Development and validation of risk prediction algorithms based on cone beam computed tomography and intraoral scanning fusion data to estimate whether biological width is at a low value: a cross-sectional study Authors: Ying Yuan, Shuo Yang, Xu Wang, Lisha Di, Wulin He Published: 2026 ABSTRACT BackgroundBiological width (BW) safeguards the periodontal ligament and alveolar bone. A reduced BW value may lead to periodontal complications. This study aims to develop and validate digital risk prediction algorithms to assess whether BW is low in adults with good oral hygiene.MethodsFour thousand four hundred twenty sites of 740 teeth from 55 adults with good oral hygiene were used for model development and validation. A mixed-effects logistic regression model was used to derive a risk prediction equation for low BW, incorporating CBCT, intraoral scanning (IOS), and periodontal probing data. Predictors included tooth type, tooth site, gingival margin (GM)-cementum-enamel junction (CEJ) distance, CEJ-alveolar ridge crest (ARC) distance, and gingival thickness (GT). The outcome measure was the classification of BW as low if it was less than 1.13 mm, an operational threshold. Sulcus depth (SD) was measured clinically. GM-CEJ, GT, and D (the distance between GM and ARC) were measured digitally. BW was calculated by subtracting D and SD.ResultsThe raw prediction model is represented by the equation logit(P-raw) = 8.40–2.17*GM-CEJ − 2.25*CEJ-ARC + βSite + βTooth+ (1 | PatientID), where P-raw denotes the probability of low BW. The re-calibration model is logit(P-cal) = 0.1028–0.7159* logit(P-raw). The intra-class correlation coefficient (ICC) was 0.165, indicating that 16.5% of the individual variation was attributable to the patient level. The raw model with an area under the curve (AUC) of 0.867 indicates good discrimination. The calibration curve suggested good internal calibration. The external feasibility test of the model yielded an AUC of 0.68, demonstrating moderate discrimination. The external feasibility test of the recalibration model indicated acceptable calibration. The model’s visual nomogram enables individual site risk calculation for low BW.ConclusionThe developed and validated risk prediction models, based on CBCT and IOS data fusion, may assist to estimate low BW. The inclusion of predictors such as GM-CEJ, CEJ-ARC, tooth site, and tooth type assists clinicians in identifying high-risk tooth sites, thereby facilitating detection of periodontal tissue damage. Future research could develop linear regression models to solve limitations, including the definition of the cutoff and the limited external validation. Background Periodontal examination plays a key role in diagnosing and treating periodontal diseases. The gingiva is the most easily accessible periodontal tissue for monitoring. Biological width (BW) represents the vertical morphology of the gingiva and is defined by Cohen as the junctional epithelium and supra-crestal connective tissue attachment that occupy the space between the base of the gingival crevice (BGC) and the alveolar ridge crest (ARC). BW is used as a historical and operational term conceptually referring to supracrestal tissue attachment (STA) [1]. BW is a protective barrier for the periodontal ligament and supporting bone [2, 3]. Any violation of BW can lead to complications, such as gingival inflammation and alveolar bone loss [4]. Since there is limited data on predictors of BW at the site level, this study aims to develop a non-invasive and precise BW measurement method using digital techniques, combined with clinical methods. Histological studies by Gargiulo [5] concluded that the most consistent component of the dento-gingival junction is BW, with an average length of 2.04 mm, which is not direct clinical guidelines. A clinical trial conducted by Hamasni [2] calculated BW by subtracting sulcus depth (SD) from the gingival margin (GM)-ARC distance. Their results indicated a mean BW of 1.13 mm. Subsequently, the concept of the biological zone acknowledges a dynamic range of physiological tolerance rather than a fixed vertical dimension, provided that the envelope of function is respected [6]. Since junctional epithelium and connective tissue can change over time and in response to environmental factors, it is inadvisable to treat BW as a static measurement [3, 4]. Schmidt et al. [3] reported that the mean BW in humans without a history of periodontal disease ranged from 1.5 to 2.7 mm. Significant heterogeneity in measurement methods and outcomes exists across studies. Factors such as tooth types (BW was smaller around anterior teeth than around posterior teeth), tooth sites (junctional epithelium was larger at mesial/distal sites than buccal/oral sites), the presence of restorations, and history of periodontal diseases can influence BW. Therefore, measuring BW for each individual is essential for providing accurate references for treatment, such as margin placement and gingivectomy. Although several methods for measuring BW exist, no standardized technique has been established. Histological sections have been used to obtain BW measurements, which require a tissue biopsy [3, 5]. BW can also be assessed through transgingival probing; however, this technique is limited to surgical procedures due to its invasiveness. Galgali et al. [7] introduced the parallel profile radiographic (PPR) technique, which allows radiographic evaluation of BW. The mean BW using TGP was 1.6 mm, while the mean BW obtained via the PPR technique was 1.72 mm. However, this method is restricted to anterior teeth and requires a complex clinical procedure. Digital technologies offer a promising alternative to address the limitations of conventional methods. The fusion of CBCT and IOS data through registration and superimposition enables a complementary approach for non-invasive 3D reconstruction of tooth, bone, and soft tissue, facilitating the measurement of the distance from the GM to the ARC [8–10]. Yang et al. [11] developed a CBCT-based 3D visualization system that constructed 3D gingival thickness models, providing millimeter-precision spatial quantification for implant surgical planning. While registration errors vary between