Tooth-level risk modeling and prediction of fixed orthodontic bracket bonding failure: a real-world retrospective cohort study

Abstract

Objective

Based on real-world retrospective cohort data, this study aims to identify risk factors for fixed orthodontic bracket bonding failure, develop a risk prediction model, and create a visualized online predictive tool for individualized clinical risk management.

Methods

A total of 5,808 tooth positions from 286 patients undergoing fixed orthodontic treatment during January 2021-January 2024 were analyzed. Using tooth position as the analysis unit and patient as the clustering unit, a multivariable Logistic regression model was constructed via generalized estimating equations (GEE) to identify independent risk factors for bracket bonding failure. A standard Logistic regression model was established for prediction. Model performance was evaluated using ROC curve, Hosmer-Lemeshow calibration test, and Bootstrap resampling. A web-based dynamic prediction tool was developed based on the prediction equation.

Results

Bonding failure occurred at 548 tooth positions (9.44%), involving 239 patients. The GEE model identified plaque index, gingival index, hard food biting, bruxism, molar position, ceramic bracket, and use of intermaxillary elastics as independent risk factors (all P < 0.05). The prediction model exhibited an AUC of 0.613 (95% CI: 0.588–0.633). The Hosmer-Lemeshow test yielded a P-value of 0.419, and the Bootstrap-corrected AUC was 0.6027. The developed online dynamic tool provided real-time outputs including the predicted failure probability for a single tooth position, corresponding risk level, and clinical recommendations.

Conclusion

Bracket bonding failure is multifactorial, linked to oral hygiene, habits, tooth position, and bracket type. The well-calibrated prediction model and accompanying online tool can aid clinicians in risk assessment and support personalized treatment planning.

Introduction

Fixed orthodontic therapy is a primary clinical modality for malocclusion correction, the efficacy of which relies on the stable bonding of brackets to the tooth surface. Bracket bonding failure not only prolongs treatment duration, increases appointment frequency and healthcare costs, but may also disrupt the planned sequence and direction of tooth movement, thereby compromising treatment efficiency and patient experience. Previous research reports bracket bonding failure rates ranging approximately from 5% to 10%, with significant variation across different centers, bonding systems, and patient populations [1]. Bracket debonding, as a common yet potentially preventable treatment-interrupting event, remains of significant clinical relevance to identify its risk factors.

The factors influencing bracket bonding stability are complex, involving multiple dimensions such as patient characteristics, tooth position, operator skill, and treatment environment. Existing literature suggests that poor oral hygiene, inadequate plaque control, and gingival inflammation may increase stress on the bracket or compromise the adhesive interface [2]. Anatomical factors, including tooth type, may also lead to varying degrees of occlusal interference or mechanical stress concentration [3]. However, the combined impact of these factors in a real-world clinical setting lacks substantial supporting data. Moreover, most prior investigations focused on isolated risk factors, lacking a systematic and comprehensive risk prediction tool.

In recent years, real-world data (RWD)-based cohort studies have gained prominence as a valuable clinical research methodology, offering a more authentic reflection of clinical treatment processes and outcomes, thus providing a reliable data foundation for predictive model development [4]. Nevertheless, research on predictive models specifically for fixed bracket bonding failure remains limited, particularly studies employing rigorous statistical methods that account for the clustered nature of the data, using the tooth as the analysis unit and the patient as the clustering unit. Concurrently, systematically established, individualized risk prediction tools providing real-time clinical decision support are notably absent.

Addressing this evidence gap and clinical need, this study utilized real-world retrospective cohort data to systematically evaluate risk factors for fixed bracket bonding failure. A multivariable predictive model was constructed using generalized estimating equations (GEEs). Furthermore, a web-based dynamic prediction tool was developed to enable quantitative risk assessment and visual representation of bracket bonding failure risk, thereby supporting individualized orthodontic treatment planning.

Materials and methods

Study population

A retrospective cohort design was employed. Clinical data from 286 patients who underwent fixed orthodontic treatment at our hospital between January 2021 and January 2024 were collected. The study protocol was reviewed and approved by the Ethics Committee of Changsha Stomatological Hospital (Approval No.: 2025 − 173). The study adhered strictly to the ethical principles of the Declaration of Helsinki. Given the retrospective design and no compromise to patient privacy, the Ethics Committee granted a waiver for informed consent. All patient data were anonymized and de-identified prior to analysis.

