https://orcid.org/0000-0003-3819-2864
Figures
Abstract
Background
Advances in digital orthodontics and artificial intelligence (AI) planning have the potential to enhance treatment precision, but randomized evidence based on the Peer Assessment Rating (PAR) index remains limited.
Methods
In this single-center, parallel-group randomized controlled trial registered retrospectively in the Chinese Clinical Trial Registry (ChiCTR2500108499), 140 patients aged 12–35 years with Angle Class I malocclusion were randomized to receive an AI-assisted digital workflow (Digital and AI group) or conventional fixed appliances (Conventional group). PAR scores were assessed at baseline (T0), 6-month intervals (T1), and immediately after treatment completion (T2) by calibrated, blinded examiners following British Standards Institute criteria. Analyses followed the intention-to-treat principle, applying independent t-tests, χ²/Fisher’s exact tests, repeated-measures mixed-effects models, and multivariable linear regression. Effect sizes were expressed as mean difference (MD) or relative risk (RR) with 95% confidence intervals (CI).
Results
Baseline PAR scores did not differ significantly between groups (MD = 0.63, 95% CI: –0.13 to 1.40; p = 0.105). At T2, the Digital and AI group had lower mean PAR scores (4.88 ± 0.45) than the Conventional group (7.81 ± 0.70; MD = 2.93, 95% CI: 2.73–3.13; p < 0.001). A higher proportion of patients in the Digital and AI group achieved ≥70% PAR reduction (82.9% vs 50.0%; RR = 1.66, 95% CI: 1.27–2.17; p < 0.001). Repeated-measures mixed-effects analysis showed significant effects of intervention, time, and their interaction (all p < 0.001), indicating different improvement trajectories between groups. Multivariable regression identified allocation to the Digital and AI group, higher baseline PAR, and younger age as independent predictors of greater PAR reduction. No severe adverse events occurred, and no participants were lost to follow-up between T0 and T2.
Conclusions
Under controlled trial conditions, the AI-assisted digital workflow produced greater short-term improvements in PAR-based occlusal outcomes than conventional fixed appliances. These findings suggest a potential benefit of integrating an AI-assisted digital system into orthodontic practice; however, conclusions are limited to short-term occlusal changes, and further multicenter studies with longer follow-up, patient-reported outcomes, and economic evaluation are warranted.
Citation: Xiaoting X, Ismail NA, Alhoot MA, Ying L, Abbas Helmi RA, Song Q, et al. (2026) Randomized controlled trial comparing AI-assisted digital and conventional orthodontics: Superior PAR reduction and occlusal outcomes. PLoS One 21(5): e0347499. https://doi.org/10.1371/journal.pone.0347499
Editor: Spyridon N. Papageorgiou, University of Zurich, SWITZERLAND
Received: September 8, 2025; Accepted: April 1, 2026; Published: May 4, 2026
Copyright: © 2026 Xiaoting et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: The datasets generated and analyzed during the current study are not publicly available due to ethical and legal restrictions related to patient privacy. De-identified research data may be made available upon reasonable request, subject to approval by the Ethics Committee of the Affiliated Hospital of North Sichuan Medical College. Researchers who meet the criteria for access to onfidential data may submit requests via email at ethics@nsmc.edu.cn.
Funding: The author(s) received no specific funding for this work. Management and Science University provided support for the publication of this article.
Competing interests: The authors declare no conflicts of interest related to this study.
1. Introduction
Malocclusion affects a substantial proportion of adolescents and adults worldwide, compromising masticatory efficiency, facial aesthetics and quality of life [1]. Although fixed-appliance therapy has matured technically, treatment planning and finishing still depend largely on the clinician’s judgement, which introduces variability in outcome and cost [2]. To reduce such subjectivity, the Peer Assessment Rating (PAR) index was introduced and has become a benchmark for quantifying changes in alignment and occlusal relationships [3,4].
Digital dentistry is reshaping this landscape. High-resolution intra-oral scanners produce faithful virtual casts within seconds, eliminating distortion from conventional impressions and facilitating chair-side simulations [5]. Combined with computer-aided design platforms, these scanners enable end-to-end digital workflows that begin with three-dimensional diagnostics and end with appliance delivery [6]. Additive manufacturing extends the continuum by fabricating customised brackets, aligners and auxiliaries with micron-level accuracy, supporting truly patient-specific mechanics [7–9]. While laboratory studies repeatedly confirm the linear precision of printed devices, clinical correlates measured with standardised occlusal indices remain sparse.
A second, equally disruptive layer of innovation stems from artificial intelligence (AI). Deep-learning algorithms now localise anatomical landmarks, forecast individual tooth trajectories and even recommend mid-course corrections, promising shorter treatment times and fewer refinements [10–12]. Landmark-segmentation networks such as Dilated-Tooth-Seg-Net achieve sub-millimetre errors on complex three-dimensional meshes, bringing fully automated progress monitoring within reach [13]. Nevertheless, most available reports are proof-of-concept explorations or retrospective case series; prospective evidence that links these technological gains to meaningful PAR improvements is conspicuously lacking [14,15].
Recent meta-analyses underline this gap, citing substantial heterogeneity among existing studies and calling for rigorously designed randomised controlled trials (RCTs) with harmonised outcome reporting [16]. Meanwhile, chair-side metal printing of bespoke appliances has become technically feasible, yet its clinical superiority over traditional laboratory methods remains unverified [17]. Real-time tele-monitoring platforms and algorithm-guided alerts may enhance compliance, but their influence on end-of-treatment occlusion has never been quantified in a controlled setting [18]. The same evidence deficit is evident in broader applications of patient-specific additive manufacturing for cranio-facial rehabilitation [19].