different software platforms, some studies have reported registration errors of less than 0.2 mm with markers and less than 0.43 mm without markers, although prosthetic artifacts can increase the error rate [12, 13]. This study measured BW by combining CBCT and IOS data, with a limited number of metal prostheses. Ogawa et al. [8] utilized a combination of CBCT and IOS to measure palatal soft tissue thickness, and applied multiple linear regression algorithms for more precise measurements. Vishakha et al. [14] developed a multiple linear regression algorithm to assess the quantitative effects of periodontal variables, such as pocket depth (PD), on gingival recession. Abdulkarim et al. [15] used a similar digital method as the present study to measure STA and analyzed the associations between STA and factors like gingival thickness (GT) and tooth position. There is a lack of recent research on predicting BW using digital methods combined with regression algorithms. Given that relying on mean BW values may obscure the true clinical scenario [3, 4], this study aims to develop and validate risk prediction algorithms using digital methods based on CBCT and IOS data fusion to estimate the likelihood of low BW in adults with good oral hygiene. Methods Study subjects To develop and validate risk prediction equations using digital methods to estimate whether BW is at a low value, 55 patients (15 males and 40 females) were recruited from the Stomatological Hospital of Southern Medical University in China from September 2021 to December 2022. All procedures involving human participants in this study were conducted in accordance with the Declaration of Helsinki 1975. The study protocol was approved by the Ethics Committee of the Stomatological Hospital of Southern Medical University (No. 2021-28), and written informed consent was obtained from each participant. Inclusion criteria were as follows: (1) age between 20 and 32 years; (2) good oral hygiene, with a full-mouth plaque score ≤ 20% and full-mouth bleeding on probing (BOP) ≤ 10%, and no detectable supragingival or subgingival calculus; (3) at least 10 natural teeth in both the upper and lower dentition; (4) teeth without caries, root canal treatment history, restorations, implants, or devices causing CBCT artifacts; (5) no partial removable dentures or orthodontic palate devices worn. Exclusion criteria included: (1) patients with systemic illnesses (e.g., diabetes mellitus, cancer, bone metabolism disorders, HIV/AIDS), pregnant or breastfeeding women; (2) a history of periodontal surgery, gingival hyperplasia, or medication causing gingival hyperplasia in the past 3 months; (3) severely misaligned or malformed teeth; (4) a history of allergies to alginate, acrylic resins, or resin adhesives. For the logistic regression models, the required sample size is 5–20 cases per independent variable [8, 16]. The dependent variable for the regression analysis was BW, while the independent variables were the 4–7 factors affecting BW identified in previous studies [3], including six tooth sites,, four tooth quadrants, four tooth types, gingival phenotype (quantified by GT), and periodontal condition (gingival recession (GM-cementoenamel junction (CEJ) ), alveolar bone resorption (CEJ-ARC)) [3, 8, 17]. After transforming categorical variables into dummy variables, the number of independent variables ranged from 4 to 17. Therefore, at least 80–340 sites were needed. For subjects with only one site available, a minimum of 40 subjects would be required. However, each subject was expected to contribute 8–16 teeth, ensuring at least 1,920 sites, a sufficient sample size. Consequently, accounting for a 10% dropout rate, 45 subjects were included for model development. Additionally, 10 subjects were selected for external feasibility testing. Tooth quadrants(Upper/Lower (U/L)), left/right (Le/R)), types(incisors, cuspids, premolars, molars), and sites may influence BW; therefore, mesial incisors, cuspids, first premolars, and first molars from the upper left, upper right, lower left, and lower right quadrants were chosen for measurement in this study [3, 8, 17]. For subjects whose teeth did not meet the inclusion criteria, the corresponding ipsilateral teeth were selected. Six dental sites around each selected tooth (mesio-buccal (Mes-B), mid-buccal (Mid-B), disto-buccal (Dis-B), mesio-lingual (Mes-L), mid-lingual (Mid-L), disto-lingual (Dis-L) were measured. A single-blinded design was implemented, with all measurements performed by the same examiner. Subjects received unique identifiers, and the examiner remained blinded to demographic details to minimize potential bias. The digital and clinical procedures for this study are depicted in Fig. 1. Fig. 1Digital and clinical processes for BW measurement and model development: the blue measurement items represent the independent variables of the model. The orange measurement items represent the dependent variables of the model Digital and clinical processes for BW measurement and model development: the blue measurement items represent the independent variables of the model. The orange measurement items represent the dependent variables of the model Evaluation of SD, D, BW, GT BW refers to the width of the attached soft tissue on the coronal side of the ARC, specifically the distance between the bottom of the BGC and the ARC. BW, along with SD, constitutes the supra-crestal gingival tissue (SGT) [18]. This study combined the clinical probing method and digital technology to measure BW. Clinical evaluation of SD with a guide plate SD was clinically measured as the distance between the GM and BGC. The procedure for fabricating the guide plate is outlined as follows (Fig. 2): Fig. 2The process of making the guide plate: (a) marking models; (b) trimming the guide plate for the selected tooth; (c) filling marking materials into the guide plate The process of making the guide plate: (a) marking models; (b) trimming the guide plate for the selected tooth; (c) filling marking materials into the guide plate Impressions of the upper and lower jaws were taken using alginate impression material (Kulzer Dental, Heraplast®, Germany) from 55 subjects, and gypsum models were