This study is a post hoc analysis of a previously published multicenter, double-blind, randomized, sham-controlled superiority trial (KCT0009493), in which an active TMD DTx app was compared with a visually identical sham app over 6 weeks in a 1:1 allocation at 2 tertiary care oral and maxillofacial surgery clinics in South Korea (Hallym University Sacred Heart Hospital and Hallym University Dongtan Sacred Heart Hospital), with assessments at baseline and weeks 2, 4, and 6 [14]. Detailed eligibility criteria, randomization procedures, intervention content, blinding, ethical approvals, and primary efficacy analyses are reported in the original publication [14]. The present analysis focused on response heterogeneity within the trial by examining predictors of clinically meaningful response, behavioral mediators, and moderators of treatment effect. Patients and members of the public were not involved in the design, conduct, or reporting of the parent trial or the present secondary analysis.

No separate sample size calculation was performed for this post hoc analysis; the analytic sample was determined by the parent trial, which enrolled 102 participants based on a target of 38 participants per group (2-tailed α=.05, 90% power) with an allowance for 25% dropout. In the parent trial, participants were randomly assigned (1:1) by site using stratified block randomization with an independently generated allocation list, and both participants and investigators were blinded to group assignment. The sham application was visually and functionally identical to the active DTx but lacking therapeutic content. Full procedures are reported in the parent trial [14].

Both groups received treatment as usual and were provided with a smartphone app for 6 weeks. The active DTx (Clickless DTx TMD-01) delivered TMD-related education, guided jaw exercises, behavioral habit tracking, guided meditation, and medication guidance with reminder alarms. The sham application was visually and functionally identical in interface and navigation but lacked all therapeutic content, permitting only weekly input of passive user data. Components common to both groups included a jaw health diary for symptom self-monitoring, visual feedback via weekly reports, and supportive check-in calls during weeks 1, 3, and 5, so that the 2 applications differed only in therapeutic content while self-monitoring, feedback, and human contact were held constant. Both applications were locked before trial initiation, and participants accessed them on personal smartphones, with daily use recommended throughout the 6-week period. Full details of intervention development and content are reported in the parent trial and Multimedia Appendix 1 [14].

This manuscript reports a post hoc secondary analysis of a multicenter, sham-controlled randomized clinical trial. The present secondary analysis was reported in accordance with the CONSORT (Consolidated Standards of Reporting Trials) 2025 guidelines and the CONSORT-EHEALTH (V1.6.1; CONSORT of Electronic and Mobile Health Applications and Online Telehealth, Checklist 1) extension for web-based and mobile health interventions [21,22]. The parent trial was approved by the Institutional Review Boards of Hallym University Sacred Heart Hospital (HALLYM IRB 2024-04-001-006) and Hallym University Dongtan Sacred Heart Hospital (HDT IRB 2024-04-001-004) and was conducted under the supervision of the Korean Ministry of Food and Drug Safety (MFDS) in accordance with the Declaration of Helsinki and the International Council for Harmonization Good Clinical Practice (ICH-GCP) guidelines. The parent trial was prospectively registered on the Clinical Research Information Service (KCT0009493; registration date, May 31, 2024), before enrollment of the first participant on June 20, 2024 [14]. All participants provided written informed consent face-to-face prior to enrollment in the parent trial. The institutional review boards granted an exemption from review for the present secondary analysis of deidentified data, which did not require additional informed consent. Participant data were pseudonymized and stored in a secure, password-protected system accessible only to authorized investigators, with no personally identifiable information included in the analysis. Participants received compensation of approximately KRW 500,000 (Korean won; approximately US $350; KRW 1=US $0.00072 as of June 20, 2024) upon trial completion. All data were managed and locked by an independent contract research organization before the analysis to ensure data integrity and analytical independence. No images in this manuscript or its supplementary materials allow for the identification of individual participants.

This analysis integrated trial-derived datasets comprising DTx usage logs, module-completion records, and clinical outcome files. Data were cleaned, standardized by participant identifiers and visit labels, and merged using pseudonymized patient IDs. Engagement variables were derived from server logs over the 6-week intervention period and linked to repeated clinical measurements at each scheduled visit.