Against this background, the present parallel-group RCT was undertaken to compare AI-assisted digital orthodontics with conventional fixed-appliance therapy, using percentage reduction in PAR score as the primary end-point. It was hypothesised that algorithm-guided planning would deliver clinically and statistically superior occlusal outcomes, thereby providing a data-driven rationale for integrating digital-AI workflows into routine practice.
2. Materials and methods
2.1. Study design
This investigation was conducted as a single-center, prospective, parallel-group randomized controlled trial (RCT) comparing AI-assisted digital orthodontics with conventional fixed-appliance therapy. The study was designed in accordance with the CONSORT 2010 guidelines to ensure methodological rigor and transparent reporting.
Although the study was prospectively designed and conducted, the formal trial registration in the Chinese Clinical Trial Registry (ChiCTR2500108499) was completed after initiation of participant enrollment; therefore, the registration is classified as retrospective.
2.2. Participants
Between September 2023 and June 2025, 140 eligible participants were consecutively recruited from, and received treatment at, the Department of Stomatology, Affiliated Hospital of North Sichuan Medical College. Written informed consent was obtained from all participants, and consent for minors was provided by their legal guardians. Inclusion criteria were aged 12–35 years; Angle Class I malocclusion diagnosed according to the British Standards Institute (BSI) criteria [20]; no systemic disease affecting bone metabolism; no previous orthodontic treatment; and ability to attend scheduled follow-ups. Exclusion criteria included pregnancy, active periodontal disease, severe psychological disorders, or other contraindications identified by the attending orthodontist. All radiographic examinations followed institutional ALARA guidelines, and advanced imaging such as CBCT was obtained only when clinically indicated.
2.3. Sample size estimation
The required sample size was estimated using G*Power version 3.1 for a two-sided independent samples t-test. The calculation was based on a significance level of α = 0.05, statistical power (1–β) of 0.80, and a medium effect size of Cohen’s d = 0.60, selected according to effect magnitudes reported in previous randomized clinical trials evaluating PAR-based orthodontic outcomes. The unit of analysis was the change in total PAR score (T0–T2).
Applying the standard formula with critical values Zα/2 = 1.96 and Zβ = 0.84 indicated that at least 51 participants per group were required [21,22]. To account for an anticipated 10% attrition rate, the target sample size was increased to 70 participants per group, resulting in a total of 140 participants.
2.4. AI-assisted digital workflow
The digital workflow incorporated a suite of AI-assisted tools embedded within the 3D planning software (3Shape Ortho Analyzer) and the DentalMonitoring™ platform. Machine-learning algorithms were used to automate tooth segmentation, anatomical landmark identification, arch-form estimation, and preliminary tooth-movement simulations. These AI-generated outputs provided decision-support information for treatment planning, but all final diagnostic interpretations and movement prescriptions were determined by the treating orthodontists, ensuring that no autonomous decisions were made.
During active treatment, remote monitoring was performed using DentalMonitoring™, which employs convolutional-neural-network classifiers to detect issues such as bracket failure, wire displacement, plaque accumulation, or soft-tissue irritation. When threshold-based alerts were triggered, clinicians reviewed the AI-flagged images and arranged earlier clinical visits when necessary, primarily for scheduling purposes. These alerts supplemented, but did not replace, standard in-office evaluations, and did not trigger predefined treatment modifications, alter biomechanical strategies, or constitute a prespecified study outcome.
This workflow therefore represents a clinician-directed digital system augmented by AI-based decision-support tools, rather than a standalone or autonomous AI intervention. The combined effects of AI-assisted planning, CAD/CAM appliance fabrication, and remote monitoring could not be isolated within the present study design. CBCT imaging was performed in selected cases in both groups when clinically indicated for diagnostic clarification and treatment planning, while routine orthodontic assessment primarily relied on conventional two-dimensional radiography in accordance with standard orthodontic practice.
2.5. Randomization and allocation concealment
The randomization sequence was generated in advance by an independent statistician using the R package blockrand (version 1.5) in R version 4.2.2, with a fixed random seed to ensure reproducibility. Block randomization with a block size of four was applied to maintain balanced allocation between groups. Allocation concealment was achieved using sequentially numbered, opaque, sealed envelopes (SNOSE), prepared by a research assistant not involved in participant recruitment or assessment.
Envelopes were opened sequentially by the treating clinician only after all baseline measurements had been completed. Due to the nature of the intervention, blinding of participants and treating clinicians was not feasible. However, outcome assessors and data analysts remained blinded to group allocation, and all datasets were de-identified before statistical analysis.
2.6. Interventions
Digital and AI-assisted Orthodontics Group (DAOG): Patients in the DAOG underwent intraoral scanning with TRIOS 4 (3Shape, Denmark). Cone-beam computed tomography (CBCT) was obtained only when clinically indicated and in accordance with the ALARA principle. Digital models were processed in 3Shape Dental System (v2.22.0.0; 3Shape, Denmark) to generate virtual setups and biomechanical simulations. AI-assisted tools integrated in the planning software provided automated tooth segmentation, landmark identification, and preliminary movement suggestions; however, all AI-generated outputs were reviewed and modified as necessary by the treating orthodontist, who retained full decision-making authority. Custom brackets and auxiliaries were produced chairside using stereolithography-based 3D printing (Form 3B + ; Formlabs, USA). Remote monitoring was performed using DentalMonitoring™ (DentalMonitoring, Paris, France), which enabled automated screening of intraoral images for potential issues (e.g., bracket failure, hygiene concerns) and generated alerts when deviations exceeded software-defined thresholds. All alerts were reviewed and validated by clinicians before determining whether an in-person visit was required. This workflow represents an integrated digital system that combines CBCT-supported diagnostics, CAD/CAM appliance customization, and AI-assisted remote monitoring; therefore, the independent contribution of each component cannot be isolated within this trial design.