created. Six measurement sites (Mid-B, Mid-L, Mes-B, Mes-L, Dis-B, Dis-L) were marked on the models.Acrylic clear retainers were fabricated on the gypsum models, trimmed to the contour height of the selected teeth, and used as full-arch guide plates. The guide plate around the unselected teeth was left untrimmed to ensure a proper fit for retention.Vertical grooves were then created in the guide at the six measurement sites around the selected teeth, parallel to the long axis of each tooth. This step standardized the placement and angulation of the probe. The grooves were made using dental high-speed handpieces (NSK VIP - II, Japan) fitted with a cylindrical bur (MANI DIA-BURS, Japan), with a needle tip diameter of 0.5 mm. The grooves were subsequently filled with heated dental gum (Qingpu Dental Clinic, Shanghai) as an imaging material, ensuring consistency between clinical probing sites and imaging measurement sites. Impressions of the upper and lower jaws were taken using alginate impression material (Kulzer Dental, Heraplast®, Germany) from 55 subjects, and gypsum models were created. Six measurement sites (Mid-B, Mid-L, Mes-B, Mes-L, Dis-B, Dis-L) were marked on the models. Impressions of the upper and lower jaws were taken using alginate impression material (Kulzer Dental, Heraplast®, Germany) from 55 subjects, and gypsum models were created. Six measurement sites (Mid-B, Mid-L, Mes-B, Mes-L, Dis-B, Dis-L) were marked on the models. Acrylic clear retainers were fabricated on the gypsum models, trimmed to the contour height of the selected teeth, and used as full-arch guide plates. The guide plate around the unselected teeth was left untrimmed to ensure a proper fit for retention. Acrylic clear retainers were fabricated on the gypsum models, trimmed to the contour height of the selected teeth, and used as full-arch guide plates. The guide plate around the unselected teeth was left untrimmed to ensure a proper fit for retention. Vertical grooves were then created in the guide at the six measurement sites around the selected teeth, parallel to the long axis of each tooth. This step standardized the placement and angulation of the probe. The grooves were made using dental high-speed handpieces (NSK VIP - II, Japan) fitted with a cylindrical bur (MANI DIA-BURS, Japan), with a needle tip diameter of 0.5 mm. The grooves were subsequently filled with heated dental gum (Qingpu Dental Clinic, Shanghai) as an imaging material, ensuring consistency between clinical probing sites and imaging measurement sites. Vertical grooves were then created in the guide at the six measurement sites around the selected teeth, parallel to the long axis of each tooth. This step standardized the placement and angulation of the probe. The grooves were made using dental high-speed handpieces (NSK VIP - II, Japan) fitted with a cylindrical bur (MANI DIA-BURS, Japan), with a needle tip diameter of 0.5 mm. The grooves were subsequently filled with heated dental gum (Qingpu Dental Clinic, Shanghai) as an imaging material, ensuring consistency between clinical probing sites and imaging measurement sites. Clinical evaluation of SD was performed by the same experienced periodontist using a standardized William’s periodontal probe (Silver by Hu-Friedy Goldman Fox, Italy) at a force of approximately 20 g at each selected site, while wearing the guide plate (Fig. 3). Measurements were rounded to the nearest 0.5 mm. The SD at Mid-B, Mid-L, Mes-B, Mes-L, Dis-B, and Dis-L were measured parallel to the long axis of each tooth. These measurement sites were defined according to modifications of previous studies [19]. Fig. 3Periodontal probing with the positioning guide plate Periodontal probing with the positioning guide plate Digital evaluation of D ang GT with a guide plate D was measured digitally as the distance between GM and ARC at six measurement sites around each selected tooth. IOS data acquisition Digital measurements were obtained prior to clinical assessments to avoid potential inaccuracies caused by factors such as increased blood flow and changes in gingival status following periodontal probing. To capture a digital impression of the 3D surface of both teeth and soft tissue, an IOS of the patient’s upper and lower full dental arches was performed by the same experienced clinician. The soft tissue data, obtained using the intraoral scanner (Trios3, 3Shape, Copenhagen, Denmark), included at least all keratinized gingiva. The digital impressions were saved in Standard Tessellation Language (STL) file format or Data Communication Module (DCM) format. CBCT data acquisition 3D data of the bone and teeth were acquired using CBCT (NewTom VG, Verona, Italy) with settings of 110 kV and 6 mAs, utilizing a field of view of 15 × 15 cm and a voxel size of 0.3 mm. During the CBCT imaging, subjects were instructed to wear the previously mentioned guide plates, which were marked with imaging grooves, to ensure alignment with the clinical examination sites and directions. All subjects wore protective lead clothing to minimize radiation exposure. The resulting images were saved in Digital Imaging and Communications in Medicine (DICOM) format. Registration and fusion of IOS and CBCT data 3D data of the bone, teeth, and soft tissue obtained from CBCT and IOS were imported and geometrically superimposed using automatic registration technology in implant planning software (3Shape TRIOS, Copenhagen, Denmark). The STL models of the dentition were aligned with the DICOM models of the 3D virtual head, based on contrast information at the tooth crown margin. After alignment, registration errors were visualized using a 3D threshold color-coded map, generated by surface comparison between CBCT and IOS. The green area indicates an error of less than 0.25 mm, the red area indicates an error range between − 0.25 and 0.25 mm, and the blue area represents errors between 0.25 and 1 mm. Manual adjustments were then made in the non-green areas to reduce the error to less than 0.25 mm (Fig. 4). Fig. 4Registration and fusion of CBCT and IOS in TRIOS software Registration and fusion of CBCT and IOS in TRIOS software Digital evaluation of D and GT in a reconstruction model A 3D reconstruction model