Inclusion criteria were: (1) Patients receiving fixed orthodontic appliance therapy; (2) Availability of complete baseline data and follow-up records; (3) Documented bonding status for each bracket/tooth position throughout the treatment course. Exclusion criteria were: (1) History of prior orthodontic treatment or bracket bonding; (2) Presence of systemic diseases severely affecting oral health (e.g., severe diabetes, history of head and neck radiotherapy); (3) Missing information on bonding materials or appliances precluding bracket status assessment; (4) Inadequate follow-up duration to determine bonding failure; (5) Incomplete records or high proportion of missing key information. Ultimately, 286 patients corresponding to 5,808 tooth positions were included. Bracket bonding failure was defined as complete detachment of the bracket from the tooth surface requiring re-bonding during the treatment period. Multiple failures at the same tooth position were counted as a single event.

Sample size calculation: Based on preliminary investigation and previous literature, the estimated bonding failure rate for fixed brackets is approximately 5–10% [5]. This study aimed to enroll 286 patients, corresponding to roughly 4,000–6,000 tooth positions. Conservatively estimating a 5% failure rate would yield approximately 200–300 failure events; an 8% estimate yields about 320–480 events. The study planned to construct a multivariable risk prediction model incorporating approximately 10–15 independent variables. According to the empirical rule of an events per variable (EPV) ratio ≥ 10, a minimum of 100–150 failure events is required. Given the anticipated total number of tooth positions and failure rate, the expected number of events far exceeds this minimum, ensuring adequate statistical power for constructing a multivariable GEE model and conducting internal validation. Therefore, the sample size was deemed sufficient and robust.

Data sources and variable definitions

Data were extracted from the electronic health record system and the imaging database of the Department of Orthodontics. All data were structured and anonymized prior to extraction. Patient-level variables included age, gender, Angle’s classification, overbite and overjet (mm), plaque index (PLI), gingival index (GI), detrimental habits (hard food biting, bruxism). For continuous variables, averages from initial and follow-up visits were used. Tooth-level variables included tooth type (incisor, canine, premolar, molar), jaw (maxillary/mandibular), side (left/right), bracket material (stainless steel/ceramic), bracket structure (self-ligating/conventional), use of bite plate, and use of intermaxillary elastics. For statistical analysis, the “tooth position” was the unit of analysis, with “patient” designated as the clustering unit to account for intra-patient correlation across multiple teeth. All metal brackets used in this study were conventional stainless-steel brackets featuring a mechanically retentive woven mesh base, whereas ceramic brackets employed a micro-mechanically retentive base design according to the manufacturer’s specifications. All bracket bonding procedures were performed by multiple orthodontic practitioners from the same clinical team, following a standardized bonding protocol routinely used at our center.

Outcome measure

The primary outcome was bracket bonding failure (Yes/No). A failure event was recorded if a bracket at a specific tooth position completely debonded and required re-bonding during treatment. Considering that repeated failures at the same tooth likely reflect the same underlying cause and are not statistically independent, multiple failures per tooth were counted as a single bonding failure event [6].

Statistical analysis

Data analysis was performed using R software (version 4.3.2). Continuous variables were described as mean ± standard deviation or median (interquartile range) based on their distribution, with between-group comparisons conducted using t-tests or Mann-Whitney U tests. Categorical variables were summarized as frequencies and percentages, compared using Chi-square (χ²) or Fisher’s exact tests. To identify potential risk factors for bonding failure, univariable analysis was first performed. Variables with P < 0.10 and clinical relevance were entered into the multivariable model. Prior to multivariable analysis, variance inflation factor (VIF) was used to diagnose multicollinearity among candidate independent variables. A VIF > 5 (or > 10) indicated significant collinearity, prompting variable selection or combination based on clinical and statistical considerations. Given the correlated outcomes among multiple teeth within the same patient, a multivariable Logistic regression model was constructed using the GEE method, specifying the patient as the clustering unit and an exchangeable working correlation structure to estimate the marginal effects of variables, calculating odds ratios (ORs) and their 95% confidence intervals (CIs). While the GEE model is primarily used for inference, the development of a risk prediction tool requires individual-level predictions. Therefore, a standard Logistic regression model (glm) with the same variable set as the GEE model was fitted to generate predicted probability values and drive the online prediction tool. Model discrimination was assessed via the receiver operating characteristic (ROC) curve and the area under the curve (AUC). Calibration was evaluated using calibration curves and the Hosmer-Lemeshow (HL) test. Internal validation was performed using the Bootstrap resampling method to obtain bias-corrected AUC and calibration estimates. All statistical tests were two-sided, with P < 0.05 considered statistically significant.

