This study was a single-center, 2-arm, parallel-group RCT reported in accordance with the CONSORT (Consolidated Standards of Reporting Trials) 2010 guidelines. It was conducted at the Chinese National Medical Center of Orthopaedics in Shanghai, China (affiliated with Shanghai Sixth People’s Hospital) from June 2023 to June 2024.

Eligible patients awaiting primary THA at the study center were identified from surgical waiting lists and screened by two physicians according to predefined inclusion criteria. Patients meeting eligibility were enrolled prior to surgery. Baseline (preoperative) evaluations for all participants were performed by a single physiotherapist after consent. On the first postoperative day, participants were randomly allocated in a 1:1 ratio to the study groups by a trial coordinator (allocation details and sequence described later). Two independent physiotherapists (not involved in initial enrollment nor outcome assessment) subsequently contacted each participant to implement the allocated rehabilitation intervention. All participants then underwent a 12-week home-based rehabilitation program according to their group assignment and were followed for a total of 24 weeks postsurgery. The 24-week follow-up duration was selected to capture both the acute rehabilitation phase, during which sensor-mediated benefits were expected to be most pronounced, and the later recovery period, during which convergence of outcomes was anticipated based on prior literature in musculoskeletal rehabilitation. Outcome assessments were conducted at baseline (presurgery) and at 6, 12, and 24 weeks after surgery by two physiotherapists in the outpatient clinic who were blinded to group allocation to minimize observer bias. These outcome evaluators, who were not involved in providing the interventions, collected all functional performance measures and assisted participants with completing self-reported outcome questionnaires using an online system during the clinic visits. Participant adherence to the digital training intervention was monitored through system usage logs; if a physiotherapist noted that a patient was not engaging with the digital rehabilitation platform as prescribed, they promptly inquired about the reasons for nonadherence to address any barriers to participation.

Ethics approval for this study was obtained from the Ethics Committee of Shanghai Sixth People’s Hospital (approval number: 2016KY-109(K)), and the study was conducted in accordance with institutional guidelines and international ethical standards, including the principles of the Declaration of Helsinki. Written informed consent was obtained from all participants prior to enrollment, and participation was voluntary. All data were anonymized to protect participants’ privacy and maintain confidentiality, ensuring that no personal identifiers were included in the analytical dataset. No compensation was provided to participants.

This single-center trial enrolled patients scheduled for primary THA from the waiting list of the orthopedic surgery outpatient clinic of a tertiary hospital. Eligible patients were identified and consented preoperatively, and baseline assessments were completed before surgery. Potential participants were informed about the study 2 weeks prior to their admission and formally screened for eligibility on the same day of their admission. Those meeting all inclusion criteria were invited to participate, and written informed consent was obtained before enrollment.

The inclusion and exclusion criteria are shown in Textbox 1. Patients meeting the criteria were enrolled by the research coordinator and randomized to begin the assigned home-based rehabilitation program. All participants were recruited from a single high-volume arthroplasty center, ensuring a consistent surgical protocol and postoperative care across the study cohort. Each participant’s eligibility was verified against the criteria prior to randomization, and a screening log was maintained to record any exclusions or refusals.

All participants in both the intervention and control groups engaged in a 12-week home-based rehabilitation exercise program following THA. On the second postoperative day, each patient received an in-person orientation: One of the technicians assisted them with installing a dedicated mobile app (Joymotion software, Shanghai Medmotion Medical Management Co, Ltd) on their smartphone and provided instructions on its use. This app-based platform delivered a “Digital Physiotherapeutic THA-Specific Training System,” following the standardized post-THA training protocol (see Multimedia Appendix 1), and it recorded each patient’s exercise performance and completion rates. The supervising physiotherapist prescribed individualized daily exercise tasks through the app and could adjust the program remotely as needed. Both groups were scheduled for weekly teleconsultation sessions; at an appointed time each week, a physiotherapist from the rehabilitation center initiated a 2-way video call via the app to monitor progress, provide coaching, and address any issues. Notably, the frequency and duration of physiotherapist contact (including teleconsultations) were identical for the two groups to ensure parity in clinical guidance and support.