was created, ensuring the accurate positioning of the long axis of each selected tooth through panoramic, orthogonal, and sagittal views. The direction of the axis was further verified through a clip view (Fig. 5). A virtual implant was placed to assist in locating the long axis of the tooth, rather than to simulate a real dental implant. Based on the position and direction of the developing groove of the guide plate, the distance from GM to ARC at the six measurement sites—Mid-B, Mid-L, Mes-B, Mes-L, Dis-B, Dis-L—was measured by the same examiner as D, parallel to the long axis of each tooth (Fig. 6 (a), (b)). Measurements were rounded to the nearest millimeter. Fig. 5Virtual implants assist in locating tooth long axes in different planes: (a) panoramic view; (b) axial view; (c) orthogonal view; (d) tangent plane view; (e) clip view Virtual implants assist in locating tooth long axes in different planes: (a) panoramic view; (b) axial view; (c) orthogonal view; (d) tangent plane view; (e) clip view Fig. 6Depth measurement: (a) location of the ARC plane perpendicular to the tooth’s long axis; (b) measurement of D; (c) measurement of GM-CEJ; (d) measurement of CEJ-ARC Depth measurement: (a) location of the ARC plane perpendicular to the tooth’s long axis; (b) measurement of D; (c) measurement of GM-CEJ; (d) measurement of CEJ-ARC In the same reconstruction model, GT at the CEJ at the Mid-B site of each selected tooth was measured as G1 (Fig. 7). The measurements of the reconstructed model were repeated after 72 h. The mean values of these two measurements were used as the final values. Fig. 7GT measurement at the CEJ level at the Mid-B site of each tooth as G1 GT measurement at the CEJ level at the Mid-B site of each tooth as G1 Calculation of BW BW, defined as the distance between BGC and ARC, was calculated by subtracting D and SD measurements for the six measurement sites per tooth (Fig. 8). Fig. 8BW calculation [4]: the digital measurement value D minus the clinical measurement value SD gives the distance between BGC and ARC, representing BW BW calculation [4]: the digital measurement value D minus the clinical measurement value SD gives the distance between BGC and ARC, representing BW Statistical analysis: the prediction model development and validation Development of the prediction model The distribution of the data was evaluated using the Kolmogorov-Smirnov test, which indicated a skewed distribution for BW. Consequently, a cut-off value of 1.13 mm, an operational threshold, based on the average BW from similar research, was selected [2]. BW values below 1.13 mm were considered indicative of a low BW risk. This threshold converted BW from a continuous to a binary variable. Subsequently, the dependent variable was whether BW was low, while GM-CEJ, CEJ-ARC, GT, tooth site, tooth type, Le/R, and U/L were treated as planned independent variables to derive the risk prediction equations for low BW. All independent variables were included in the model, and covariate coefficients were estimated using the maximum likelihood method, with a significance level of α = 0.05. Finally, a nomogram was generated to visualize the model. Validation of the prediction model The same dataset was first used for internal validation of the model, assessing both discrimination and calibration. Discrimination, commonly measured by the C-index, reflects how well the predicted outcomes align with the actual results. For mixed-effects logistic regression, this study plotted the receiver operating characteristic (ROC) curve. The area under the ROC curve (AUC) was selected as the C-index. An AUC value closer to 1 indicates better model discrimination. Calibration was evaluated using the Hosmer-Lemeshow test and calibration curves. For preliminary external feasibility testing, data from tooth sites of 10 additional subjects were used, with discrimination and calibration as the primary evaluation metrics. The methods used for the external feasibility testing were the same as those applied in internal validation. Results In this study, a total of 4,420 sites from 740 teeth of 55 adults were selected. For the development of the BW prediction model, data from 460 teeth (117 central incisors, 115 canines, 118 premolars, and 110 molars) from 45 subjects were used. These included 2,740 depth measurement sites, distributed as follows: 453 Mes-B, 457 Mid-B, 460 Dis-B, 455 Mes-L, 457 Mid-L, and 458 Dis-L. For the external feasibility testing of the BW prediction model, a separate set of 280 teeth (80 incisors, 80 premolars, 80 molars, and 40 incisors) from 10 subjects was employed, corresponding to 1,680 depth measurement sites (280 each for Mes-B, Mid-B, Dis-B, Mes-L, Mid-L, and Dis-L). The measurements taken at each selected site included SD, GM-ARC, GT, GM-CEJ (a composite indicator of gingival index and attachment loss (AL)), and CEJ-ARC. BW values for each site were calculated by subtracting D and SD measurements. The tooth types and sites were recorded in the periodontal chart. All measurements represent the mean values of two repeated assessments made by the same examiner, with an interval of more than 72 h. Correlation analysis The Kolmogorov-Smirnov test for BW (P < 0.001) revealed a non-normal distribution, leading to the use of Spearman’s rank correlation for analysis between BW and other factors. As shown in Table 1, BW exhibited positive correlations with GM-CEJ, CEJ-ARC, and tooth site, with rank-sum correlation coefficients (rs) of 0.24, 0.28, and 0.08 (P < 0.001), respectively. It also showed negative correlations with GT and tooth type, with rs of -1.33 and − 0.23 (P < 0.001), respectively. No correlation was found between BW and Le/R, U/L (P ≧ 0.01), indicating no significant difference in BW between right/left and upper/lower lateral teeth. Consequently, Le/R and U/L were excluded as dependent variables in the subsequent regression analysis. Table 1Spearman rank correlation analysis of BW and preset influencing factorsVariablesrs P BW(mm)GM-CEJ(mm)0.24***< 0.001CEJ-ARC(mm)0.28***< 0.001GT(mm)-1.33***< 0.001Tooth site0.08***< 0.001Tooth type-0.23***< 0.001U/L-0.060.01Le/R0.020.44***P < 0.001 Table 1Spearman rank correlation analysis of BW and