Results

Comparison of clinical characteristics between failure and non-failure groups

This study included 286 fixed orthodontic patients, totaling 5,808 tooth positions. Among them, 239 patients experienced bonding failure at least once, corresponding to 548 failure tooth positions (tooth-level failure rate: 9.44%). The remaining 47 patients had no bracket failure events during follow-up. Overall, 5,260 tooth positions did not experience bonding failure. Given the different levels of analysis, patient-level and tooth-level variables are presented separately in Tables 1 and 2.

Table 1.

Comparison of clinical characteristics (patient as unit, n = 286)

Variable category	Indicator	Stratification	Failure group (n = 239)	Non-failure group (n = 47)	χ²/t	P
Demographics	Age (years)	—	21.48 ± 6.91	22.04 ± 6.91	-1.812	0.07
	Gender	Female	154 (64.4%)	30 (63.8%)	0.005	0.943
		Male	85 (35.6%)	17 (36.2%)
Malocclusion	Angle’s classification	Class I	155 (64.9%)	28 (59.6%)	0.721	0.698
		Class II	51 (21.3%)	11 (23.4%)
		Class III	33 (13.8%)	8 (17.0%)
	Overbite (mm)	—	3.17 ± 1.44	3.15 ± 1.46	0.263	0.793
	Overjet (mm)	—	3.26 ± 1.47	3.17 ± 1.50	1.377	0.169
Oral hygiene	PLI	—	1.30 ± 0.49	1.25 ± 0.50	2.280	0.023
	GI	—	1.07 ± 0.49	1.01 ± 0.48	2.688	0.007
Detrimental habits	Hard food biting	No	188 (78.7%)	41 (87.2%)	1.833	0.176
		Yes	51 (21.3%)	6 (12.8%)
	Bruxism	No	187 (78.2%)	43 (91.5%)	4.569	0.033
		Yes	52 (21.8%)	4 (8.5%)

PLI Plaque index, GI Gingival index

Table 2.

Comparison of clinical characteristics (tooth as unit, n = 5,808)

Variable category	Indicator	Stratification	Failure teeth (n = 548)	Non-failure teeth (n = 5,260)	χ²/t	P
Tooth-related features	Tooth type	Incisor	179 (32.7%)	1,887 (35.9%)	9.590	0.022
		Canine	131 (23.9%)	1,314 (25.0%)
		Premolar	134 (24.5%)	1,314 (25.0%)
		Molar	104 (19.0%)	745 (14.2%)
Bracket features	Bracket material	Metal	422 (77.0%)	4,303 (81.8%)	7.220	0.007
		Ceramic	126 (23.0%)	957 (18.2%)
	Bracket structure	Self-ligating	194 (35.4%)	2,093 (39.8%)	3.820	0.051
		Conventional	354 (64.6%)	3,167 (60.2%)
Treatment-related factors	Bite plate	No	446 (81.4%)	4,091 (77.8%)	3.580	0.059
		Yes	102 (18.6%)	1,169 (22.2%)
	Intermaxillary elastics	No	312 (56.9%)	3,438 (65.4%)	15.040	< 0.001
		Yes	236 (43.1%)	1,822 (34.6%)

Patient-level comparisons (Table 1) revealed that the failure group had significantly higher PLI and GI scores than the non-failure group (both P < 0.05), suggesting an association between oral hygiene status and bonding failure. No significant differences were observed between the groups for age, gender, Angle’s classification, overbite, or overjet (all P > 0.05).