Participants allocated to the inertial measurement unit (IMU) intervention group received the home-based digital exercise program via the smartphone app with full telerehabilitation features. The app provided daily interactive exercise instructions, feedback on training performance, and real-time video or audio communication with the physiotherapist. The rehabilitation regimen was automatically tailored to each patient’s baseline characteristics and functional status, and the physiotherapist could further fine-tune the difficulty or exercises through a clinician interface. In addition to the app-based program, this group also used a wearable inertial sensor system to enhance exercise monitoring and feedback. Five small wireless IMU sensors were affixed to the patient’s body (one at the waist and one on the front of each thigh and each lower leg) and were registered and calibrated with the app during the initial setup. During each training session, these sensors captured the patient’s movements and projected them onto a virtual avatar visible in the app. The system’s software continuously compared the patient’s avatar movements with the pre-set correct posture for each exercise. If a movement did not reach the target position or was performed with improper form, the patient received immediate feedback from the app, prompting them to repeat the motion correctly. This real-time feedback mechanism ensured that patients in the IMU group could correct their technique on the spot, helping them perform the exercises as intended. Throughout the 12-week period, the assigned physiotherapists remotely monitored the IMU group’s exercise sessions via the app’s data logs. The sensor system transmitted detailed performance data to the therapist interface, allowing the physiotherapist to review each patient’s adherence, exercise form, and progress. Any necessary modifications to the exercise program or additional guidance were provided based on these observations.

Participants in the control group followed the same 12-week home exercise program using the digital platform but without the aid of wearable sensors. Like the IMU group, control group patients accessed daily THA-specific exercises and instructional videos through the mobile app, which tracked their exercise completion and provided basic performance feedback such as summaries of repetitions completed or time spent exercising. They received the same schedule of weekly physiotherapist-led teleconference sessions via the app’s two-way video interface, ensuring equal contact time and guidance as the intervention group. The exercise content and progression for control participants were also individualized to their baseline function and could be adjusted remotely by the physiotherapist in the same manner as for the IMU group. However, unlike the IMU group, the control group did not receive real-time motion feedback on their exercise technique. They performed the prescribed exercises with the app’s guidance alone (see Multimedia Appendix 1), without the virtual avatar or sensor-based corrections. All other aspects of the rehabilitation program were kept consistent between the two groups.

The primary outcome was the Hip Disability and Osteoarthritis Outcome Score (HOOS), measured at 24 weeks postsurgery as a comprehensive patient-reported index of hip-related health status [7]. Secondary outcomes were assessed at follow-up visits and included the 5 HOOS subscale scores (pain, symptoms, activities of daily living [ADL], sport/recreation, and hip-related quality of life), objective functional performance tests (the timed up and go [TUG] test and the Berg Balance Scale), generic health-related quality of life measures (36-item Short Form Health Survey [SF-36] physical and mental component summary scores), and psychological status measures (Hospital Anxiety and Depression Scale [HADS] anxiety subscale [HADS-A] and depression subscale [HADS-D]) [8-11]. Patient satisfaction with the rehabilitation program was evaluated at 12 weeks and 24 weeks using a 4-item Likert-scale questionnaire addressing convenience, ease of use and access, perceived helpfulness, and communication with the care provider (each item scored from 1 to 5, with higher scores indicating greater satisfaction).

Cost outcomes were analyzed from a societal perspective capturing direct medical costs (including the cost of the wearable sensor hardware and setup for the IMU group), direct nonmedical costs, and opportunity costs (lost wages and productivity) obtained from hospital billing records and participant self-reports. All costs were converted to 2024 Chinese Yuan (¥), with costs incurred before 2024 adjusted to 2024 price levels using the national consumer price index and any costs occurring beyond 1 year postsurgery discounted at an annual rate of 3%. A currency exchange rate of ¥1=US $0.15 is applicable.