App engagement was quantified using objective usage and compliance logs. The variables included total app logins, exercise completion counts, meditation module use, jaw health diary entries, behavioral therapy module interactions, educational video views, article reads, notification checks, and pain record entries, all aggregated over the 42-day intervention period.

Clinical outcomes were assessed at 4 scheduled time points: baseline (V1), week 2 (V2), week 4 (V3), and week 6 (V4; end point). Pain intensity was measured using a Visual Analog Scale (VAS; 0‐100 mm). The jaw function was assessed by a blinded examiner using the maximum mouth opening (MMO; mm) and the Jaw Functional Limitation Scale-20 (JFLS-20). Behavioral and psychological domains were assessed using the OBC and Patient Health Questionnaire-4 (PHQ-4) and scored according to standard procedures for the DC/TMD [23,24].

Statistical analyses were performed on the per-protocol population (N=93; DTx: n=44; sham: n=49). Because the per-protocol population was defined a priori to exclude participants with missing primary outcome data and no item-level missing data were observed among completers, all analyses were conducted on complete cases without imputation, consistent with the parent trial protocol [14]. Normality was evaluated using the Shapiro-Wilk test. Between-group comparisons were performed using independent-samples t tests or Wilcoxon rank-sum tests as appropriate. A 2-sided significance threshold of α=.05 was used. No adjustments for multiple comparisons were made, which was consistent with the exploratory framework, and effect sizes with 95% CIs were reported. Analyses were conducted using Stata (version 17; StataCorp LLC) and Python (version 3.14; Python Software Foundation) with the pandas, SciPy, and Statsmodels packages.

Responder status was defined by ≥30%, ≥50%, and ≥70% reduction in pain VAS score from baseline to week 6. Logistic regression models were used to estimate the odds of response for each threshold, with treatment assignment (DTx=1 and sham=0), age, and change in MMO, OBC, and PHQ-4 scores as independent variables. The results were reported as odds ratios (ORs) with 95% CIs.

Within the DTx arm (n=44), participants were classified as OBC modifiers (n=21) or nonmodifiers (n=23) based on the median OBC change score from baseline to week 6 (median Δ=−8). Participants with OBC changes below the median (ie, greater behavioral improvement) were designated as modifiers. Independent-samples t tests were used to compare clinical outcomes and DTx usage metrics between the subgroups.

Causal mediation analysis used the potential outcomes framework as implemented in the Mediate command in Stata (StataCorp LLC) [25,26]. The treatment variable was DTx assignment (vs sham), the mediator was OBC modifier status (binary), and the outcomes were (1) continuous VAS change from baseline to week 6 and (2) binary responder status at each threshold. The natural indirect effect (NIE), natural direct effect (NDE), and total effect were estimated using bootstrap-based 95% CIs (1000 iterations). This framework relies on several identifying assumptions, including no unmeasured treatment-mediator, mediator-outcome, or treatment-outcome confounding, no mediator-outcome confounder affected by treatment, and correct model specification. Because treatment was randomized, confounding of the treatment-mediator and treatment-outcome paths was minimized by design, leaving mediator-outcome confounding as the principal residual concern. Given the nonrandomized nature of the mediator, results were interpreted as exploratory and associative rather than confirmatory causal inference.

Logistic regression models were reestimated using week 4 change scores to characterize early treatment effects and their temporal dynamics. Probit regression models were used for robustness checks. A linear regression moderator analysis examined the treatment×PHQ-4 High (≥3) interaction on VAS change. The multidomain responder analysis (triple responder: VAS +OBC improvement+MMO improvement) is shown in Multimedia Appendix 2. Figure 1 summarizes the analytical workflow, linking trial-derived data sources to the per-protocol cohort (N=93) and the 5 complementary analyses described below.