Tooth-level comparisons (Table 2) showed significant differences in tooth type, bracket material, and use of intermaxillary elastics between the groups (all P < 0.05). Specifically, molar positions had a higher proportion of failures, ceramic brackets failed more frequently than metal brackets, and tooth positions subjected to intermaxillary elastics were more prone to bonding failure. Differences in bracket structure and use of bite plates showed borderline statistical significance (P ≈ 0.05).

Multivariable GEE analysis of independent risk factors for bracket bonding failure

Using bracket bonding failure (0 = No, 1 = Yes) as the dependent variable, variables with statistical significance or potential clinical relevance from univariable analysis (PLI, GI, hard food biting, bruxism, tooth type, bracket material, intermaxillary elastics) were included in a multivariable GEE Logistic regression model. The patient was specified as the clustering unit with an exchangeable working correlation structure. The results identified PLI, GI, hard food biting, bruxism, molar tooth position, ceramic bracket, and use of intermaxillary elastics (all P < 0.001) as independent risk factors for bracket bonding failure. Premolar and incisor positions showed no significant difference compared to canines (all P > 0.05), as seen in Table 3.

Table 3.

Multivariable GEE Logistic regression analysis for bracket bonding failure (n = 5,808)

Variable	β	SE	Wald	OR	95% CI	P
PLI	0.186	0.093	4.006	1.205	1.003–1.445	0.045
GI	0.234	0.096	5.966	1.264	1.047–1.525	0.015
Hard food biting (Yes vs. No)	0.373	0.121	9.523	1.452	1.146–1.840	0.002
Bruxism (Yes vs. No)	0.661	0.116	32.294	1.937	1.542–2.434	< 0.001
Tooth type: Molar	0.353	0.149	5.649	1.423	1.064–1.904	0.017
Tooth type: Premolar	0.031	0.133	0.054	1.031	0.795–1.339	0.816
Tooth type: Incisor	-0.059	0.116	0.259	0.943	0.752–1.183	0.611
Bracket material: Ceramic (vs. Metal)	0.286	0.111	6.604	1.332	1.070–1.657	0.010
Intermaxillary elastics (Yes vs. No)	0.357	0.086	17.099	1.430	1.207–1.694	< 0.001

Note: PLI Plaque index, GI Gingival index, SE Standard error, OR Odds ratio, CI Confidence interval

Discrimination, calibration, and internal validation of the bracket bonding failure risk prediction model

The ROC curve for the Logistic prediction model showed an AUC of 0.613 (95% CI: 0.588–0.633), indicating modest discrimination (Fig. 1A). The calibration analysis, assessed by the HL test, yielded a P-value of 0.419, and the calibration plot demonstrated good agreement between predicted and observed probabilities (Fig. 1B), suggesting reliable overall model fit.

Fig. 1.

Model performance validation. Note: (A) ROC curve of the predictive model; (B) Calibration plot of the predictive model; (C) Bootstrap validation of model performance. ROC, Receiver operating characteristic; AUC, Area under the curve; CI, Confidence interval; HL test, Hosmer-Lemeshow test

To further assess model robustness and generalizability, internal validation was performed using Bootstrap resampling (1,000 repetitions). The Bootstrap average optimism was 0.0103, resulting in a corrected AUC of 0.6027, which was close to the original AUC, indicating good model stability. The distribution of Bootstrap AUC values also showed minimal performance fluctuation across different samples (Fig. 1C).

Development of a web-based dynamic prediction tool

Based on the final multivariable Logistic regression model, an interactive online dynamic prediction tool was developed using R (version 4.3.2) and the shiny application package (RStudio, Posit Software). The model was fitted using the glm() function. Input variables included oral hygiene indicators (PLI, GI), detrimental oral habits (hard food biting, bruxism), tooth type, bracket material, and use of intermaxillary elastics. All pre-processed variables served as predictors fed into the server-side logic module of the shiny application, where the model calculated the real-time bonding failure probability.

The tool interface was built using shiny layout functions [fluidPage(), sidebarLayout(), mainPanel()]. User input was collected via numericInput() and selectInput() widgets, and predictions were generated using the predict() function. Outputs were displayed in a card-based format, including the predicted failure risk (as a percentage), a corresponding risk level (low, medium, high) automatically categorized based on the risk value, and pertinent clinical recommendations. Embedded cascading style sheet (CSS) enhanced the visual presentation for clinical and publication purposes (Fig. 2). This tool visualizes and structures the model calculation process, enhancing its interpretability in chairside scenarios. It enables clinicians to assess bonding failure risk for individual tooth positions in real-time, facilitating adjustments in bonding strategy, follow-up frequency, or oral hygiene instructions. Although the model’s discriminative ability is moderate, the tool holds clear value for clinical risk alerting, individualized strategy formulation, and patient communication.