Conventional Orthodontics Group (COG): Patients in the COG received alginate impressions, panoramic radiographs, and lateral cephalograms for diagnosis and treatment planning. All patients were treated with 0.022-inch slot preadjusted edgewise appliances (MBT prescription) and attended follow-up visits at four-week intervals for appliance activation and adjustments.
All treatments were delivered by the same team of board-certified orthodontists with at least five years of clinical experience. Visit schedules, follow-up protocols, and retention regimens were standardized across both groups.
2.7. Retention protocol
“Both groups followed the institution’s standardized retention protocol, which consisted of maxillary vacuum-formed retainers and mandibular fixed lingual retainers. Importantly, all PAR assessments (T2) were completed immediately after appliance debonding and prior to retainer delivery. Consequently, retention appliances did not influence any measured outcomes and should not be considered as contributing to between-group differences.”
2.8. Outcome measures
The primary endpoint was the percentage reduction in Peer Assessment Rating (PAR) score between baseline (T0) and treatment completion (T2), calculated as:
Baseline PAR (T0) was obtained from pretreatment digital or plaster models prior to any orthodontic intervention, whereas final PAR (T2) was assessed immediately after debonding and before retainer delivery.
The PAR index is a validated occlusal assessment tool comprising five weighted components:(1) upper anterior alignment, (2) lower anterior alignment, (3) overjet, (4) overbite/open bite, and (5) buccal occlusion, each scored according to the British Standards Institute (BSI) criteria and multiplied by established weighting factors. To enhance reproducibility, all PAR component scores were assessed using standardized BSI criteria, with weighting factors applied according to the validated scoring protocol. A ≥ 70% reduction in total PAR is classified as “greatly improved” occlusion. To assist readers, annotated examples of the PAR components are provided in S1 Fig in S1 Date.
Secondary endpoints included PAR scores at 6 months (T1-6m) and 12 months (T1-12m), as well as scores of individual PAR components (anterior alignment, overjet/open bite, buccal occlusion, and midline). Three calibrated orthodontists, blinded to group allocation, independently assessed plaster or digital models using the BSI-revised PAR index. Inter-rater reliability was high (intraclass correlation coefficient, ICC = 0.92). Digital models were further evaluated with Dolphin Imaging software (version 11.95 SP2; Dolphin Imaging & Management Solutions, USA).
2.9. Data management
Clinical data and images were stored in an AES-256 encrypted electronic health record system, accessible only to authorized study personnel. Allocation codes were concealed from analysts until database lock. All datasets were de-identified before analysis, and no participant was lost to follow-up or excluded due to protocol deviations between T0 and T2. Adverse events were prospectively recorded throughout the study period, and no severe events occurred.
Extended follow-up data at 18 and 24 months were excluded because of incomplete availability across both groups. All cleaned and de-identified datasets used for statistical analysis are preserved on a secure institutional server and are available upon reasonable request in accordance with PLOS ONE data-sharing policies.
2.10. Statistical analysis
All analyses followed the intention-to-treat (ITT) principle. No participants were lost to follow-up between T0 and T2; therefore, the per-protocol (PP) analysis yielded results identical to the ITT analysis and is reported as a sensitivity check. Missing data (<5%) were handled using multiple imputation (m = 5, predictive mean matching), applied only to non-primary covariates with minor missingness.
Data normality was assessed with the Shapiro–Wilk test. Continuous variables (mean ± SD) were compared between groups using independent-samples t-tests and within groups using paired-samples t-tests. Categorical variables (n, %) were compared using χ² tests or Fisher’s exact test when expected cell counts were <5.
Longitudinal changes in PAR scores were evaluated using a repeated-measures mixed-effects model, with group, time, and the group × time interaction specified as fixed effects. This approach was selected instead of repeated-measures ANOVA because it accommodates missingness, relaxes sphericity assumptions, and provides unbiased estimates under ITT assumptions.
Exploratory subgroup and interaction analyses (e.g., age × treatment) were performed as prespecified but underpowered analyses and interpreted cautiously.
Effect sizes were reported as Cohen’s d for continuous outcomes and relative risk (RR) with 95% confidence intervals for categorical outcomes. All statistical analyses were conducted using SPSS Statistics 29.0 (IBM, USA) and R version 4.2.2 (R Foundation for Statistical Computing, Austria). All tests were two-sided, and p-values <0.05 were considered statistically significant.
2.11. Ethics approval and trial registration
This trial was approved by the Ethics Committee of the Affiliated Hospital of North Sichuan Medical College, China (File Number: 2023ER435−1), and endorsed by the Ethics Committee of Management and Science University (MSU), Malaysia. The study protocol specifying the primary and secondary outcome measures was finalized prior to participant enrolment and approved by the institutional ethics committee (protocol version 1.0, July 2, 2023). The study was conducted in accordance with the Declaration of Helsinki and relevant national regulatory requirements. Written informed consent was obtained from all participants prior to enrollment, with consent for minors provided by their legal guardians. The trial was retrospectively registered in the Chinese Clinical Trial Registry (ChiCTR2500108499) after the first participant was enrolled in September 2023.
3. Results
3.1. Participant flow
Between September 2023 and June 2025, a total of 162 individuals were assessed for eligibility. Of these, 22 were excluded (15 did not meet inclusion criteria, 5 declined to participate, and 2 for other reasons). The remaining 140 participants were randomized in a 1:1 ratio to the Digital and AI group (n = 70) or the Conventional group (n = 70). All randomized participants received the allocated intervention, completed follow-up at T1-6m, T1-12m, and T2 (completion). No participant was lost to follow-up, and no protocol deviations occurred during the study period. All participants were included in the intention-to-treat analyses. The participant flow is illustrated in Fig 1.
A total of 162 individuals were screened for eligibility. Twenty-two were excluded (15 not meeting inclusion criteria, 5 declined to participate, and 2 for other reasons). The remaining 140 were randomized equally to the Digital and AI group (n = 70) or the Conventional group (n = 70). All participants completed the allocated intervention, attended follow-up visits at T1-6m, T1-12m, and T2, and were included in the final analyses.
