinical periodontal characteristics of participants stratified by periodontal statusSociodemographic (DI)OverallPeriodontal Status, n (%)p -valuePeriodontal health26 (33.8)Gingivitis25 (32.5)Periodontitis26 (33.8)n ( %)77 (100.0)26 (33.8)25 (32.5)26 (33.8)0.987‡Age (mean ± SD)37.60 ± 10.1432.23 ± 8.67a35.96 ± 9.33a44.54 ± 8.44b< 0.001†Gender, n (%)0.014‡ Male20 (26.0)3 (11.5)5 (20.0)12 (46.2) Female57 (74.0)23 (88.5)20 (80.0)14 (53.8)Ethnicity n (%)0.849‡ Malay44 (57.1)13 (50.0)17 (68.0)14 (53.8) Chinese27 (35.1)11 (42.3)7 (28.0)9 (34.6) Indian6 (7.8)2 (7.7)1 (4.0)3 (11.5) OthersEducation level, n (%)0.281 Low21 (27.3)6 (23.1)5 (20.0)10 (38.5) High56 (72.7)20 (76.9)20 (80.0)16 (61.5)Clinical parameters (mean ± SD) Number of teeth26.01 ± 3.3826.96 ± 2.29a26.88 ± 1.79a24.23 ± 4.62b0.001† Probing pocket depth (mm)3.03 ± 1.471.86 ± 0.24a2.19 ± 0.22b5.01 ± 0.57c< 0.001† Clinical attachment level (mm)1.31 ± 1.950.03 ± 0.06a0.07 ± 0.18a3.78 ± 1.41b< 0.001† Bleeding on probing (%)25.51 ± 23.924 ± 2.05a24.35 ± 12.49b48.15 ± 23.89c< 0.001† Plaque score (%)30.96 ± 20.0713.74 ± 10.13a33.98 ± 15.32b45.29 ± 19.14b< 0.001††Kruskal-Wallis test ‡ Chi square testDifferent alphabets denoted statistically significant differences between groups Sociodemographic profile and clinical periodontal characteristics of participants stratified by periodontal status †Kruskal-Wallis test ‡ Chi square test Different alphabets denoted statistically significant differences between groups Overview of SROH questionnaire responses and salivary biomarker concentrations Among the 77 respondents to the SROH questionnaire (Table 2), 41.6% reported a history of gum disease. Slightly more than half (54.5%) rated their overall teeth and gum health as good to excellent, with 45.5% described it as fair or poor. Frequent gum bleeding was reported by 53.2% of participants, whereas 46.8% indicated they never experienced bleeding. Only 13.0% had received professional gum treatment. Approximately one in five reported loose teeth (19.5%), bone loss (18.2%), or concerns about teeth appearance (20.8%), although majority was unaware of having these conditions. Table 2Distribution of self-reported oral health (SROH) responses across periodontal health, gingivitis, and periodontitis groupsSelf-reported oral health questionsResponsesOverall77(100.0)Periodontal Status, n (%)Periodontal health26 (33.8)Gingivitis25 (32.5)Periodontitis26 (33.8)Q1. Gum diseaseYes32 (41.6%)2 (7.7%)10 (40.0%)20 (76.9%)No33 (42.9%)22 (84.6%)8 (32.0%)3 (11.5%)Don’t know12 (15.6%)2 (7.7%)7 (28.0%)3 (11.5%)Refused0 (0%)0 (0%)0 (0%)0 (0%)Q2. Teeth gumExcellent1 (1.3%)1 (3.8%)0 (0%)0 (0%)healthVery good8 (10.4%)5 (19.2%)2 (8.0%)1 (3.8%)Good30 (39.0%)15 (57.7%)9 (36.0%)6 (23.1%)Fair21 (27.3%)3 (11.5%)10 (40.0%)8 (30.8%)Poor14 (18.2%)0 (0%)4 (16.0%)10 (38.5%)Don’t know3 (3.9%)2 (7.7%)0 (0%)1 (3.8%)Refused0 (0%)0 (0%)0 (0%)0 (0%)Q3. BleedinggumsNever22 (29.9%)12 (46.2%)8 (32.0%)3 (11.5%)Hardly ever13 (16.9%)7 (26.9%)3 (12.0%)3 (11.5%)Sometimes36 (46.8%)7 (26.9%)13 (52.0%)16 (61.5%)Fairly often3 (3.9%)0 (0%)0 (0%)3 (11.5%)Very often2 (2.6%)0 (0%)1 (4.0%)1 (3.8%)Q4. GumtreatmentYes10 (13.0%)4 (15.4%)3 (11.5%)3 (11.5%)No65 (84.4%)22 (84.6%)22 (84.6%)22 (84.6%)Don’t know2 (2.6%)0 (0%)1 (3.8%)1 (3.8%)Refused0 (0%)0 (0%)0 (0%)0 (0%)Q5. Loose teethYes15 (19.5%)0 (0%)1 (4.0%)14 (53.8%)No57 (74.0%)24 (92.3%)22 (88.0%)11 (42.3%)Don’t know5 (6.5%)2 (7.7%)2 (8.0%)1 (3.8%)Refused0 (0%)0 (0%)0 (0%)0 (0%)Q6. Bone lossYes14 (18.2%)1 (3.8%)4 (16.0%)9 (34.6%)No61 (79.2%)25 (96.2%)21 (84.0%)15 (57.7%)Don’t know2 (2.6%)0 (0%)0 (0%)2 (7.7%)Refused0 (0%)0 (0%)0 (0%)0 (0%)Q7. TeethYes16 (20.8%)0 (0%)2 (8.0%)14 (53.8%)appearanceNo57 (74%)26 (100%)20 (80.0%)11 (42.3%)Don’t know4 (5.2%)0 (0%)3 (12.0%)1 (3.8%)Refused0 (0%)0 (0%)0 (0%)0 (0%)Q8. Floss useNever28 (36.4%)8 (30.8%)8 (32.0%)12 (46.2%)1–7 days43 (55.8%)17 (65.4%)13 (52.0%)13 (50.0%)Missing/Refused6 (7.8%)1 (3.8%)4 (16.0%)1 (3.8%)Q9. MouthwashNever41 (53.2%)16 (61.5%)12 (48.0%)13 (50.0%)1–7 days31 (40.3%)9 (34.6%)10 (40.0%)12 (46.2%)Missing/Refused5 (6.5%)1 (3.8%)3 (12.0%)1 (3.8%) Distribution of self-reported oral health (SROH) responses across periodontal health, gingivitis, and