Adherence to the rehabilitation program was measured using multiple indicators collected throughout the 24-week study period. These included the number of prescribed exercise sessions completed, the number of weekly app-based assessments attended, and the number of scheduled weekly follow-up phone calls successfully completed in each group.

Adverse events (AEs) and serious AEs (SAEs) were closely tracked as part of safety monitoring from enrollment through the full 24-week trial. Participants were instructed to report any AEs or SAEs either through the study’s mobile app (which included a “Report a Problem” feature) or via a phone call to the research team. For this study, an AE was defined as any untoward or unintended medical occurrence during the trial, regardless of whether it was considered related to the intervention. An SAE was defined using standard criteria as any adverse event that resulted in death, was life-threatening, required or prolonged hospitalization, caused significant disability or incapacity, or captured any other important medical event that necessitated intervention to prevent such outcomes. Follow-up assessments were conducted at 6 weeks, 12 weeks, and 24 weeks postsurgery in the outpatient clinic. Two trained physiotherapists conducted all outcome measurements at each visit. Participants completed patient-reported outcome measures, including HOOS, HADS, and SF-36, using paper questionnaires. All performance-based outcome tests were administered in person by the physiotherapists during each follow-up visit.

The sample size was determined to provide sufficient power to detect a clinically meaningful difference in the primary outcome (HOOS) between groups. Based on published data in THA populations, an improvement of roughly 8 to 10 points on the HOOS represents the minimum clinically important difference in patient-reported hip function after surgery. We therefore set a 10-point difference in HOOS as the target effect size for this trial. To inform the variability estimate, we reviewed prior studies reporting HOOS outcomes and found that the SD of change scores is approximately 15 to 20 points in this context. Using these assumptions and a 2-sided significance level of .05 with 80% statistical power, we calculated that 86 patients per group would be required to detect a 10-point difference in HOOS between the intervention and control arms. This calculation was based on a 2-sample comparison of means for the primary outcome. To account for an anticipated attrition rate of about 25% (loss to follow-up and dropouts), the sample size was inflated accordingly. We aimed to recruit approximately 115 participants per arm to ensure that the study would remain adequately powered for the primary endpoint after accounting for potential dropouts. This sample size was also expected to provide sufficient power for key secondary outcomes while maintaining a feasible recruitment target.

An independent statistician generated the allocation sequence using block randomization with randomly varying block sizes (between 4 and 6), achieving a 1:1 allocation ratio to the 2 study groups. Allocation concealment was maintained by placing assignments in sequentially numbered, opaque, sealed envelopes. Randomization was implemented after participants completed baseline assessments. At that time, the enrolling investigator opened the next envelope in sequence to reveal the participant’s group assignment. The allocation list was kept secure by the statistician and was not accessible to investigators or clinicians responsible for recruitment and treatment, ensuring that allocations remained concealed until the point of assignment. Due to the nature of the intervention, participants and treating physiotherapists were not blinded to group allocation; however, all outcome assessors and data analysts remained blinded to group assignments throughout the study to minimize the risk of bias.

Baseline characteristics, outcome measures, and treatment adherence data were summarized using appropriate descriptive statistics (mean and SD for continuous variables; frequencies and percentages for categorical variables). We applied the Shapiro-Wilk test to formally evaluate normality [12]. A nonsignificant Shapiro-Wilk test (P>.05) indicated no evidence against normality. Any major deviations would have prompted use of transformations or nonparametric methods, but no such issues arose in this study. All hypothesis tests were 2-sided with a significance level of 5%. Intention-to-treat (ITT) analyses were primary, including all participants as randomized.