3.2. Baseline characteristics
Baseline demographic and clinical characteristics of the participants are presented in Table 1. The two groups were broadly comparable in terms of age, sex distribution, and baseline PAR scores. A small numerical difference was observed in overjet values; however, this difference was clinically negligible and not expected to influence subsequent analyses.
3.3. Primary outcome: Total PAR reduction
From baseline (T0) to treatment completion (T2), total PAR scores decreased in both groups, with lower post-treatment scores observed in the Digital and AI group compared with the Conventional group (4.88 ± 0.45 vs 7.81 ± 0.70; mean difference [MD] = −2.93, 95% CI −3.13 to −2.73; p < 0.001; Table 2). These between-group differences reflect the integrated digital workflow rather than the isolated contribution of AI algorithms, which could not be separated within the constraints of this trial design. A repeated-measures mixed-effects model showed significant effects of intervention (p < 0.001), time (p < 0.001), and their interaction (p < 0.001), indicating distinct improvement trajectories between groups.
The proportion of participants achieving at least a 70% reduction in PAR was significantly higher in the Digital and AI group (82.9%) than in the Conventional group (50.0%), corresponding to a risk ratio [RR] = 1.66 (95% CI 1.27–2.17; p < 0.001; Table 3, Fig 2). These results indicate differential occlusal improvement between groups but do not imply enhanced treatment efficiency, chairside time reduction, or workflow acceleration, as such metrics were not collected in this study.
(A) Stacked bar charts showing the proportions of participants achieving significant improvement (≥70%), partial improvement (30–70%), and no improvement (<30%) in the Digital and AI group and the Conventional group at T1-6m, T1-12m, and T2. (B) Time trend of the proportion of participants achieving significant improvement (≥70%) in PAR scores from baseline (T0) to T2. The Digital and AI group consistently showed higher rates than the Conventional group. Error bars represent 95% confidence intervals, and p values indicate between-group differences at each timepoint. This figure reflects differences in occlusal improvement only and should not be interpreted as evidence of enhanced treatment efficiency, reduced chairside time, or accelerated workflow, as such metrics were not collected in this study. Repeated-measures mixed-effects analysis demonstrated significant intervention (p < 0.001), time (p < 0.001), and group × time interaction effects (p < 0.001).
To further illustrate individual patient-level distributions, scatter plots of baseline versus post-treatment PAR scores at T2 are provided in S2 Fig in S1 Data. The plots demonstrate that most patients in the Digital and AI group clustered in the “greatly improved” region, whereas patients in the Conventional group were more widely distributed across the “partial improvement” and “no improvement” categories.
All PAR values and improvement metrics were rechecked for internal consistency, and corrected values are reported in Table 2 and Table 3.
3.4. Secondary outcomes: PAR component scores
Across follow-up, both groups demonstrated reductions in PAR component scores. At treatment completion (T2), the Digital and AI group achieved significantly lower scores than the Conventional group in the upper anterior segment (0.65 ± 0.53 vs 1.39 ± 0.59; p < 0.001), lower anterior segment (1.02 ± 0.50 vs 1.45 ± 0.65; p < 0.001), overbite/open bite (1.07 ± 0.55 vs 1.69 ± 0.52; p < 0.001), and buccal occlusion (2.01 ± 1.23 vs 3.34 ± 1.37; p < 0.001). No significant between-group difference was observed for midline alignment at T2 (0.38 ± 0.30 vs 0.42 ± 0.35; p = 0.469).
These findings are summarized in Table 4 and illustrated in Fig 3, which depict the trajectories of PAR component improvements over time for both groups.
(A) Upper anterior; (B) Lower anterior; (C) Overbite/Open bite; (D) Overjet; (E) Midline alignment; (F) Buccal occlusion. Data are presented as mean ± SD. Significant differences between groups at T2 are marked in each panel.
Consistent with the primary outcome, these between-group differences should be interpreted as the result of the integrated digital workflow—including advanced imaging, appliance customization, and remote monitoring—rather than being attributed to AI algorithms alone. The component-level improvements indicate differential occlusal changes between groups but do not provide evidence regarding treatment efficiency, chairside workload, stability, or other clinical outcomes not assessed in this trial.
3.5. Subgroup analysis of PAR score improvement by age
Subgroup analyses were conducted to explore whether treatment effects varied across different age categories (12–18, 19–25, and 26–35 years) (Table 5, Fig 4).In the youngest subgroup (12–18 years), patients in the Digital and AI group achieved a significantly higher PAR score improvement rate compared with those in the Conventional group (82.4 ± 6.2% vs 72.7 ± 6.4%; mean difference = 9.7%, 95% CI 7.6–11.8%; p < 0.001). Similar differences were observed in the 19–25 years subgroup (79.2 ± 5.7% vs 69.3 ± 5.9%; mean difference = 9.9%, 95% CI 7.9–11.9%; p < 0.001) and the 26–35 years subgroup (75.5 ± 6.5% vs 66.1 ± 6.3%; mean difference = 9.4%, 95% CI 7.0–11.8%; p < 0.001).
Violin plots show the distribution of PAR improvement rates (%) at T2 for age subgroups (12–18, 19–25, 26–35 years). Each dot represents an individual participant. Box plots indicate interquartile ranges with medians, and black diamonds represent means with 95% confidence intervals. Horizontal lines and p values indicate between-group differences (independent-samples t-test). Subgroup analyses were exploratory and underpowered; therefore, non-significant interaction results should not be interpreted as evidence of uniform treatment effects across age groups.