periodontitis groups With respect to oral hygiene practices, 60.6% of respondents reported flossing at least once per week, and 56.9% reported weekly mouthwash use. The remaining seldom or never used these oral hygiene aids. Salivary biomarkers were quantified using ELISA, and their concentrations were compared across the periodontal health, gingivitis and periodontitis groups (Table 3). Table 3Comparison of salivary biomarker concentrations among periodontal health, gingivitis, and periodontitis groupsSalivary biomarkersOverallPeriodontal healthGingivitisPeriodontitisP-valueIL-1β (pg/mL)3.67 ± 5.910.22 ± 0.47a4.85 ± 6.62b5.98 ± 6.68b < 0.001 IL-6 (pg/mL)6.91 ± 12.732.09 ± 4.41a10.43 ± 19.21b8.35 ± 8.96b < 0.001 TNF-α (pg/mL)20.84 ± 57.4415.13 ± 33.7316.78 ± 40.5230.46 ± 84.670.092MMP-8 (ng/mL)119.09 ± 132.5639.91 ± 53.7a91.6 ± 78.32b224.73 ± 160.51c < 0.001 MMP-9 (ng/mL)187.34 ± 175.5475.91 ± 88.98a151.95 ± 113.26b332.8 ± 193.05c < 0.001 MT (pg/mL)1517.31 ± 1105.431327.56 ± 1045.181431.86 ± 1043.631789.23 ± 1206.310.458Data was presented as mean ± standard deviation. Intergroup differences were analysed with Kruskal Wallis test with post-hoc Dunn test. Different alphabets between groups indicated statistically significant differencesAbbreviation: IL-1β Interleukin-1 beta, IL-6 Interleukin-6, TNF-α Tumour necrosis factor-alpha, MMP-8 Matrix metalloproteinase-8, MMP-9 Matrix metalloproteinase-9, MT Metallothionein Comparison of salivary biomarker concentrations among periodontal health, gingivitis, and periodontitis groups Data was presented as mean ± standard deviation. Intergroup differences were analysed with Kruskal Wallis test with post-hoc Dunn test. Different alphabets between groups indicated statistically significant differences Abbreviation: IL-1β Interleukin-1 beta, IL-6 Interleukin-6, TNF-α Tumour necrosis factor-alpha, MMP-8 Matrix metalloproteinase-8, MMP-9 Matrix metalloproteinase-9, MT Metallothionein Levels of IL-1β and IL-6 showed a similar pattern, with significantly lower concentrations observed in the periodontal health group compared with both gingivitis and periodontitis groups (P < 0.001). Similarly, MMP-8 and MMP-9 concentrations were significantly higher in periodontitis patients compared with the gingivitis or periodontal health groups (P < 0.001). However, there was no significant differences were found among the three groups for TNF-α and MT (P > 0.05). Development predictive models for periodontitis To develop robust predictive models for periodontitis using ML algorithms (RF and NB), input variables were strategically selected to capture diverse and relevant aspects of disease risk. For modelling purposes, participants with periodontal health or gingivitis were grouped as non-periodontitis, while those with periodontitis were grouped as the comparator group, reflecting the irreversible tissue destruction characteristics of the disease. Three feature sets were summarised and assembled for analysis, drawing sociodemographic data (Table 1), SROH responses (Table 2) and salivary biomarker measurements (Table 3). Sociodemographic variables (for example, age, gender, ethnicity, education level) provided contextual and risk stratification factors. Self-reported oral health responses (e.g., history of gum disease, self-rated teeth/gum health, gum bleeding, gum treatment, loose teeth, bone loss, teeth appearance concerns, floss use, mouthwash use) captured individual perceptions and symptoms associated with periodontal status. Quantified salivary biomarker concentrations (e.g., IL-1β, IL-6, TNF-α, MMP-8, MMP-9, MT) contributed objective biochemical indicators of inflammation and disease progression. Classification performance The classification performance of the NB, RF and logistic regression algorithms using sociodemographic items (DI), SROH items (SI) and biomarker items (BI), either individually or in combination, is presented in Table 4. All models achieved an AUC above 0.6, with the highest AUC recorded by RF model trained on the combined SI + BI feature set (AUC: 0.856). Performance metrics, including ROC area, Kappa, precision, recall, F1-score, RMSE, and MAE, are summarised in Table 