Of the 102 randomized participants, 93 met the per-protocol criteria (DTx: n=44; sham: n=49) and were included in the analysis. As previously reported, the DTx group demonstrated significantly greater improvements at week 6 in VAS pain (between-group difference, −23.78 mm; P<.001), MMO (+4.93 mm; P<.001), JFLS-20 (−19.68; P=.001), and OBC (−5.84; P=.008), with no significant difference in PHQ-4 (−0.36; P=.40). Baseline characteristics and week-6 outcomes are provided in Multimedia Appendices 3 and 4. Harms were systematically monitored in the parent trial, in which no serious harms or adverse events were reported in either group. The present secondary analysis did not include an additional assessment of harms.

Across all 3 response thresholds, DTx assignment was a strong predictor of responder status (Table 1). The adjusted OR for DTx was 5.39 (95% CI 1.72‐16.94; P<.01) at the ≥30% threshold, 3.21 (95% CI 1.14‐8.99; P=.03) at ≥50%, and 3.45 (95% CI 1.12‐10.63; P=.03) at ≥70%. MMO improvement was independently associated with higher odds of response at all thresholds (ORs 1.13‐1.15; all P<.05). Changes in PHQ-4 scores and age were not significant predictors at any threshold.

OBC change exhibited a threshold-dependent gradient: each unit reduction in OBC score was associated with increased odds of response (OR 0.94 at ≥30% and ≥50%; P=.03; OR 0.91 at ≥70%; 95% CI 0.85‐0.98; P=.01). The association strengthened as the response criterion became more stringent, with improved model fit (Pseudo R²=0.32 at ≥70% vs 0.25-0.27 at lower thresholds). Estimates were robust to model specification, with probit regression yielding directionally consistent treatment effects across all thresholds (Multimedia Appendix 5).

Under a stricter multidomain responder definition requiring simultaneous VAS, OBC, and MMO improvement, DTx participants showed 8.55-12.32-fold higher odds of triple response than sham across thresholds (all P<.01; Multimedia Appendix 2).

Within the DTx arm, OBC modifiers (n=21) showed greater pain reduction than nonmodifiers (n=23; VAS change: mean −45.71, SD 17.34 vs mean −22.61, SD 26.75 mm; difference: −23.11 mm; 95% CI −36.97 to −9.24; P<0.01) and greater improvement in jaw function (JFLS-20: mean −42.52, SD 26.28 vs mean −22.22, SD 28.72; difference −20.31; 95% CI −37.11 to −3.50; P=.02). PHQ-4 improvement did not differ significantly between the subgroups (difference −1.07, 95% CI −2.16 to 0.01; P=.05). MMO change also showed no significant difference (difference −0.64, 95% CI −3.88 to 2.61; P=.69).

Despite these clinical differences, no significant between-group differences were observed across any objective DTx usage metric, including total logins, exercise completions, meditation sessions, jaw diary entries, educational video views, or behavioral therapy module uses (all P>.05), with a nonsignificant trend toward fewer behavioral therapy entries among the modifiers (P=.10; Table 2). The waterfall plot illustrates that the OBC modifier cases were concentrated among participants with the greatest VAS reduction (Figure 2).

For continuous VAS change, the NIE through OBC modifier status was −6.91 mm (95% CI −13.23 to −0.58; P=.03), indicating that approximately 29.1% of the total treatment effect (−23.78 mm) was mediated through behavioral modification. The NDE remained significant (−16.87 mm; 95% CI −27.30 to −6.45; P<.01).

A threshold-dependent pattern emerged for the binary outcomes. The NIE was 0.09 (95% CI −0.01 to 0.18; P=.07) at ≥30%, 0.11 (95% CI −0.00 to 0.22; P=.05) at ≥50%, and 0.13 (95% CI 0.01‐0.25; P=.03) at ≥70%, with the mediated proportion increasing from 19.6% to 25.6% to 30.2%. The NDE remained significant across all models (all P<.01; Table 3 and Figure 3).