For the primary outcome (HOOS at 24 weeks postrandomization), an adjusted 2-sided analysis was conducted in the ITT population. The primary between-group comparison of HOOS at 24 weeks was based on a linear mixed effects model treating the repeated HOOS measurements as the dependent outcome. This model included fixed effects for treatment group, time, and the treatment-by-time interaction, with baseline HOOS score, age, and sex entered as covariates to improve precision. A random intercept for each participant was incorporated to account for within-subject correlation over time. As this was a single-center trial, no random center effect was needed. The adjusted mean difference in HOOS at 24 weeks between the 2 groups was estimated from this model with a 2-sided 95% CI. The primary hypothesis was evaluated using the P value associated with the treatment group effect (difference in outcomes between groups) at the 24-week time point. Model assumptions (such as normality of residuals and homogeneity of variance) were checked, and analyses were robust to these assumptions or used appropriate transformations if necessary.

Continuous secondary outcomes were analyzed with similar linear mixed effects models appropriate for their distribution. Each continuous outcome measured over time was examined for overall group differences and trajectory over the follow-up period. These models included the same fixed effects (group, time, and their interaction) and participant-level random intercepts, adjusting for baseline values when available. From each model, we obtained estimates of the mean outcome in each group at each follow-up time and the mean differences between groups with 95% CIs. Group differences at each time point were derived from interaction terms or post hoc contrasts. For binary categorical outcomes, comparisons between groups were made using χ² tests. These dichotomous results were reported as risk differences or relative risks with corresponding 95% CIs. All significance tests were 2-sided, and a P value <.05 was considered statistically significant for each individual outcome analysis. To control for inflated Type I error across secondary endpoints, a Bonferroni correction was applied to all secondary outcome comparisons. The primary outcome (HOOS overall at 24 weeks) was tested at a 2-sided significance level of .05 as the single confirmatory analysis. For secondary outcomes, the family-wise alpha was divided by the total number of secondary comparisons (47 tests across 15 secondary outcome measures at up to 3 time points, plus 2 nonprimary time points for HOOS overall), yielding a Bonferroni-adjusted significance threshold of P<.001. Both unadjusted and Bonferroni-adjusted significance levels are reported for transparency.

A comprehensive cost-effectiveness analysis was also performed alongside the clinical outcomes. Total health care costs for each participant over the 24-week follow-up were calculated, incorporating intervention costs and any other relevant medical expenditures. The difference in mean total cost between the 2 groups was then evaluated using appropriate between-group comparisons (with nonparametric bootstrap methods applied to account for skewed cost data). We plotted the joint uncertainty in costs and effects on a cost-effectiveness plane to visualize the difference in mean cost against the difference in effectiveness (primary outcome improvement) for the digital rehabilitation versus control. If the digital intervention demonstrated lower average cost while yielding no worse outcome, it was considered a dominant strategy (more cost-effective). In cases where the digital program had higher cost or marginally different outcomes, an incremental cost-effectiveness ratio (ICER) was computed to quantify the additional cost per unit of benefit gained. Uncertainty in the cost-effectiveness results was explored by nonparametric bootstrapping (10,000 resamples), which produced a cloud of incremental cost-effect pairs on the cost-effectiveness plane and allowed construction of cost-effectiveness acceptability curves. These acceptability curves illustrate the probability that the digital program is cost-effective compared with the control across a range of willingness-to-pay thresholds for improvement in outcome.

Missing data were handled using both ITT and per-protocol (PP) analysis strategies. In the ITT analysis, all participants were retained in their originally assigned groups for analysis, regardless of protocol adherence or missing follow-up data. The linear mixed effects model used all available observations from each participant under a missing-at-random assumption, thereby inherently accommodating incomplete follow-up without requiring formal imputation. In contrast, the PP analysis included only those participants who completed all scheduled follow-up assessments, thereby using a complete-case approach. It is important to note that missingness in our dataset was holistic at the follow-up level: If a patient was lost to follow-up at a given time point, all variables for that visit were missing for that patient. There was no item-level missingness within an attended visit. Consequently, missing data occurred only at the participant-visit level, and all available data from each participant’s other time points were still used in the mixed model under the ITT approach. Because entire follow-up visits were missing for dropouts and not selective outcome omissions, we did not perform any statistical imputation for missing values. In addition, given this pattern of whole-visit missingness, a formal test for data missing completely at random (such as the Little MCAR test) was not applicable, as that test is designed for detecting patterns of sporadic item-level missingness across variables rather than the uniform visit-level dropout observed in this trial.