4. For clinical interpretability in a screening context, sensitivity and specificity were also considered. The RF model trained on SI + BI achieved a sensitivity of 0.701, indicating its ability to correctly identify periodontitis cases. Table 4Performance metrics of logistic regression, random forest, and Naive Bayes models across different feature sets for periodontitis classificationModelAlgorithmPerformance IndicatorROC AreaKappaPrecision(%)RecallF1-MeasureRMSE↓MAE↓DINaive Bayes0.6790.20146.750.4680.4670.4490.376Random Forest0.6550.22148.050.4810.4810.4820.358Logistic0.6880.27851.940.5190.5170.4510.279BINaive Bayes0.7510.37558.440.5840.5760.4920.279Random Forest0.7920.43462.330.6230.6190.4070.317Logistic0.7750.39559.740.5970.5920.4180.309SINaive Bayes0.7330.45363.630.6360.6240.4410.290Random Forest0.6650.24049.350.4940.5000.4780.362Logistic0.6830.37558.440.5840.5760.4530.343SI + DINaive Bayes0.7780.49266.230.6620.6530.4370.275Random Forest0.7630.35757.140.5730.5700.4220.356Logistic0.7440.37758.440.5890.5890.4570.299SI + BINaive Bayes0.7950.45161.040.6100.6050.4750.255Random Forest0.8560.55170.130.7010.6940.3710.292Logistic0.7240.35757.140.5170.5740.4770.297DI + BINaive Bayes0.6410.37658.440.5840.5810.4960.282Random Forest0.7670.39559.740.5970.5880.41830.334Logistic0.6890.31854.540.5450.5480.4920.333SI + DI+BINaive Bayes0.8010.43462.340.6230.6170.4740.254Random Forest0.8460.55270.130.7010.6970.3790.303Logistic0.6110.23949.350.4940.4810.56460.346DI Sociodemographic Items, SI SROH Items, BI Biomarker Items Performance metrics of logistic regression, random forest, and Naive Bayes models across different feature sets for periodontitis classification DI Sociodemographic Items, SI SROH Items, BI Biomarker Items Predictor contribution was evaluated using the same feature‑selection pipeline embedded within the 10‑fold cross‑validation framework as described in the Methodology section. Feature selection was performed independently within the training portion of each fold, and no single fixed predictor set was derived from the full dataset. Instead, feature relevance was summarised based on selection frequency across cross‑validation folds, providing an indication of predictor stability and robustness. Using this approach, several predictors were consistently retained across folds, including Q5, Q7, IL‑1β, and IL‑6 (selected in 100% of folds), as well as MMP8 and MMP9 (90%). Q1 also demonstrated high stability (80% of folds), whereas Q2 and Q3 were selected less frequently (40% and 30% of folds, respectively), suggesting more moderate and context‑dependent contributions. In contrast, predictors that were rarely or never selected (e.g. Q4, Q6, Q8, Q9, TNF‑α, and MT) contributed limited independent information and were excluded from most model iterations. The most frequently selected predictors are therefore reported to enhance transparency and interpretability of the results. Overall, RF models consistently outperformed both NB and logistic regression across different feature combinations. Models incorporating both SROH items and biomarker items demonstrated superior performance in terms of precision, RMSE, MAE and AUC compared with models using individual feature sets. Model comparison The performance of the RF models improved with the integration of SI and BI features, achieving an AUC of 0.856, precision of 70.13%, RMSE of 0.371, and MAE score of 0.292. Among all tested models, the RF trained on the SI + BI dataset demonstrated the strongest classification performance. The classification performance of the NB, RF, and logistic regression algorithms is further illustrated in the confusion matrices (Fig. 1). RF achieved the highest overall precision (70.13%), outperforming NB (61.04%) and logistic regression (57.14%). In addition to the primary binary classification, multiclass confusion matrices are presented (Fig. 1) as an exploratory analysis to evaluate the model’s ability to differentiate periodontal health states. Confusion matrix of algorithms of SROH plus biomarker Based on the ROC curve shown in Fig. 2, the RF algorithm demonstrated the best classification performance for predicting periodontitis, consistent with the