preset influencing factorsVariablesrs P BW(mm)GM-CEJ(mm)0.24***< 0.001CEJ-ARC(mm)0.28***< 0.001GT(mm)-1.33***< 0.001Tooth site0.08***< 0.001Tooth type-0.23***< 0.001U/L-0.060.01Le/R0.020.44***P < 0.001 Spearman rank correlation analysis of BW and preset influencing factors ***P < 0.001 Mixed-effects logistic regression model to estimate whether biological width is at a low value Development of the prediction model The mean BW value reported across various studies exhibits considerable variation. The primary objective of this regression model is to predict low BW values, aiming to estimate the probability of a low BW when the values of the independent variables are known. In this study, an operational cutoff value of 1.13 mm for BW, derived from a similar method [2], was selected. BW was then transformed into a binary variable, referred to as BW-recode. BW values ≥ 1.13 mm were classified as non-low, with a label value of 0, while BW values < 1.13 mm were classified as low, with a label value of 1. Based on the correlation analysis, the independent variables selected for the mixed-effects logistic regression model were GM-CEJ, CEJ-ARC, GT, tooth site, and tooth type, with BW-recode serving as the dependent variable. Tooth type was numerically encoded as follows: incisor = 1, canine = 2, premolar = 3, molar = 4. Tooth site was encoded as: Mes-B = 1, Mid-B = 2, Dis-B = 3, Mes-L = 4, Mid-L = 5, and Dis-L = 6. These encodings were used to create dummy variables for the analyses, where Dis-B and molar were considered the reference group. The raw mixed-effects logistic regression model (Eq. 1) was derived by excluding statistically insignificant independent variables (Table 2). Table 2Mixed-effects logistic regression analysis model table of BW-recodeVariablesCoefficientORCI-lowerCI-upper P Intercept8.40 ***4430.721399.6814025.61< 0.001GM-CEJ (mm)-2.17 ***0.120.090.14< 0.001CEJ-ARC (mm)-2.25 ***0.110.080.14< 0.001GT (mm)0.181.200.662.170.553SiteMid-B-1.37 ***0.250.170.39< 0.001SiteMes-B-0.400.670.441.020.058SiteDis-L-0.66 ***0.520.350.770.001SiteMid-L-1.75 ***0.170.120.26< 0.001SiteMes-L-0.66 ***0.520.350.760.001Toothpremolar-0.61 **0.540.360.82< 0.01Toothcanine-1.35 ***0.260.170.40< 0.001Toothincisor-1.52 ***0.220.150.33< 0.001**P < 0.01; ***P < 0.001 Table 2Mixed-effects logistic regression analysis model table of BW-recodeVariablesCoefficientORCI-lowerCI-upper P Intercept8.40 ***4430.721399.6814025.61< 0.001GM-CEJ (mm)-2.17 ***0.120.090.14< 0.001CEJ-ARC (mm)-2.25 ***0.110.080.14< 0.001GT (mm)0.181.200.662.170.553SiteMid-B-1.37 ***0.250.170.39< 0.001SiteMes-B-0.400.670.441.020.058SiteDis-L-0.66 ***0.520.350.770.001SiteMid-L-1.75 ***0.170.120.26< 0.001SiteMes-L-0.66 ***0.520.350.760.001Toothpremolar-0.61 **0.540.360.82< 0.01Toothcanine-1.35 ***0.260.170.40< 0.001Toothincisor-1.52 ***0.220.150.33< 0.001**P < 0.01; ***P < 0.001 Mixed-effects logistic regression analysis model table of BW-recode **P < 0.01; ***P < 0.001 The regression coefficient values for GM-CEJ, CEJ-ARC, tooth site, and tooth type were − 2.17, -2.25, and other negative numbers, respectively. This indicates that GM-CEJ, CEJ-ARC, tooth site, and tooth type negatively influence BW-recode, while GT has no significant effect on BW-recode (P > 0.05).1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm{logit(P-raw)}=8.40-2.17\ast\mathrm{GM-CEJ}-2.25\ast\mathrm{CEJ-ARC}+\beta\mathrm{Site}+\beta\mathrm{Tooth}+\left(1\mid\mathrm{PatientID}\right)$$\end{document} (where P-raw represents the probability that the BW-recode is a low value, i.e., BW < 1.13 mm). Internal validation and visualization of the model Random effects Random effects Random effects The standard deviation of the random intercept for patients was 0.8077. The intra-class correlation coefficient (ICC) was approximately 0.165, indicating that about 16.5% of the individual variation was attributable to the patient level. (2) Discrimination (2) Discrimination Discrimination Model discrimination was used to assess the prediction capability of the model. As shown in Fig. 9, the ROC curve is plotted, and the AUC, also referred to as the C statistic, is calculated to evaluate the model’s success in predicting low BW values. The C of this model is 0.867, indicating that the logistic regression model successfully distinguishes low BW values with a probability of 86.70%. Fig. 9The ROC curve of the BW mixed-effects logistic regression model: model with an AUC of 0.867, indicates good model discrimination The ROC curve of the BW mixed-effects logistic regression model: model with an AUC of 0.867, indicates good model discrimination (3) Calibration (3) Calibration Calibration As shown in Fig. 10, the calibration plot demonstrated good internal calibration, with decile-specific observed event rates closely matching predicted probabilities of low BW along the ideal line. Fig. 10Calibration curve for the BW mixed-effects logistic regression model: the curves closely align with the ideal line, suggesting good model calibration Calibration curve for the BW mixed-effects logistic regression model: the curves closely align with the ideal line, suggesting good model calibration (4) Model visualization (4) Model visualization Model visualization For further model visualization, a nomogram was constructed. As shown in Fig. 11, the four significant variables in the regression model—GM-CEJ, CEJ-ARC, tooth site, and tooth type—are represented on the vertical axis of the nomogram, with each independent variable corresponding to a line segment labeled with a scale. The length of each line segment reflects the magnitude of each factor’s contribution to the outcome of low BW. The value of each independent variable is associated with a score at the top, and the sum of the scores from all independent variables yields the total score at the bottom. Finally, the probability of low BW at an individual site can be determined from the total score corresponding to the bottom line segment. For example, if GM - CEJ and CEJ - ARC are 3 mm and 2 mm, the corresponding single - item points are 21.5 and 79. For the Dis - B site of a molar, the corresponding points are 6 and 5. The total score is 111.5. Drawing a vertical line from 111.5 on the total points