Fig. 2.

Schematic diagram of the dynamic prediction tool for fixed bracket bonding failure risk

Discussion

Bracket bonding failure during fixed orthodontic treatment is a common adverse event impacting treatment progress and patient experience, influenced by multifaceted factors including patient oral hygiene, occlusal forces on the tooth, bracket material, operator technique, and individual habits [7, 8]. While prior studies have explored isolated factors contributing to bracket failure, systematic evaluations based on clinical RWD are relatively scarce, and individualized risk prediction models are lacking. Utilizing large-scale RWD with the tooth as the analysis unit and the patient as the clustering unit, this study systematically evaluated multidimensional determinants of bonding failure and, for the first time, developed a dynamic risk tool combining GEE for inference and Logistic regression for prediction.

The findings confirmed a significant association between patient oral hygiene status and bonding failure. Both PLI and GI were independent risk factors, consistent with previous research indicating that plaque/biofilm accumulation and gingival inflammation contaminate the adhesive interface, weakening bond strength and increasing failure probability [9]. Studies have noted that inadequate moisture control, gingival crevicular fluid seepage, and plaque coverage can significantly reduce composite resin bonding efficacy [10–12]. Our data reinforce the importance of oral hygiene in the reliability of fixed appliances and suggest the need for enhanced follow-up and hygiene instruction for patients with poor oral health.

Hard food biting and bruxism were also confirmed as significant risk factors. Previous research indicates that bruxism imposes repetitive non-physiological overload, subjecting the bracket-adhesive interface to sustained shear or torsional forces [13, 14], predisposing it to failure. Our study similarly observed a markedly increased failure risk in patients with bruxism. Additionally, hard food biting introduces localized instantaneous high stress, which can disrupt the adhesive interface. The mechanism for both relates to an imbalance in forces within the tooth-bracket-adhesive system [15, 16], suggesting the need for stronger bonding protocols or protective measures in such patients.

Regarding tooth-related factors, our study found molar positions to be more prone to failure, aligning with existing literature. Challenges in saliva control, higher masticatory forces, and limited operative access at molar sites are primary contributors to bonding failure [17]. Premolar and incisor positions did not show significant differences in our study, possibly due to relatively lower and more evenly distributed orthodontic force loads compared with molar regions. In fixed appliance therapy, incisors and premolars are primarily subjected to controlled tipping or sliding forces during alignment, whereas molars experience higher occlusal loads and greater masticatory forces, which may increase stress at the bracket–adhesive interface. These biomechanical differences may partly explain the higher bonding failure observed at molar positions.

In addition to tooth position and bracket material, previous studies have shown that adhesive type and bonding technique may also influence bracket failure rates. Conventional light-cured composite resins generally demonstrate higher bond strength and lower failure rates compared with glass ionomer-based adhesives, particularly in regions exposed to higher functional loads [18]. With respect to bonding technique, both direct and indirect bonding methods have been widely used in orthodontic practice, and most clinical studies report comparable bracket failure rates when standardized bonding protocols are followed. In the present real-world retrospective study, detailed information on specific adhesive systems and bonding techniques was not consistently available and therefore could not be included as independent variables. This reflects routine clinical practice and represents a limitation. Nevertheless, all bonding procedures were performed following standardized protocols within the same clinical team, which may have mitigated technique-related variability.

Concerning bracket material, ceramic brackets carried a higher failure risk compared to metal brackets, consistent with prior reports. The mechanical retention of ceramic bracket bases to adhesive is often inferior to the mesh base of metal brackets, and their brittleness may lead to stress concentration and interfacial separation. Furthermore, ceramic brackets are more sensitive to tooth surface preparation and light-curing conditions [19], potentially elevating failure risk.