A formal test of the group × age interaction did not reach statistical significance (p = 0.74). Because the trial was powered only for the primary between-group comparison and not for detecting interaction effects, the subgroup analyses are exploratory and underpowered. Therefore, the non-significant interaction should not be interpreted as evidence of uniform effects across age groups. Younger patients tended to show slightly higher absolute improvement rates, but these trends require cautious interpretation, and the Digital and AI group showed numerically greater improvement than the Conventional group across all age categories.
3.6. Multivariable regression analysis
Multivariable linear regression analyses were conducted to identify independent predictors of PAR score improvement (Table 6, Fig 5). The analysis included treatment modality (Digital and AI versus Conventional), baseline PAR score, age group, and gender as independent variables.
Data are shown as β-coefficients with 95% confidence intervals (CI). Predictors included treatment modality, baseline PAR score, age group, and gender. Significant predictors are indicated when CIs do not cross zero.
Treatment modality was a highly significant predictor, with the Digital and AI group showing greater improvement compared with the Conventional group (β = 8.62, 95% CI: 6.75–10.49; p < 0.001). Baseline PAR scores were also positively associated with improvement (β = 0.32, 95% CI: 0.11–0.53; p = 0.003).
Age group significantly influenced improvement, with lower improvements observed in older subgroups compared with the 12–18 years reference group (β = −1.25, 95% CI: −2.35 to −0.15; p = 0.026 for 19–25 years; β = −4.83, 95% CI: −6.17 to −3.49; p < 0.001 for 26–35 years). Gender was not significantly associated with PAR score improvement (β = 0.46, 95% CI: −0.68 to 1.60; p = 0.429).
3.7. Sensitivity analysis
Sensitivity analyses were performed after excluding extreme outliers, defined as values exceeding 1.5 times the interquartile range (IQR), to assess the robustness of the findings. In this sensitivity analysis, final PAR scores at treatment completion (T2) were used as the outcome measure, rather than percentage PAR reduction, to evaluate whether between-group differences persisted after removal of extreme values. Results are summarized in Table 7.
After exclusion, the mean final PAR score at T2 remained significantly lower in the Digital and AI group compared with the Conventional group (5.02 ± 0.37 vs 7.66 ± 0.59; mean difference = −2.64, 95% CI −2.80 to −2.48; p < 0.001). The proportion of patients achieving ≥70% improvement also remained consistently higher in the Digital and AI group (85.1%) than in the Conventional group (53.7%; p < 0.001), with no patients in the Digital and AI group falling into the < 30% improvement category.
Paired boxplots with individual data points further illustrated within-group changes before and after outlier exclusion (Fig 6). Both groups demonstrated significant internal improvements after excluding outliers (Conventional: p < 0.001; Digital and AI: p < 0.01). These findings indicate that the primary results were robust to sensitivity testing. Taken together, these sensitivity analyses confirm that the primary findings are robust to exclusion of extreme values and to alternative outcome specifications, and are consistent with the main analyses based on percentage PAR reduction.
Paired boxplots illustrating PAR improvement rates (%) before and after outlier exclusion in the Digital and AI group and the Conventional group. Each dot represents an individual participant. Lines represent paired observations. Statistical significance of within-group differences was determined using paired t-tests (**p < 0.01, **p < 0.001).
3.8. Adverse events
No serious adverse events were reported in either group during the study period. Minor events, such as bracket debonding and wire irritation, occurred occasionally and were managed with routine chairside adjustments. The incidence of these events did not differ significantly between the Digital and AI group and the Conventional group. No protocol deviations occurred, and no participants were lost to follow-up between T0 and T2.
4. Discussion
This parallel-group randomised controlled trial provides prospective evidence that a fully digital, AI-integrated orthodontic workflow is associated with clinically relevant reductions in PAR scores. Compared with conventional fixed-appliance therapy, the digital–AI arm achieved an average PAR improvement that exceeded the minimal clinically important difference previously proposed for high-quality treatment finishes [20]. The superiority was uniform across anterior alignment, overjet, overbite/open-bite and buccal occlusion, and it persisted after multivariable adjustment for baseline severity, sex and age—findings that expand earlier cohort observations linking appliance accuracy to final occlusal quality [13,23,24]. Although statistically significant, these differences should be interpreted in light of established clinical relevance thresholds. However, these findings should be interpreted strictly within the scope of PAR-based occlusal assessment and within the context of an integrated digital workflow involving intraoral scanning, virtual setup, customised appliance fabrication, and AI-assisted monitoring. As the PAR index does not capture treatment efficiency, patient-reported outcomes, or biological side effects, the observed differences reflect short-term occlusal outcomes rather than overall treatment superiority or isolated AI effects.
High-resolution intraoral scanners capture full-arch geometry with minimal distortion, producing virtual casts that faithfully replicate occlusal relationships. These digital datasets allow orthodontists to visualize spatial discrepancies, simulate biomechanics, and formulate treatment plans in true three dimensions with sub-millimetre precision [6,25].Complementing these hardware advancements, modern artificial intelligence modules enhance treatment by automating cephalometric and dental landmark identification, predicting individualized tooth movement trajectories based on baseline malocclusion, and issuing real-time alerts when deviations exceeded predefined clinical thresholds, such as delayed tooth movement trajectories, potential appliance failures (e.g., bracket detachment), or discrepancies in treatment progress. These systems support early intervention, minimize cumulative errors, and reduce chairside time through remote supervision [26–28]. Nonetheless, the present study did not measure appointment frequency, workflow efficiency, or chairside time; therefore, no conclusions regarding improved efficiency can be drawn from our data despite the mechanistic plausibility.
Deep-learning architectures such as Dilated-Tooth-Seg-Net demonstrate sub-millimetre segmentation errors on complex dental meshes, further reducing human variability in model analysis [29,30]. Collectively, these innovations enhance precision at every stage of treatment—from diagnosis to finishing—reducing the need for mid-course refinements and accounting for the ΔPAR improvements observed in this study. Still, the observed improvements likely reflect the combined influence of high-resolution imaging modalities, including CBCT when clinically indicated, CAD/CAM-based appliance customization, digital planning, and clinician oversight; because these elements operate together as an integrated workflow, the present trial cannot isolate the independent contribution of AI.