scale gives a reading of 0.32 on the probability of the low BW line, indicating a 32% probability of a low BW value. Fig. 11Visualization of the logistic regression prediction model using a nomogram: the risk probability of low BW value is calculated by summing the points of the independent variables Visualization of the logistic regression prediction model using a nomogram: the risk probability of low BW value is calculated by summing the points of the independent variables External feasibility testing of the model A total of 280 teeth (80 incisors, 80 premolars, 80 molars, and 40 cuspids) from 10 subjects were used for the external feasibility testing of the BW prediction model, comprising 1,680 depth measurement sites (280 for each of Mes-B, Mid-B, Dis-B, Mes-L, Mid-L, and Dis-L). The raw model was poorly calibrated in the external feasibility testing. After recalibration, the calibration was significantly improved, while the discrimination remained stable. The re-calibration equation is as follows:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm{logit(P-cal)}=0.1028-0.7159\ast\mathrm{logit(P-raw)}$$\end{document} (where P-cal represents the probability that the BW-recode is a low value, calculated by the recalibrated model). External feasibility testing of model discrimination External feasibility testing of model discrimination External feasibility testing of model discrimination As shown in Fig. 12, the external feasibility testing ROC curve was plotted with an AUC of 0.68, indicating that the probability of correctly distinguishing low BW sites using this mixed-effects logistic regression model is 68.00%. Fig. 12External feasibility testing ROC curve of the mixed-effects logistic regression model: the AUC of the raw model (a) and re-calibration model (b) were both 0.68, indicating moderate model discrimination External feasibility testing ROC curve of the mixed-effects logistic regression model: the AUC of the raw model (a) and re-calibration model (b) were both 0.68, indicating moderate model discrimination (2) External feasibility testing of model calibration (2) External feasibility testing of model calibration External feasibility testing of model calibration The calibration of the raw model was assessed using the Hosmer-Lemeshow test, with P < 0.001, indicating a poor fit. The calibration of the re-calibration equation was significantly improved (Hosmer-Lemeshow test P = 0.74). The calibration curve was plotted, as shown in Fig. 13. Fig. 13External feasibility testing calibration curve of the BW mixed-effects logistic regression model: (a) The raw model shows poor calibration; (b) The calibration curve of the re-calibration model lies on either side of the ideal line, indicating acceptable calibration External feasibility testing calibration curve of the BW mixed-effects logistic regression model: (a) The raw model shows poor calibration; (b) The calibration curve of the re-calibration model lies on either side of the ideal line, indicating acceptable calibration Discussion In recent years, multidimensional virtual patients can be reconstructed using multimodal data from both oral hard and soft tissues [20]. However, the current technologies have limited capability in measuring deep gingival structures. The characterization of the gingiva is a crucial indicator of periodontal health, influencing both oral aesthetics and the outcomes of treatments [21]. Although BW has traditionally been considered a fixed value [5], it can change dynamically due to various factors. In the present study, following the approach of Hamasni et al. [22], the BW was derived from digitally measured D combined with clinically measured SD, leading to the development of a prediction model. The model was conducted using CEJ-ARC, GM-CEJ, and individual indices, enabling non-invasive assessment of BW. Correlation analysis Regarding the correlation between BW and other factors: (1) In this study, BW showed no correlation with the left and right sides of the dental arch, but a positive correlation with GM-CEJ and CEJ-ARC. Additionally, maxillary BW was greater than mandibular BW. (2) Tooth type exhibited a negative correlation with BW, suggesting that anterior BW was greater than posterior BW, which contrasts with several earlier studies [3, 23, 24]. The discrepancy may be due to potential AL in this study, with more connective tissue being preserved in the anterior region. Dental sites showed a positive correlation with BW, indicating that BW followed the pattern: buccal < lingual, proximal-medial < central < distal-medial. (3) This study found a negative correlation between GT and BW, which is different from previous studies [24, 25]. This may be due to the fact that the metrics used in this study quantified thickness 1 mm laterally from the root of the GM, as opposed to classifying the gingiva as thick or thin. In summary, the correlation analysis conducted in this study examined individual factors, thus necessitating multiple regression modeling for a more comprehensive analysis. Development and validation of the BW mixed-effects logistic regression model Similar models and model development Multiple linear regression models have been utilized to enhance the accuracy of CBCT and IOS measurements of GT [8, 26]. A cross-sectional study employed multiple logistic regression models to predict periodontitis based on variables such as age, gender, and BOP [27]. Santos M et al. [28] developed a model to predict periodontal inflammation by measuring cuspal width at the free GM level to estimate mandibular canal height (MCh) dimensions, with a standard error of only 0.4 mm. Given that the clinical measurement of BW is invasive, the present study integrated digitized indices such as CEJ-ARC, GM-CEJ, GT, and other accessible variables to construct a predictive BW model. The goal is to integrate the digitized approach with the prediction model as a complementary adjunct to enhance clinical decision-making. Model validation The ROC curves demonstrated good differentiation, with an internal validation AUC of 0.867, indicating strong diagnostic value. The external feasibility testing, however, showed moderate discrimination, with an AUC of 0.68. The