We evaluated 279 patients scheduled for primary THA, and 240 were randomized. The flowchart of our study is shown in Figure 1. The baseline characteristics were comparable between the IMU and control groups. Men accounted for 62 (62/120, 51.7%) participants in the IMU group and 56 (56/120, 46.7%) participants in the control group (P=.44); the mean age was 70 years versus 69 years (P=.46); mean BMI was 22 kg/m² versus 22 kg/m² (P=.31; Table 1). Paracetamol or nonsteroidal anti-inflammatory drugs were taken by 17 (17/120, 14.2%) in the intervention group versus 18 (18/120, 15%) in the control group (P=.86). In addition, 62% (74/120, 61.7% vs 72/120, 60%) had the left hip as the surgical side (P=.79). There were no significant differences in functional tests nor patient-reported outcome measures.

Sensitivity analysis using the PP population showed similar results as those of the ITT population (Table S1 in Multimedia Appendix 2).

The primary outcome (HOOS) improved significantly more in the IMU group than in the control group at all follow-ups. The estimated between-group difference in the overall HOOS (IMU minus control) was 2.021 (95% CI 1.445‐2.598) points at 6 weeks (P<.001), 1.721 (95% CI 0.896‐2.545) points at 12 weeks (P<.001), and 1.298 (95% CI 0.313‐2.282) points at 24 weeks (P=.01). All 5 HOOS subscales showed a pattern of greater early improvement with the IMU intervention. At 6 weeks, 4 of the 5 subscales (pain, symptoms, sport/recreation, and hip-related quality of life) had significant between-group differences after Bonferroni correction (all P<.001), with coefficients ranging from 2.334 to 3.327 points. The HOOS ADL subscale difference at 6 weeks (1.208 points, 95% CI 0.283 to 2.134; P=.01) did not survive Bonferroni correction. At 12 weeks, none of the individual HOOS subscale differences remained significant after correction for multiple comparisons (unadjusted P values ranged from .002 to .047). By 24 weeks, all subscale differences had further narrowed and were not statistically significant at either the unadjusted or adjusted level (all P>.13), except for HOOS pain, which had a nominal difference of 1.501 (95% CI 0.084 to 2.918) points (unadjusted P=.04) that did not survive Bonferroni correction. In terms of objective function, the IMU group completed the TUG faster than controls by 1.198 (95% CI 0.789 to 1.607) seconds at 6 weeks (P<.001) and 1.131 (95% CI 0.557 to 1.705) seconds at 12 weeks (P<.001), both of which remained significant after Bonferroni correction. The 0.901-second (95% CI 0.237 to 1.565) difference at 24 weeks was nominally significant (unadjusted P=.008) but did not survive Bonferroni correction. Berg Balance Scale scores were numerically higher in the IMU group at all time points (0.300 to 0.431 points), but none of these differences reached the Bonferroni-adjusted threshold (unadjusted P values of .004, .02, and .16 at 6 weeks, 12 weeks, and 24 weeks, respectively). For generic patient-reported outcomes, the SF-36 physical component summary was higher in the IMU group by 3.821 (95% CI 3.315 to 4.326) points at 6 weeks, 2.626 (95% CI 1.913 to 3.338) points at 12 weeks, and 1.949 (95% CI 1.144 to 2.753) points at 24 weeks (all P<.001; all surviving Bonferroni correction). The SF-36 mental component summary showed a similar pattern (3.239, 95% CI 2.881 to 3.597; 2.272, 95% CI 1.771 to 2.773; and 1.726, 95% CI 1.117 to 2.334 points at 6 weeks, 12 weeks, and 24 weeks, respectively; all P<.001; all surviving Bonferroni correction). Psychological outcomes also consistently favored the IMU arm across all time points and remained significant after Bonferroni correction: The IMU group had lower HADS-A scores by 0.537 (95% CI 0.470 to 0.604), 0.418 (95% CI 0.310 to 0.527), and 0.363 (95% CI 0.220 to 0.506) points at 6 weeks, 12 weeks, and 24 weeks, respectively (all P<.001), and lower HADS-D scores by 0.501 (95% CI 0.435 to 0.566), 0.370 (95% CI 0.260 to 0.481), and 0.307 (95% CI 0.163 to 0.450) points at 6 weeks, 12 weeks, and 24 weeks, respectively (all P<.001). Patient satisfaction with the rehabilitation program was high in both groups, with no significant between-group differences in any domain at 12 weeks or 24 weeks (all P>.23). Detailed results of the primary and secondary outcomes are shown in Tables 2 and 3. Sensitivity analyses in the PP population yielded results consistent with the ITT findings, reinforcing the robustness of the observed treatment effects (Tables S2-S3 in Multimedia Appendix 2).