Our data align with, yet substantively extend, earlier reports. Tele-monitoring systems were shown to reduce appointment frequency but did not quantify occlusal quality [18,31]; automated PAR scoring proved reproducible but was not linked to treatment strategy [3,32]. By integrating virtual planning, additive manufacturing, and AI-assisted monitoring within a controlled study framework, the present trial indicates that this integrated approach may contribute to procedural efficiency and improved occlusal outcomes. Beyond orthodontics, similar engineering principles have enhanced results in craniofacial rehabilitation and endoscopic surgery, highlighting the cross-disciplinary effectiveness of patient-specific design [19,33,34]. The presence of active digital remote monitoring may also have enhanced patient adherence, introducing a potential Hawthorne effect that could partly influence the observed group differences.
Sensitivity analyses indicated that the findings were robust. Excluding extreme responders (> 1.5 × IQR) had no significant impact on the effect estimates, aligning with recommended statistical practices for analyzing complex survey data [35]. Age-stratified results further showed that adolescents tended to show greater improvements, which may be related to higher alveolar turnover at younger ages [36]—while adults still enjoyed clinically relevant gains, suggesting that algorithm-guided planning can partly compensate for slower periodontal remodelling. Because the study was not powered to detect age-by-treatment interaction effects, all age-stratified findings should be regarded as exploratory and interpreted with caution. Biological plausibility is supported by evidence showing that cytokine networks enhance bone remodeling in response to digitally guided force application [37] and that systemic conditions such as hypophosphatasia modulate hard-tissue response [38,39]. These mechanisms remain theoretical in the context of this study and were not directly evaluated; thus, they should not be interpreted as causal explanations for the observed group differences.
From a clinical viewpoint, the digital–AI workflow may contribute to shorter chair time, reduced laboratory inventories, and potentially higher patient satisfaction—outcomes echoed by surveys on three-dimensional facial-scan accuracy and AI-driven patient engagement platforms [14,40]. Economic evaluations are underway; however, in-office metal printing already matches commercial turnaround times while allowing for rapid design iterations [17,41]. These efficiencies are particularly attractive for complex malocclusions and geographically distant patients, populations traditionally characterised by extended treatment times and higher dropout rates. Yet, because treatment time and patient-reported outcomes were not collected in this trial, such clinical advantages cannot be inferred from the present findings. Digital workflows may require additional equipment investment, training, and administrative workload, which were not assessed in this trial. Retention appliances were delivered only after outcome assessment; therefore, the retention protocol did not influence the measured PAR results and should not be interpreted as a source of bias between groups.
Despite its strengths—including a prospective design, blinded assessments, strict adherence to CONSORT guidelines, and advanced image registration protocols—the study has limitations. Because retention appliances were delivered only after outcome assessment, they did not influence the PAR-based comparisons and should not be regarded as a source of bias. It was conducted at a single academic centre; thus, operator expertise may partially account for outcomes. Differences in imaging modalities (CBCT versus 2D radiography) may introduce information bias and could have influenced diagnostic precision and treatment planning; therefore, the observed differences should be interpreted as reflecting an integrated digital workflow rather than the isolated effect of AI-assisted planning. Furthermore, retention follow-up in this study was limited to immediate post-treatment records, and no long-term assessment of post-treatment stability was performed, which restricts the interpretation of outcome durability. The 12-month follow-up does not capture long-term stability or root-resorption risk, issues of special relevance for patients with syndromic cranio-facial conditions or down-syndrome-associated periodontal vulnerability [42]. Biomarkers of bone metabolism were not measured; future work employing coarse-to-fine 3-D tooth-segmentation pipelines [43] and salivary assays could elucidate mechanistic pathways. Finally, although logistic-regression techniques appropriate for complex data were applied [35], residual confounding by psychosocial or operator variables cannot be excluded.
Future studies should therefore adopt multicentre designs, incorporate cost-utility analyses and extend observation windows to capture post-retention stability. Integrating hyperspectral bone-turnover imaging with algorithm-guided force systems may reveal how digital mechanics interact with cytokine-mediated remodelling [37]. Parallel advances in additive manufacturing for maxillofacial prosthodontics [19,44,45] suggest that an engineering-biology pipeline will continue to personalise orthodontic intervention.
5. Conclusions
Under controlled trial conditions, the AI-assisted digital orthodontic workflow produced statistically significant and clinically interpretable short-term improvements in PAR scores compared with conventional fixed appliances. These benefits likely reflect the combined influence of high-precision imaging, customized appliance fabrication, and AI-supported monitoring, and cannot be attributed to AI alone.
Because outcomes were assessed immediately after treatment and only Class I cases from a single center were included, the durability and generalizability of these findings remain limited. Future multicenter studies with extended post-retention follow-up and cost–utility analyses are required to determine the long-term clinical value of this integrated digital workflow.
Supporting information
S1 Data. S1 Fig. PAR scoring scheme illustrated with standardized line drawings.