results presented in Table 4. Superior accuracy and a larger area under the curve (AUC = 0.856) were observed when trained on the combined SROH and biomarker (SI + BI) features. ROC Curve of three types of algorithms of SROH plus biomarker Across all feature combinations, models integrating SROH responses with salivary biomarkers consistently outperformed those relying on questionnaire and sociodemographic variables alone. Discussion Combining SROH responses with salivary biomarkers enhanced model discrimination, highlighting the benefits of integrating subjective and objective indicators for early disease detection. Our results demonstrated that ML-based models, particularly RF, outperformed logistic regression, achieving an AUC of 0.856 compared to 0.775 for logistic regression. The highest AUC was attained by the RF model incorporating SI and BI, underscoring the strength of this combined approach. These findings support the potential for integrating patient-reported outcomes with biochemical indicators to improve early detection of periodontal disease using AI-driven models. This non-invasive approach is particularly valuable in primary care or resource-limited settings, facilitating earlier referral and targeted intervention. At a population level, these models may optimise resource allocation, reduce under-diagnosis, and mitigate the clinical and economic burden of advanced periodontitis. In the context of screening, sensitivity and specificity are particularly relevant, as they reflect the ability of the model to correctly identify individuals with disease while minimising false reassurance. Recent Asian evidence supports the use of self-reported periodontal measures for population-based screening. A multiethnic Singapore study showed that selected SROH items combined with demographic variables achieved moderate discrimination for severe periodontitis (AUC ≈ 0.76), with tooth mobility, age, and ethnicity as key predictors. Consistent with these findings, SROH variables contributed meaningfully to the present study, although population-specific validation remains essential due to variability across settings [43]. Previous SROH studies predominantly used logistic regression [26], which despite accommodating multiple predictors, it is limited by its binary outcome framework and may not capture the complexity of periodontal disease progression. Binary classification may oversimplify the spectrum of periodontal disease. In the present study, grouping periodontal health and gingivitis as non-periodontitis cohort, though methodologically necessary, that may have masked subtle differences in disease status. According to the latest classification scheme [40–42], periodontitis is stratified into stages (I - IV) and grades (A - C), reflecting disease severity and complexity. Logistic regression assumes a linear relationship between predictors and outcome, which may not hold in complex biological systems. In contrast, RF models, by capturing non-linear interactions, offer superior discrimination for periodontal disease classification. AI models have demonstrated promise in detecting dental conditions, often using dental radiographs or intraoral photographs [44]. Yet, integrating SROH questionnaires with salivary biomarkers in AI-based models remains underexplored. A study by Deng et al. [25] evaluated different tools for periodontitis detection in 408 subjects. They assessed the CDC/AAP self-reported questionnaire alone and found limited accuracy for early-stage disease (Stages I/II, AUC 0.608), but moderate-to-high accuracy for overall periodontitis and Stages III/IV (AUC 0.803–0.870). The CDC/AAP questionnaire shows limited utility for early detection of periodontitis, which aligns with the inherent difficulty of diagnosing the disease at an early stage, as clinically relevant symptoms typically emerge only in advanced stages. Adding age and smoking status to the responses substantially improved detection of advanced disease (sensitivity 95.7%; specificity 89.0%; AUC 0.953) though early-stage detection remained limited [25]. In a subsequent study, the same research group examined oral rinse active-matrix metalloproteinase-8 (aMMP-8) point-of-care test (POCT) for differentiating periodontal health, gingivitis, and periodontitis stages. The aMMP-8 POCT demonstrated high specificity (> 90%) but low sensitivity (< 40%), limiting its suitability as a standalone screening tool for periodontitis despite significant differences observed [45]. When combined with subject characteristics, aMMP-8 improves periodontitis detection, supporting its use with other risk factors or sensitive tests. Adjustment for number of teeth present (NTP) further enhanced diagnostic performance and showed strong predictive ability for Stage IV disease (sensitivity 89.7%; specificity 73.6%; AUC 0.856). Combining aMMP-8 with age and smoking further improved accuracy (sensitivity 82.5%; specificity 84.4%; AUC 0.883). These findings suggest that both self-reported questionnaires and POCTs hold promise for detecting advanced periodontitis, though their utility in early-stage disease requires further validation [45]. More recently, they evaluated the diagnostic value of non-conventional predictors, including SROH responses, self-reported gingival bleeding during brushing, and aMMP-8 POCT. Both conventional logistic regression and RF classifiers were tested. While logistic regression allowed only binary classification, RF enabled multiclass discrimination and achieved superior predictive accuracy with lower misclassification rates. Across 100 iterations, the RF algorithm demonstrated strong performance: for periodontal health, sensitivity was 89.3%, specificity 92.3%, and accuracy 91.8%. For gingivitis and periodontitis, both sensitivity and specificity exceeded 91%, with overall accuracy reaching 92.4% [46]. However, approximately 15% of periodontitis cases were misclassified as periodontal health or gingivitis when using SROH data alone. To address this, the authors also explored weighting strategies to reduce this misclassification, but these adjustments introduced trade-offs in specificity for healthy and gingivitis cases [46]. In our study, the RF model trained on combined SROH, and biomarker data misclassified 29.9% of periodontitis cases as either periodontal health or gingivitis, a higher rate than that 15% reported by the study in Hong Kong [46]. Several factors may account for this difference. First, while our model incorporated both SROH data and salivary biomarker (BI), Deng et al. [25] relied solely on SROH. Although biomarkers have the potential to enhance discrimination, their levels may be influenced by transient inflammation or systemic conditions, thereby reducing specificity for periodontitis and increasing overlap with gingivitis profiles. Second, our sample may have included a greater proportion of borderline or clinically ambiguous cases, where disease presentation closely resembles adjacent stages, making accurate classification more difficult. Third, differences in case definitions, population characteristics, and threshold settings may also have contributed to performance variation. While binary classification aligns with screening objectives, the multiclass analysis provides additional insight into the model’s ability to differentiate disease states. Detecting early or moderate periodontitis using limited non-invasive inputs remains challenging. Multi-biomarker panels or additional clinical parameters may enhance predictive accuracy. ML-based models demonstrate strong potential for periodontitis prediction, consistent with a recent systematic review that confirmed AI as a reliable tool for enhancing dental diagnosis, and clinical decision-making [47]. External validation in diverse populations is essential before clinical implementation. For example, in one study that evaluated 10 algorithms, internal validation achieved excellent performance (AUC > 0.95; accuracy > 95%), but these results were not replicated in external validation, where predictive performance dropped substantially. The authors concluded that larger sample size and the inclusion of more complex predictors are needed for ML to achieve its full potential in periodontitis prediction [48]. ML approaches have demonstrated strong potential for periodontal disease prediction. Using nationally