moderate performance may be attributed to the following factors: (1) Patient random effect: ICC indicated that 16.5% of the individual variation was attributable to the patient level. (2) Potential omission of important independent variables from the model, necessitating the inclusion of additional variables in future iterations; and (3) the use of the external feasibility testing data from previous patient information, which may have differed from the standards applied in the model’s development. Therefore, future validation should incorporate a new set of data with stricter standards. The Hosmer-Lemeshow test indicated poor calibration, and the discrepancies observed in the calibration results may partly be due to the limitations of the test itself, particularly in large sample sizes [29]. After recalibration, the calibration was significantly improved. Since existing studies did not define low BW, we adopted the mean BW value of 1.13 mm from a similar method [2] as a cutoff. However, its generalizability may be influenced by inherent biological variation across different populations. The variability in defining the cutoff for low BW affected the calibration. Furthermore, the limitations of the small preliminary external feasibility testing sample and the age-restricted cohort are also factors affecting the test. Given the model’s limited prediction accuracy and calibration, subsequent studies should incorporate a larger external validation sample and different demographic cohorts to optimize the model [30]. Model visualization The application of the regression model requires statistical expertise to calculate the probability of a low BW value, which may limit its generalizability in clinical settings. To address this, the model was visualized using a nomogram. The independent variables are assigned point values, and the sum of these points determines the total score, which can then be converted into the probability of a low BW value. This provides a more accessible method to use the model for early warning of low BW values [31, 32]. It is important to note that the narrow range of the predictive bands on the nomogram suggests that the model’s strong predictive ability is confined to a specific range. Therefore, clinicians should use it in conjunction with clinical judgment to account for other potentially important risk factors. Risk of Bias (ROB) evaluation The Prediction Model Risk of Bias Assessment Tool (PROBAST) was utilized for evaluating the risk of bias (ROB). “Y” indicates low ROB, and “N” indicates high ROB [33]. The assessment revealed the following: (1) Participants: Y – This study is a diagnostic prediction model designed to predict low BW values, and a cross-sectional design was appropriately used. The inclusion criteria were not restricted to individuals with healthy periodontal tissue, which helped avoid selection bias; (2) Predictors: Y – The evaluation criteria for the independent variables were clear, consistent, and applied uniformly across all included tooth sites, independent of the outcome of low BW; (3) Outcomes: Y – The outcome definitions were meaningful, standardized, and consistently applied. N – Thresholds for predicting low BW have not been clearly reported. In this study, the cutoff was based on the mean value from previous studies, which introduces some risk of bias. Additionally, there is a potential adulteration bias due to the correlation between BW measured using the same digital method and the independent variables GM-CEJ and CEJ-ARC; (4) Statistical analyses: Y – The study had an adequate sample size with no missing values, which is essential for ensuring the distinction and discrimination power of the model. Overfitting was avoided through the use of corrected indicators. N – Some bias may have been introduced due to the exclusion of extreme BW values. Limitations Inclusion Criteria: Gingival health in this study was defined as ≤ 10% BOP with SD ≤ 3 mm in patients with an intact periodontium and SD ≤ 4 mm in patients with periodontal loss [34]. Subjects were included based on gingival health, and periodontal health-related independent variables were incorporated into the model, including GM-CEJ as an indicator of gingival recession, CEJ-ARC to reflect alveolar bone resorption, and GM-ARC to assess the size of the SGT. However, other indicators, such as AL, which may also influence BW [3], were not included and should be considered in future models. Additionally, the study focused on subjects aged 20–32 years, so the current model is limited to this age range. Future studies should stratify BW characteristics by age group and periodontal condition [35].Sample Size: The sample size used to develop the model meets the requirements for the prediction model. However, a larger sample size will be necessary for broader clinical application.Experimental Design: To further validate the prediction model, it is crucial to collect new multi-center data for the external feasibility testing. All measurements in this study were completed by a single examiner, and future research should involve multiple examiners to improve the reproducibility of results. Repeated measurements and averaging of results will enhance intra- and inter-examiner consistency. Furthermore, future studies may use algorithms to address low-contrast edges and partial volume effects in CBCT.BW Prediction Model: In this study, to minimize trauma, BW is defined as (GM–CEJ + CEJ–ARC) − SD, and GM–CEJ and CEJ–ARC are also included as predictors; the outcome and predictors are not fully independent. Also, due to the inherent limitations in the precision of periodontal probing [36], clinical BW could be used as an alternative outcome variable to generate a similar model. The current study has a limitation in analyzing how SD variability affects the BW model. The single - time point data collection inherently fails to capture the dynamic SD. Future longitudinal studies are needed to improve the understanding of this relationship. Inclusion Criteria: Gingival health in this study was defined as ≤ 10% BOP with SD ≤ 3 mm in patients with an intact periodontium and SD ≤ 4 mm in patients with periodontal loss [34]. Subjects were included based