Schematic illustration of the five components of the Peer Assessment Rating (PAR) index used to evaluate occlusal outcomes. (A) Upper anterior alignment and (B) lower anterior alignment represent crowding or spacing of the maxillary and mandibular anterior segments, respectively. (C) Overjet reflects the anteroposterior relationship of the anterior teeth. (D) Overbite/open bite depicts the vertical relationship of the anterior teeth. (E) Buccal occlusion illustrates posterior occlusal relationships, incorporating anteroposterior, vertical, and transverse dimensions. S2 Fig. Scatter plots of baseline versus post-treatment PAR scores at T2. Each dot represents an individual participant. Regions indicate categories of change (worse/no difference, improved, greatly improved). Most participants in the Digital and AI group clustered within the “greatly improved” region, whereas those in the Conventional group were more widely distributed across the “partial improvement” and “no improvement” categories. S1 File. Study protocol. S2 File. CONSORT 2010 checklist of information to include when reporting a randomised trial.
https://doi.org/10.1371/journal.pone.0347499.s001
(ZIP)
References
- 1. Graf I, Bock NC, Bartzela T, Röper V, Schumann U, Reck K, et al. Quality of orthodontic care—A multicenter cohort study in Germany: Part 1: Evaluation of effectiveness of orthodontic treatments and predictive factors. Journal of Orofacial Orthopedics. 2022;83.
- 2. Chambers DW, Thakkar D. Consistency of orthodontists’ clinical decisions: A systematic review, meta-analysis, and theory development. Am J Orthod Dentofacial Orthop. 2022;161(4):497-509.e4. pmid:35101324
- 3. Hannebauer A, Wesemann C, Bartzela T, Bister D, Bumann A. Comparison of automated digital Peer Assessment Rating compared with measurements performed by orthodontists, dental students, and assistants using plaster, additive manufactured, and digital models. Eur J Orthod. 2022;44(5):588–94. pmid:35731637
- 4. Richmond S, Shaw WC, O’Brien KD, Buchanan IB, Jones R, Stephens CD, et al. The development of the PAR Index (Peer Assessment Rating): Reliability and validity. Eur J Orthod. 1992;14(2):125–39. pmid:1582457
- 5. Christopoulou I, Kaklamanos EG, Makrygiannakis MA, Bitsanis I, Perlea P, Tsolakis AI. Intraoral scanners in orthodontics: A critical review. International Journal of Environmental Research and Public Health. 2022.
- 6. Leonardi R. Digital technologies: From scientific to clinical applications in orthodontic and dental communities. Applied Sciences (Switzerland). MDPI; 2022.
- 7. Tsolakis IA, Gizani S, Tsolakis AI, Panayi N. Three-dimensional-printed customized orthodontic and pedodontic appliances: A critical review of a new era for treatment. Children. 2022.
- 8. Thurzo A, Urbanová W, Novák B, Waczulíková I, Varga I. Utilization of a 3D Printed Orthodontic Distalizer for Tooth-Borne Hybrid Treatment in Class II Unilateral Malocclusions. Materials (Basel). 2022;15(5):1740. pmid:35268969
- 9. Xiaoting X, Ismail AN, Ying L, Abbas RA, Song Q, Ruiting L. A comparative study on the effects of intelligent diagnosis in orthodontics and conventional orthodontic treatment. Frontiers in Health Informatics. 2024;13.
- 10. Nordblom NF, Büttner M, Schwendicke F. Artificial intelligence in orthodontics: Critical review. J Dent Res. 2024;103(6):577–84. pmid:38682436
- 11. Kazimierczak N, Kazimierczak W, Serafin Z, Nowicki P, Nożewski J, Janiszewska-Olszowska J. AI in orthodontics: Revolutionizing diagnostics and treatment planning-A comprehensive review. J Clin Med. 2024;13(2):344. pmid:38256478
- 12. Xiaoting X, Ismail NA, Abbas RA, Ruiting L. AI-Assisted digital orthodontics enhances predictability of craniofacial hard and soft tissue remodeling: Evidence from a randomized controlled trial. Perinat J. 2025;33(1).
- 13. Krenmayr L, von Schwerin R, Schaudt D, Riedel P, Hafner A. Dilated tooth SegNet: Tooth segmentation network on 3D dental meshes through increasing receptive vision. Journal of Imaging Informatics in Medicine. 2024;37.
- 14. Tomášik J, Zsoldos M, Oravcová Ľ, Lifková M, Pavleová G, Strunga M. AI and face-driven orthodontics: A scoping review of digital advances in diagnosis and treatment planning. AI. 2024;:158–76.
- 15. Al Turkestani N, Bianchi J, Deleat-Besson R, Le C, Tengfei L, Prieto JC, et al. Clinical decision support systems in orthodontics: A narrative review of data science approaches. Orthod Craniofac Res. 2021;24 Suppl 2(Suppl 2):26–36. pmid:33973362
- 16. Alam MK, Kanwal B, Shqaidef A, Alswairki HJ, Alfawzan AA, Alabdullatif AI, et al. A Systematic review and network meta-analysis on the impact of various aligner materials and attachments on orthodontic tooth movement. J Funct Biomater. 2023;14(4):209. pmid:37103299
- 17. Graf S, Tarraf NE, Kravitz ND. Three-dimensional metal printed orthodontic laboratory appliances. Semin Orthod. 2021;27.
- 18. Caruso S, Caruso S, Pellegrino M, Skafi R, Nota A, Tecco S. A knowledge-based algorithm for automatic monitoring of orthodontic treatment: The dental monitoring system. Two cases. Sensors (Basel). 2021;21(5):1856. pmid:33799930
- 19. Kouhi M, de Souza Araújo IJ, Asa’ad F, Zeenat L, Bojedla SSR, Pati F, et al. Recent advances in additive manufacturing of patient-specific devices for dental and maxillofacial rehabilitation. Dent Mater. 2024;40(4):700–15. pmid:38401992
- 20. Richmond S, Shaw WC, Roberts CT, Andrews M. The PAR Index (Peer Assessment Rating): Methods to determine outcome of orthodontic treatment in terms of improvement and standards. Eur J Orthod. 1992;14(3):180–7. pmid:1628684
- 21. Faul F, Erdfelder E, Buchner A, Lang A-G. Statistical power analyses using G*Power 3.1: Tests for correlation and regression analyses. Behav Res Methods. 2009;41(4):1149–60. pmid:19897823
- 22.