representative NHANES data, Vu et al. (2025) reported that ensemble models achieved high predictive performance (AUC up to 0.845), with age and dental attendance emerging as key predictors. These findings support the ability of ML models to integrate diverse data sources and enhance risk stratification for earlier identification of individuals at risk of periodontitis [49]. In addition, ML models based solely on questionnaire and demographic data have shown modest predictive performance. Enevold et al. (2024) reported AUC values of 0.64–0.70 for predicting advanced periodontitis using self-reported measures and demographic variables in Danish cohorts. In contrast, the improved performance observed in the present study highlights the added value of incorporating salivary biomarkers, supporting the integration of subjective and biological indicators to enhance non-invasive screening accuracy [50]. A major strength of the study is combining SROH data with salivary biomarkers to create a fully non-invasive screening, while providing a comparative insight into logistic regression, RF and NB model performance. This novel analysis informs model selection by highlighting trade-offs between interpretability and predictive accuracy, thereby providing a practical framework for researchers and clinicians developing AL-driven periodontal screening tools. The use of non-invasive data and ML supports point-of-care risk stratification, particularly in underserved or primary care setting. Importantly, this approach, if proven to be valid and reliable, may eliminate the need for radiographs or periodontal probing, making it particularly suitable for large-scale, community-based screening initiatives. The findings of this study support the feasibility of a fully non-invasive screening approach for periodontitis based on self-reported outcomes and salivary biomarkers. Such an approach is particularly relevant for primary care, medical clinics, and community-based screening programmes, where periodontal probing and radiographic assessments are not routinely available. By facilitating early identification and referrals of individuals at risk of irreversible periodontal tissue destruction, this strategy may contribute to improved access to periodontal care, more efficient use of specialist resources, and reduced burden of advanced periodontal disease. This study has several limitations. Participants were recruited from a university-based centre, potentially introducing selection bias and limiting generalisability. This may have enhanced the accuracy of self-reported responses and potentially improved model performance compared to a general population setting. Periodontitis cases were analysed as a single group, potentially reducing sensitivity for detecting stage specific or severe disease. Although our sample (n = 77) was sufficient for questionnaire validation and biomarker analysis, ML model development typically requires larger datasets for optimal performance. Although classification was followed for case definition, detailed stratification on severity and staging information was not incorporated and used analytically, which may affect model discrimination across disease stages. Another limitation is the quantification of total MMP-8 and MMP-9, which include inactive pro-forms and may not fully reflect the biologically active inflammatory processes associated with periodontal tissue destruction. The active form (aMMP-8) has demonstrated diagnostic potential, with a recent systematic review reporting pooled sensitivity of 0.63 and specificity of 0.84 for detecting periodontitis [51]. Nevertheless, total MMP-8 remains capable of differentiating periodontal health from disease, as supported by previous meta-analytic and clinical evidence [52, 53]. Future studies employing assays that specifically target active MMP forms may enhance specificity for active disease processes. In addition, given the modest sample size (n = 77), feature-rich dataset, and potential for overfitting, the present findings should be interpreted as exploratory and hypothesis-generating. Ten-fold cross-validation was used to reduce overfitting; however, findings should