Among treatment-related factors, the use of intermaxillary elastics significantly increased the risk of bracket failure. Intermaxillary elastics often generate additional continuous tensile or oblique forces that may act specifically on a bracket, subjecting the adhesive interface to greater stress and increasing debonding probability [20, 21]. This finding underscores the need for close monitoring of bracket stability in patients using intermaxillary elastics.

The predictive model developed in this study had an AUC of 0.613 and a Bootstrap-corrected AUC of 0.6027, indicating limited discriminative ability. This, however, is a common observation in bracket failure prediction research. Firstly, bracket failure is a typical low-event-rate and multifactorial outcome with small effect sizes. Many crucial variables (e.g., operator variability, salivary moisture, immediate force application) are difficult to quantify fully in real-world settings, limiting model explanatory power. Secondly, random error and variability in patient compliance may attenuate model performance. Therefore, AUC values between 0.6 and 0.7 are not uncommon in orthodontic research. Importantly, our model demonstrated good calibration (HL P = 0.419) and stable internal validation, indicating reliable probability estimates. The value of a predictive model lies in risk alert and supporting clinical decision-making rather than precise diagnosis; thus, even with a moderate AUC, its clinical utility persists.

Study significance and limitations

This study developed a risk assessment model for fixed bracket bonding failure based on clinical RWD and created a web-based dynamic prediction tool. This tool enables real-time quantification and visualization of failure risk for individual tooth positions, providing chairside decision support. Clinically, this prediction tool may support early identification of tooth-level bonding failure risk, allowing orthodontists to proactively adjust treatment strategies. For high-risk tooth positions or patients, clinicians may consider intensified oral hygiene reinforcement, selection of more retentive bracket materials, modification of force application strategies, or closer follow-up scheduling. These targeted interventions may help reduce unplanned bracket debonding events, thereby minimizing treatment interruptions, additional chairside time, and the need for repeated rebonding procedures. Ultimately, such risk-guided management may contribute to shorter overall treatment duration, improved treatment efficiency, and better patient experience.

However, this study has limitations. Firstly, as a single-center retrospective study, it may be subject to data recording inconsistencies and selection bias. Secondly, several potentially important factors (e.g., operator experience, adhesive batch, light-curing protocol, salivary moisture, detailed tooth surface treatment) could not be incorporated into the model. Thirdly, the model’s discriminative ability is limited, suggesting the need to incorporate more variables or employ more complex modeling techniques (e.g., machine learning) in future research. Finally, the lack of external validation necessitates further testing of the model’s generalizability using multi-center data.

Furthermore, several patient-related behavioral factors that may influence bracket bonding failure were not systematically captured in this retrospective dataset. These include specific dietary habits, such as frequent consumption of sticky foods (e.g., chocolates or chewing gum), as well as oral parafunctional or pressure-related habits, including thumb sucking, mouth breathing, lip biting, nail biting, unilateral chewing, and smoking. These behaviors may increase mechanical stress on the bracket–adhesive interface or compromise bonding integrity, potentially affecting failure risk. The absence of these variables reflects the inherent limitations of real-world retrospective data and should be considered when interpreting the study findings.

Conclusion

Fixed orthodontic bracket bonding failure is influenced by multiple factors including oral hygiene status, detrimental habits, tooth position type, bracket material, and treatment modality. Based on RWD, this study developed a multivariable predictive model for bracket bonding failure and an online dynamic prediction tool, which can be used for risk alerting and formulating individualized orthodontic strategies. Although the model exhibits limited discriminative ability, it demonstrates good calibration and stability, offering clinicians a practical auxiliary decision-support tool.

Authors’ contributions

Yonghua Deng: conceptualization, methodology, writing-original draft Ni Quan: data curation, formal analysis Shumei Song: resources, visualization Xifei Liu: validation, supervision Zhiyong Hu: investigation, manuscript review and editing.

Funding

None.

Data availability

The data supporting this publication is available from the corresponding author upon request.

Declarations

Ethics approval and consent to participate

The study protocol was reviewed and approved by the Ethics Committee of Changsha Stomatological Hospital (Approval No.: 2025 − 173). The study adhered strictly to the ethical principles of the Declaration of Helsinki. Given the retrospective design and no compromise to patient privacy, the Ethics Committee granted a waiver for informed consent. All patient data were anonymized and de-identified prior to analysis.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.