Hanif M, Shahbaz MQ, Ahmad M. Sampling Techniques: Methods and Applications. 2017.
- 23. Choi Y, Moon W, Manso AP, Park Y-S, Lim B-S, Chung SH. Shear bond strength of orthodontic brackets bonded with primer-incorporated orthodontic adhesives and unpolymerized 3-dimensional printing materials on 3-dimensional-printed crowns. Am J Orthod Dentofacial Orthop. 2024;165(6):663–70. pmid:38520414
- 24. Coenen FA, Bartz JR, Niederau C, Craveiro RB, Knaup I, Wolf M. Orthodontic treatment quality evaluated by partially automated digital IOTN and PAR index determination: A retrospective multicentre study. Eur J Orthod. 2023;45(3):308–16. pmid:36308300
- 25. Zanjani FG, Pourtaherian A, Zinger S, Moin DA, Claessen F, Cherici T, et al. Mask-MCNet: Tooth instance segmentation in 3D point clouds of intra-oral scans. Neurocomputing. 2021;453:286–98.
- 26. Marya A, Venugopal A, Karobari MI, Chaudhari PK, Heboyan A, Rokaya D. The contemporary management of cleft lip and palate and the role of artificial intelligence: A review. TODENTJ. 2022;16(1).
- 27. Zheng J, Wang X, Zhang T, Jiang J, Wu J. Comparative characterization of supragingival plaque microbiomes in malocclusion adult female patients undergoing orthodontic treatment with removable aligners or fixed appliances: A descriptive cross-sectional study. Front Cell Infect Microbiol. 2024;14:1350181. pmid:38803569
- 28. Zhang X, Gao J, Sun W, Zhang H, Qin W, Jin Z. Evaluation of alveolar bone morphology of incisors with different sagittal skeletal facial types by cone beam computed tomography: A retrospective study. Heliyon. 2023;9(4):e15369. pmid:37113777
- 29. Tüfekçi E, Carrico CK, Gordon CB, Lindauer SJ. Does crown, root, and bone visualization in a clear aligner virtual setup impact treatment decisions?. Am J Orthod Dentofacial Orthop. 2024;165(6):671–9. pmid:38506775
- 30. Park YS, Choi JH, Kim Y, Choi SH, Lee JH, Kim KH, et al. Deep learning-based prediction of the 3D postorthodontic facial changes. J Dent Res. 2022;101(11):1372–9. pmid:35774018
- 31. Surendran A, Daigavane P, Shrivastav S, Kamble R, Sanchla AD, Bharti L, et al. The Future of Orthodontics: Deep Learning Technologies. Cureus. 2024.
- 32. Sangalli L, Alessandri-Bonetti A, Dalessandri D. Effectiveness of dental monitoring system in orthodontics: A systematic review. J Orthod. 2024;51(1):28–40. pmid:37278017
- 33. Yim S, Kim S, Kim I, Park J-W, Cho J-H, Hong M, et al. Accuracy of one-step automated orthodontic diagnosis model using a convolutional neural network and lateral cephalogram images with different qualities obtained from nationwide multi-hospitals. Korean J Orthod. 2022;52(1):3–19. pmid:35046138
- 34. He G, Ricca JM, Dai AZ, Mustahsan VM, Cai Y, Bielski MR, et al. A novel bone registration method using impression molding and structured-light 3D scanning technology. J Orthop Res. 2022;40(10):2340–9. pmid:35119122
- 35. Pellitteri F, Scisciola F, Cremonini F, Baciliero M, Lombardo L. Accuracy of 3D facial scans: A comparison of three different scanning system in an in vivo study. Prog Orthod. 2023;24(1):44. pmid:38143253
- 36. Dey D, Haque MS, Islam MM, Aishi UI, Shammy SS, Mayen MSA, et al. The proper application of logistic regression model in complex survey data: A systematic review. BMC Med Res Methodol. 2025;25(1):15. pmid:39844030
- 37. Wang J, Huang Y, Chen F, Li W. The age-related effects on orthodontic tooth movement and the surrounding periodontal environment. Frontiers in Physiology. 2024.
- 38. Inchingolo F, Inchingolo AM, Malcangi G, Ferrante L, Trilli I, Di Noia A, et al. The interaction of cytokines in orthodontics: A systematic review. Applied Sciences. 2024;14(12):5133.
- 39. Lira Dos Santos EJ, Mohamed FF, Kramer K, Foster BL. Dental manifestations of hypophosphatasia: Translational and clinical advances. JBMR Plus. 2025;9(2):ziae180. pmid:39872235
- 40. Alhoot MA, Alghadban ABS. Knowledge, perception, and attitude of parents toward dental treatment given to their primary school students in tripoli, libya. MJPHM. 2023;23(2):369–79.
- 41. Peikar M. AI driven orthodontic devices: Independent tooth movers (ITM). Semin Orthod. 2023;29.
- 42. Subramanian A, Harikrishnan S. 3D printing in orthodontics: A narrative review. J Int Oral Health. 2023;15(1):15.
- 43. Möhlhenrich SC, Schmidt P, Chhatwani S, Kniha K, Tsipkis A, Jackowski J, et al. Orofacial findings and orthodontic treatment conditions in patients with down syndrome - A retrospective investigation. Head Face Med. 2023;19(1):15. pmid:37149612
- 44. Jiang X, Xu B, Wei M, Wu K, Yang S, Qian L, et al. C2F-3DToothSeg: Coarse-to-fine 3D tooth segmentation via intuitive single clicks. Computers & Graphics. 2022;102:601–9.
- 45. Tarba CI, Cristache MA, Baciu IM, Cristache CM, Burlacu Vatamanu OE, Oancea L. Advancements in digital workflows for 3d-printed maxillofacial soft prostheses: Exploring design and materials in direct additive manufacturing: A scoping review. Applied Sciences. 2025;15(4):1701.