The mean 24-week total cost per patient was ¥87,967.32 (SD ¥8854.19) in the IMU group versus ¥94,396.30 (SD ¥10,037.74) in the control group (P=.32). The cost composition differed between groups: IMU participants had significantly lower in-hospital costs (¥69,534.95, SD ¥5394.69 vs ¥70,359.45, SD ¥5043.84; P<.001) and primary care costs (¥250.00, SD ¥38.14 vs ¥1503.18, SD ¥186.13; P<.001). Nonmedical expenses were lower with IMU, including paid home care (¥1178.73, SD ¥362.12 vs ¥1451.95, SD ¥357.74; P<.001), transportation (¥800.06, SD ¥138.89 vs ¥1968.67, SD ¥273.82; P<.001), and nutrition (¥4505.32, SD ¥479.00 vs ¥5093.08, SD ¥519.96; P<.001). Patient productivity losses were also lower (¥1599.77, SD ¥6994.72 vs ¥1829.44, SD ¥6618.91; P<.001). The detailed results are shown in Table 4.

The IMU intervention’s incremental cost was –¥6428.98 (IMU less costly). Correspondingly, most ICERs were negative, though the overall cost difference between groups was not statistically significant (¥87,967.32 vs ¥94,396.30, P=.32). For the primary outcome (HOOS overall), the incremental effect was +1.298 (95% CI 0.313 to 2.282) points, with an ICER of –¥4954 per point. TUG time decreased by 0.901 (95% CI 0.237 to 1.565) seconds (ICER ¥7136.79 per second), and the Berg Balance Scale increased by 0.300 (95% CI –0.118 to 0.718) points (ICER –¥21,451.43 per point). SF-36 physical component summary and mental component summary improved by 1.949 (95% CI 1.144 to 2.753) points (ICER –¥3299.31 per point) and 1.726 (95% CI 1.117 to 2.334) points (ICER –¥3725.41 per point), respectively. The detailed results, including for the HOOS subscales, are shown in Table 5. PP sensitivity analyses (Tables S4 and S5 in Multimedia Appendix 2) yielded similar cost-effectiveness results.

Adherence to the intervention was high in both groups, with the IMU group completing more of the 60 prescribed exercise sessions (mean 51.6, SD 9.3 vs mean 48.7, SD 10.6; P=.002), attending more of the 12 weekly app-based assessments (mean 10.5, SD 1.5 vs mean 9.5, SD 1.8; P<.001), and completing more of the 12 follow-up phone calls (mean 10.8, SD 1.6 vs mean 9.3, SD 1.8; P<.001) than the control group (Table S6 in Multimedia Appendix 2). AEs were infrequent, occurring in 7 patients (7/120, 5.8%) in the IMU group and 9 patients (9/120, 7.5%) in the control group; most AEs were minor and considered related to treatment (Table S7 in Multimedia Appendix 2).