on gingival health, and periodontal health-related independent variables were incorporated into the model, including GM-CEJ as an indicator of gingival recession, CEJ-ARC to reflect alveolar bone resorption, and GM-ARC to assess the size of the SGT. However, other indicators, such as AL, which may also influence BW [3], were not included and should be considered in future models. Additionally, the study focused on subjects aged 20–32 years, so the current model is limited to this age range. Future studies should stratify BW characteristics by age group and periodontal condition [35]. Inclusion Criteria: Gingival health in this study was defined as ≤ 10% BOP with SD ≤ 3 mm in patients with an intact periodontium and SD ≤ 4 mm in patients with periodontal loss [34]. Subjects were included based on gingival health, and periodontal health-related independent variables were incorporated into the model, including GM-CEJ as an indicator of gingival recession, CEJ-ARC to reflect alveolar bone resorption, and GM-ARC to assess the size of the SGT. However, other indicators, such as AL, which may also influence BW [3], were not included and should be considered in future models. Additionally, the study focused on subjects aged 20–32 years, so the current model is limited to this age range. Future studies should stratify BW characteristics by age group and periodontal condition [35]. Sample Size: The sample size used to develop the model meets the requirements for the prediction model. However, a larger sample size will be necessary for broader clinical application. Sample Size: The sample size used to develop the model meets the requirements for the prediction model. However, a larger sample size will be necessary for broader clinical application. Experimental Design: To further validate the prediction model, it is crucial to collect new multi-center data for the external feasibility testing. All measurements in this study were completed by a single examiner, and future research should involve multiple examiners to improve the reproducibility of results. Repeated measurements and averaging of results will enhance intra- and inter-examiner consistency. Furthermore, future studies may use algorithms to address low-contrast edges and partial volume effects in CBCT. Experimental Design: To further validate the prediction model, it is crucial to collect new multi-center data for the external feasibility testing. All measurements in this study were completed by a single examiner, and future research should involve multiple examiners to improve the reproducibility of results. Repeated measurements and averaging of results will enhance intra- and inter-examiner consistency. Furthermore, future studies may use algorithms to address low-contrast edges and partial volume effects in CBCT. BW Prediction Model: In this study, to minimize trauma, BW is defined as (GM–CEJ + CEJ–ARC) − SD, and GM–CEJ and CEJ–ARC are also included as predictors; the outcome and predictors are not fully independent. Also, due to the inherent limitations in the precision of periodontal probing [36], clinical BW could be used as an alternative outcome variable to generate a similar model. The current study has a limitation in analyzing how SD variability affects the BW model. The single - time point data collection inherently fails to capture the dynamic SD. Future longitudinal studies are needed to improve the understanding of this relationship. BW Prediction Model: In this study, to minimize trauma, BW is defined as (GM–CEJ + CEJ–ARC) − SD, and GM–CEJ and CEJ–ARC are also included as predictors; the outcome and predictors are not fully independent. Also, due to the inherent limitations in the precision of periodontal probing [36], clinical BW could be used as an alternative outcome variable to generate a similar model. The current study has a limitation in analyzing how SD variability affects the BW model. The single - time point data collection inherently fails to capture the dynamic SD. Future longitudinal studies are needed to improve the understanding of this relationship. Clinical implications The current prediction model primarily serves as an early warning tool for BW violation, offering an addition to periodontal exams for assessing the risk of periodontal diseases. Clinically, BW is often compromised to achieve better restoration outcomes, which may lead to gingival symptoms [37]. Surgical procedures, like crown lengthening or orthodontic treatments, can correct BW violations. Gingival resection or alveolar bone remodeling, based on BW considerations, can ensure the quality of restorations [38, 39]. The digital method in this study enables non-invasive assessments of BW prior to treatments. Future research will also leverage similar digital techniques to develop linear regression models, facilitating the quantitative calculation of BW. While the current prediction model shows promise, it still requires refinement. The initial model can be enhanced by incorporating additional predictors. This study combines digital methods with multiple regression modeling for periodontal measurements, offering new methodological insights for digital, non-invasive, gingival assessments. This diagnostic prediction model could eventually evolve into a prognostic tool for predicting future BW risks, incorporating time factors to assess the potential for disease development in high-risk populations, ultimately aiding in the prevention of periodontal disease. Conclusion Using fusion data from CBCT and IOS, this study developed mixed-effects logistic regression algorithms to estimate low BW values in adults aged 20–32 with good oral hygiene. The logistic regression model, incorporating GM-CEJ, CEJ-ARC, tooth site, and tooth type, demonstrated acceptable discrimination and calibration. The visualized nomogram of this model provides clinicians with a practical tool for early identification of low BW values at the dental site level, facilitating the early detection of periodontal tissue destruction and offering valuable clinical guidance. Future research could address biological width as a continuous variable to develop linear regression models, facilitating the quantitative calculation of BW.