be interpreted as preliminary and hypothesis-generating. In addition, performance results are reported primarily as point estimates without accompanying 95% confidence intervals, which limits assessment of statistical uncertainty and reflects constraints of the analytical framework used. The limited sample size also precluded stratified analyses by demographic or clinical subgroups. Furthermore, biomarker levels may be influenced by transient inflammation or unmeasured systemic factors, potentially affecting specificity. As a development and internal validation study, the primary aim was feasibility and signal detection rather than population‑level diagnostic accuracy. Larger multi‑centre studies incorporating external validation and robust uncertainty estimation are required prior to clinical implementation. We also recognise that participants were recruited from a periodontics clinic and may represent individuals with higher care-seeking behaviour and dental awareness. This could potentially enhance the accuracy of self-reported measures and introduce selection bias. This limitation has been acknowledged and discussed accordingly. To partially address these constraints, external validation was incorporated to improve the robustness of the findings, although further studies in more representative populations are warranted. In the multiclass setting, smaller per-class sample sizes may increase variability and reduce the stability of performance estimates, and these results should therefore be interpreted as exploratory. Beyond host-derived biomarkers, microbial elements such as virulence factors and microbiome dysbiosis offer significant complementary value in periodontitis detection. For instance, subgingival and salivary endotoxin [lipopolysaccharides (LPS)] activity levels correlate with disease severity, although subgingival levels demonstrate superior diagnostic utility as site-specific biomarkers [54]. Furthermore, the Salivary Microbial Dysbiosis Index (SMDI) has been shown to quantify periodontitis-associated shifts with high diagnostic accuracy, providing a convenient candidate for population-level screening [55]. Recent metagenomic analyses across cohorts from five countries confirmed this dysbiosis signature, identifying disease-enriched species comprising red complex bacteria alongside novel species (e.g., GGB4333/SGB5935), supporting microbial profiling for diagnostic enhancement [56]. Given the multifactorial nature of periodontal diseases, cumulative approaches combining bacterial burden, inflammation, and tissue degradation may further enhance diagnostic accuracy, as demonstrated by the pioneering work of a Finnish research group and subsequently validated in independent studies [57–59]. Future ML models integrating our dataset with additional microbial elements may further refine these non-invasive screening tools. The use of self‑reported periodontal assessment tools is well supported in the literature, particularly in populations with higher oral health awareness such as at dental or medical settings. Dental school–based studies remain essential for advancing periodontal science, as they allow for rigorous clinical monitoring, consistent calibration of examiners, and detailed documentation of periodontal parameters. Viana et al., (2025) validated a self‑reported periodontal screening model in a university dental clinic and Nijland et al., (2021) validated at medical setting reported acceptable diagnostic performance (AUC = 0.72) and severe periodontitis (AUC = 0.73) respectively, confirming that patient‑reported symptoms can serve as effective screening indicators [60, 61]. Future work should focus on larger, diverse cohorts, cost-effectiveness and feasibility analysis of biomarker-assisted screening, and longitudinal study to assess predictive performance for disease progression and treatment response. These steps are essential to validate clinical utility and guide broader adoption of SROH and salivary biomarker detection. Conclusion In conclusion, integrating salivary biomarkers with SROH data provides a promising approach for early, non-invasive detection of periodontitis. The RF model demonstrated