When logistic regression models were reestimated using week 4 change scores, treatment ORs were consistently higher than at the 6-week end point across all thresholds: 7.15 (95% CI 2.39‐21.32; P<.01) vs 5.39 at ≥30%, 4.48 (95% CI 1.68‐11.94; P<.01) vs 3.21 at ≥50%, and 4.91 (95% CI 1.74‐13.87; P<.01) vs 3.45 at ≥70%, representing increases of 33%, 40%, and 42%, respectively (Table 4, Figures 4A and 4B).

OBC change at Week 4 was significant for the ≥70% threshold (OR 0.92, 95% CI 0.87‐0.99; P=.02) but not at lower thresholds, indicating the behavioral effect matures between weeks 2 and 4. MMO change remained significant across all thresholds at week 4 (ORs 1.10‐1.11; all P<.05). The change in PHQ-4 score was not significant in any model. Week-4 estimates were similarly robust under probit regression, with significant treatment effects retained across all thresholds (Multimedia Appendix 6).

To explore whether baseline psychological distress moderated the treatment effect, a linear regression model with an interaction term (Treatment×PHQ-4 High [≥3]) was used with VAS change as the dependent variable (Table 5). A PHQ-4 score of ≥3 was applied as the binary threshold, consistent with established cutoff criteria indicating clinically meaningful anxiety and/or depressive symptoms [27,28]. Of the 93 participants, 65 had PHQ-4 ≥3 and 28 had PHQ-4 <3 at baseline.

In the adjusted model controlling for baseline VAS, age, and sex, the interaction term was significant (β=−19.63; P=.04), indicating the DTx effect on VAS change differed by baseline distress level. Among participants with PHQ-4 ≥3, the estimated DTx effect on VAS change was −28.09 mm (sum of treatment coefficient and interaction term: −8.46 + [−19.63]), representing greater pain reduction compared with those with PHQ-4 <3, in whom no significant treatment effect was observed (β=−8.46 mm; P=.25). The main effect of PHQ-4 High was significant (β=14.91; P=.01), indicating that higher baseline psychological distress was independently associated with greater VAS reduction regardless of treatment assignment. The adjusted model explained 49% of the variance in VAS score change (R²=0.49), compared with 24% in the unadjusted model (R²=0.24). As a sensitivity analysis, the interaction was not significant when PHQ-4 was modeled as a continuous variable (β=−4.02 per point; P=.06). The CONSORT flow diagram for the overall clinical trial is presented in Figure 5.

This post hoc analysis of a multicenter, sham-controlled randomized trial of a TMD DTx extends the parent trial’s demonstration of aggregate efficacy to the mechanisms and heterogeneity of treatment response. Building on the confirmed clinical benefit of the DTx, we examined which clinical and behavioral factors were associated with pain response, whether oral behavior change might mediate the treatment effect, and how the treatment effect evolved over time and varied with baseline psychological distress. Across all response thresholds, DTx assignment remained a strong and consistent predictor of clinically meaningful pain relief relative to sham. Three exploratory patterns emerged, including oral behavior modification appeared as a possible mediating pathway, the treatment effect appeared to be accelerated and amplified, with responders reaching clinically meaningful pain relief earlier and to a greater degree, and baseline psychological distress may play a modest moderating role. As all analyses were performed post hoc, these patterns were interpreted as associative and hypothesis-generating and are elaborated below.

In the present analysis, OBC change was the most consistent independent factor associated with response, with the association strengthening at progressively more stringent response thresholds. This gradient may indicate that behavioral modification becomes increasingly important as pain relief increases. Mediation analysis was consistent with this pattern. The proportion of the VAS treatment effect potentially transmitted via behavioral modification rose monotonically across increasingly stringent response thresholds, reaching approximately 29% on the continuous VAS scale. This positive dose-response relationship, in which greater behavioral change accompanied greater pain reduction, is consistent with OBC behavioral modification potentially acting as a partial mediating pathway rather than solely as a correlate.