superior predictive performance, compared with conventional methods, highlighting its potential for scalable, and cost-effective screening applications. Nevertheless, external validation in larger and more diverse populations remains essential before clinical implementation. Supplementary Information Supplementary Material 1. Acknowledgements The authors would like to express their gratitude to the Faculty of Dentistry, Universiti Malaya, for providing the necessary facilities and technical support that made this research possible. Abbreviations Self-Reported Oral Health Machine Learning Interleukins Tumour Necrosis Factor Metalloproteinases Metallothionein Logistic Regression Random Forest Naive Bayes Area Under the Curve Root Mean Square Error Oral Health-Related Quality of Life Diabetes Mellitus Basic Periodontal Examination Community Periodontal Index of Treatment Needs Scale-level Content Validity Index / Universal Agreement Scale-level Face Validity Index / Average Area Under the Receiver Operating Characteristic Curve Relative Centrifugal Force Enzyme-Linked Immunosorbent Assay Probing Pocket Depth Clinical Attachment Loss Bleeding On Probing Intraclass Correlation Coefficient Standard Deviation Waikato Environmental for Knowledge Analysis Sociodemographic Items SROH Items Biomarker Items Mean Absolute Error Centers for Disease Control and Prevention / American Academy of Periodontology Active Matrix Metalloproteinase Point-Of-Care Test Number of Teeth Present Authors’ contributions Conceptualization: Mohammad Tariqur Rahman, Nor Adinar BaharuddinData curation: Jan Yang Ho, Nor Shafina Mohamed Nazari, Nur Falihin Soehairq Kamal NurzamanFormal analysis: Jan Yang Ho, Abd Aziz AliasFunding acquisition: Nor Adinar BaharuddinInvestigation: Jan Yang Ho, Ainol Haniza Kherul Anuwar, Lee Peng Karen-Ng, Nor Shafina Mohamed Nazari, Nur Falihin Soehairq Kamal NurzamanMethodology: Jan Yang Ho, Lee Peng Karen-Ng, Siti Lailatul Akmar ZainuddinSupervision: Ainol Haniza Kherul Anuwar, Lee Peng Karen-Ng, Siti Lailatul Akmar Zainuddin, Mohammad Tariqur Rahman, Nor Adinar BaharuddinValidation: Jan Yang Ho, Ainol Haniza Kherul AnwarWriting – original draft: Jan Yang Ho, Ainol Haniza Kherul Anuwar, Lee Peng Karen-Ng, Abd Aziz Alias, Siti Lailatul Akmar Zainuddin, Nor Adinar BaharuddinWriting – review & editing: Jan Yang Ho, Ainol Haniza Kherul Anuwar, Lee Peng Karen-Ng, Abd Aziz Alias, Siti Lailatul Akmar Zainuddin, Mohammad Tariqur Rahman, Nor Shafina Mohamed Nazari, Nur Falihin Soehairq Kamal Nurzaman, Nor Adinar Baharuddin. Funding This work was supported by the Fundamental Research Grant Scheme (FP041-2020) by the Malaysian Ministry of Higher Education and Dental Postgraduate Research Grant (UMG006E-2023) by the Faculty of Dentistry, Universiti Malaya. Data availability The datasets generated and/or analysed during the current study are not publicly available due to ethical restrictions related to participant confidentiality and institutional approval requirements. Anonymised data may, however, be made available by the corresponding author upon reasonable request, subject to institutional approval and compliance with applicable ethical and data protection requirements. Declarations Ethics approval and consent to participate This study was approved by the Faculty of Dentistry Medical Ethics Committee, Faculty of Dentistry, Universiti Malaya (Approval No: DF RD2013/0064/2333/23105 (P). All procedures were performed in accordance with the International Conference of Harmonisation - Guidelines for Good Clinical Practice (ICH-GCP) and Helsinki Declaration. Written informed consent was obtained from all participants prior to participation in the study. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1. Data Availability Statement The datasets generated and/or analysed during the current study are not publicly available due to ethical restrictions related to participant confidentiality and institutional approval requirements. Anonymised data may, however, be made available by the corresponding author upon reasonable request, subject to institutional approval and compliance with applicable ethical and data protection requirements.