A particularly informative pattern was the dissociation between OBC-defined behavioral change and objective app usage metrics. OBC modifiers achieved significantly greater pain reduction than nonmodifiers, yet no significant differences were observed across any engagement metrics, including logins, exercises, meditation sessions, or diary entries. The nonsignificant but directionally consistent pattern of fewer behavioral therapy violation entries among modifiers provides a triangulation across 3 independent measurement modalities, including self-report OBC, digital behavioral logs, and clinical pain VAS outcomes. Of note, the behavioral therapy module log records patient-reported breaches of behavioral recommendations; therefore, fewer entries indicate better behavioral adherence rather than reduced app use. This directionality aligns the digital log with the lower OBC scores and the greater pain reduction observed in modifiers, with all 3 modalities converging on the same pattern. This suggests that meaningful behavioral change, rather than passive app engagement, may be the more relevant driver of clinical improvement in this cohort [19]. This dissociation should nonetheless be interpreted cautiously, as the present sample may be underpowered to detect modest differences in engagement metrics.

Treatment ORs were consistently higher at week 4 than at the 6-week end point across all responder thresholds, which may suggest that DTx had its strongest discriminatory effect at this intermediate time point. The temporal onset of OBC as an associated marker of response, which was absent at week 2 but emerged for the ≥70% threshold by week 4, may suggest that meaningful behavioral modification requires approximately 4 weeks to consolidate and manifest clinically [29]. Taken together, these temporal patterns may suggest that DTx responders achieve an accelerated and amplified clinical improvement than nonresponders, reaching clinically meaningful pain relief earlier and to a greater degree, with behavioral modification as a plausible temporal correlate of this response.

Baseline psychological distress (PHQ-4 ≥3) appeared to modestly amplify the treatment effect, with a larger estimated DTx effect in the elevated-distress subgroup than in the subgroup without distress. However, the interaction was not significant when PHQ-4 was modeled continuously, and the ≥3 threshold was a predefined clinical cutoff rather than a data-derived optimum; these findings are exploratory and require prospective validation [27].

The central role of OBC behavioral change aligns with and extends the biopsychosocial understanding of TMD pathophysiology [30]. Parafunctional oral behaviors, including clenching, bruxism, and habitual jaw tension, are recognized perpetuating factors that sustain mechanical loading on the temporomandibular joint and maintain pain-guarding cycles [30,31]. Longitudinal evidence from the Orofacial Pain: Prospective Evaluation and Risk Assessment (OPPERA) study associated high awake parafunction with both TMD onset and chronification [24,32]. Building on this prevalence-level evidence, our findings suggest that targeted modification of these parafunctional behaviors may yield measurable clinical improvement, moving the inference beyond a statistical between-group comparison toward a data-based mechanistic account of how parafunction-directed intervention may translate into pain reduction.

This dissociation is conceptually consistent with the framework proposed by Schmidt et al [19], who distinguish between “consumed dose,” the patient’s interaction with a DTx, and “realized dose,” the degree to which cognitive-affective and behavioral change is actually internalized by the patient, and who argue that the therapeutic effect is not a direct function of consumed dose but rather of realized dose. This convergence is consistent with the view that therapeutic benefit may depend more on behavioral adaptation, the realized dose, than on engagement volume [19]. It thus raises questions about the assumption that a greater consumed dose should predict clinical outcomes [18]. Our results provide a data-based, exploratory quantification of this distinction within a sham-controlled randomized clinical trial, suggesting that OBC-defined behavioral modification may account for approximately 29% of the total VAS treatment effect independently of any consumed dose metric. For DTx product development, optimization might prioritize facilitating real-world behavioral internalization through adaptive content delivery and habit-change reinforcement rather than maximizing session frequency or time-on-app [33-35]. These findings further caution against the use of engagement metrics such as login frequency or session duration, as surrogate indicators of adherence or therapeutic dose, given that meaningful behavioral change may occur independently of, or even inversely to, raw usage volume [19].

Because week-4 behavioral change precedes the week-6 end point, it can be interpreted as an early-treatment predictor with temporal precedence, in contrast to the concurrent change variables measured at the end point, which are more appropriately regarded as factors associated with response. The week-4 amplification has clear precedent in the broader habit-formation literature, where behavioral consolidation typically requires approximately 4 weeks to stabilize [29]. These findings may carry 2 practical implications. First, week 4 may represent a promising candidate time point at which patients who have not demonstrated a behavioral or clinical response by this time point can be identified early to receive augmented or alternative strategies before completing the full 6-week course. However, because these observations are post hoc, prospective adaptive-design or decision-analytic studies would be needed before a week-4 rule could inform clinical practice [36]. These adaptive treatment approaches are increasingly advocated within digital health interventions [36]. Second, the amplified ORs raise the prospective hypothesis that a condensed 4-week protocol might achieve equivalent or near-equivalent gains with reduced patient burden, a question that warrants dedicated evaluation in future trials [37].

The PHQ-4 moderator pattern is clinically plausible: patients with psychological comorbidity may benefit more from the self-management, mindfulness, and behavioral awareness components in the DTx intervention [38-40]. Despite evidence that 60%‐77% of patients with TMD experience psychological distress [3], routine screening remains underused in dental practice [41]. Brief instruments such as the PHQ-4 could help identify patients most likely to benefit from behavioral digital interventions [42,43].

This study has several limitations. All findings are from a post hoc secondary analysis and are associative and hypothesis-generating, not causal. Because the proposed mediator, OBC-defined behavioral change, is a postrandomization variable rather than a randomized exposure, causal mediation inference is precluded despite the use of the potential outcomes framework, and residual unmeasured confounding of the mediator-outcome relationship cannot be excluded [44]. In particular, reverse causation cannot be ruled out: patients whose pain improves may subsequently report fewer parafunctional behaviors because they feel better, are less vigilant, or perceive less need to guard the jaw [3]. Under this scenario, OBC improvement could function as both a consequence and a correlate of pain reduction rather than solely as an antecedent mediator, and the assumed direction from behavioral change to pain relief, while biologically plausible, cannot be confirmed in the present design. Baseline OBC and PHQ-4 scores were higher in the DTx group, which may have influenced the mediation and moderator analyses through regression to the mean. The OBC modifier classification relies on a data-driven median split, introducing cut-point sensitivity, and creating subgroups that may not be stable across samples. Furthermore, multiple responder, subgroup, mediation, sensitivity, and moderator analyses were conducted without formal correction for multiple comparisons; borderline findings, particularly those with P values close to .05, should therefore be regarded as exploratory signals rather than confirmatory results. Finally, the mediation estimates were sensitive to distributional assumptions and sample sizes, necessitating replication in larger independent cohorts [44].

The analysis was conducted in the per-protocol population (N=93) rather than the intention-to-treat sample, as the mechanistic question of how DTx exerts its therapeutic effect inherently concerns participants who completed the intervention as intended; intention-to-treat estimates would have diluted the engagement-outcome dissociation with noncompleters and were therefore not appropriate for this specific analytic aim [45]. This principled selection nonetheless restricts inference to completers and should be considered when generalizing the mechanistic findings to broader trial populations. The per-protocol sample size also restricted the statistical power of the subgroup and interaction analyses, particularly for the PHQ-4 moderator, and limited the ability to detect modest between-group differences in engagement metrics. In addition, only a single behavioral mediator (OBC) was examined; other plausible pathways, such as cognitive reappraisal, pain-related self-efficacy, or improved sleep, were not assessed and may also contribute to treatment response [46-48]. The 6-week follow-up precluded the assessment of long-term durability, and a longitudinal follow-up study is currently underway. The trial was conducted at 2 tertiary care centers in South Korea, limiting its generalizability to other health care settings and populations.

These findings may refine the understanding of therapeutic dose in digital interventions. While app engagement may be a necessary vehicle, it may be the patient’s internalization of that engagement into sustained behavioral change that is more closely linked to clinical benefit. In TMD, this realized behavioral adaptation, shaped by underlying psychological context and consolidated within a defined temporal window, may represent a key behavioral pathway underlying treatment response. Practically, these observations suggest that DTx development and evaluation might prioritize measures of behavioral realization over engagement volume and that mechanism-driven, precision-oriented trial designs may help clarify for whom and through which pathways these interventions work. These directions remain hypotheses to be confirmed in future prospective studies.