Monitoring Gait Recovery After Total Knee Arthroplasty Using Wearable Sensors: Responsiveness of Gait Accelerations

ABSTRACT

Functional recovery after total knee arthroplasty (TKA) varies widely among individuals, and traditional assessments often fail to detect subtle changes in real‐world walking ability. Wearable sensors offer continuous and objective tracking of gait outside of clinical settings. In this prospective, longitudinal study, thirty‐one patients undergoing unilateral TKA wore thigh‐mounted accelerometers continuously from 2 weeks before surgery through 90 days postoperatively. We analyzed temporal gait features (stride time and duty factor) and frequency‐domain metrics (power of the first five frequencies of triaxial acceleration), along with weekly EQ‐5D‐3L index scores (a summary measure of health‐related quality of life) and visual analog scale (VAS) ratings (patient's self‐reported health on a 0–100 scale). Gait acceleration profiles demonstrated significant postoperative disruptions, particularly during mid‐stance and terminal swing, followed by gradual normalization, approaching baseline by day 90. Gait temporal parameters also exhibited transient alterations and recovery trends. Using LASSO logistic regression, Fourier‐derived gait features discriminated responders and non‐responders based on improvements in EQ‐5D‐3L outcomes above the minimal clinically important difference (AUC: 0.80–0.81). Specific harmonic changes across the vertical, mediolateral, and anteroposterior axes were significantly associated with the perceived health gains. The results indicated that frequency‐based gait metrics from wearable sensors are sensitive digital biomarkers for short‐term functional improvements.

1. Introduction

Total knee arthroplasty (TKA) is a widely accepted and effective intervention for severe knee osteoarthritis (OA), with an increasing incidence due to aging and increasing prevalence of OA [1]. However, recovery trajectories vary widely, and functional outcomes frequently do not meet clinical expectations [2, 3]. This variability highlights the need for improved management strategies, particularly those that focus on monitoring functional recovery.

Conventionally, postoperative recovery following TKA is assessed using clinician‐ and patient‐reported outcome measures (PROMs) during periodic in‐clinic assessments [4]. Although valuable, these methods are subjective, episodic, and often miss real‐world mobility patterns and daily fluctuations [5]. In addition, scheduled follow‐up visits associated with these assessments require significant time and resources from both patients and healthcare providers. As healthcare systems face growing resource constraints, particularly with aging populations, scalable solutions that enable continuous, remote, and automated monitoring are becoming essential for detecting subtle yet clinically meaningful changes in recovery and supporting timely intervention [6].

Recent advances in wearable inertial sensors have made it possible to continuously and objectively monitor patients' walking patterns in their daily environments [7]. Compared to traditional assessments, wearable sensors enable stride‐by‐stride gait analysis over extended periods, providing higher‐resolution insights into functional recovery [8]. While prior studies have linked gait characteristics to clinical outcomes such as pain, function, and quality of life after TKA [9, 10, 11, 12, 13], most relied on intermittent testing or basic metrics like step count and gait speed, which may not capture short‐term functional changes. Furthermore, the association between wearable‐derived gait features and meaningful improvements in pain, mobility, and activity remains inconsistent or underreported [14]. Given that many patients experience lingering pain or delayed functional return despite technically successful TKA [15, 16], more responsive and informative metrics are needed. Detailed gait analysis approaches that capture the full stride trajectory, such as frequency‐domain analysis, have the potential to reveal subtle yet clinically relevant aspects of recovery. Frequency‐domain features can characterize rhythmicity, coordination, and energy distribution throughout the entire gait cycle, providing insights beyond those afforded by discrete metrics alone [17].

The primary aim of this study was to evaluate whether gait features derived from wearable sensors can detect short‐term, clinically meaningful changes in individuals' walking function during the early recovery period following TKA, and to assess their potential as objective tools for monitoring progress in real‐world settings. Specifically, we analyzed the trajectory of gait accelerations across the stride cycle, along with temporal parameters (stride time and duty factor) and frequency‐domain features derived from acceleration signals. These commonly used and readily available features from wearable sensor data capture a comprehensive picture of walking, including phase timing, weight bearing, gait smoothness, coordination, and rhythmicity, and reflect different aspects of functional recovery. For example, decreased stride time has been associated with recovery of gait function following knee surgery [18] as this would indicate an increase in gait speed based on the speed‐frequency power function relationship [19]. Similar utility has been demonstrated for duty factor in postsurgical gait monitoring as decreased duty factor for the surgical limb may indicate an antalgic gait [18]. By comparing week‐to‐week changes in these gait features with corresponding changes in EQ‐5D‐3L index and visual analog scale (VAS) scores, we aimed to determine their responsiveness to perceived improvements in health status. We hypothesized that changes in sensor‐derived gait features would significantly correlate with concurrent changes in EQ‐5D‐3L index and VAS scores during the early postoperative period.

2. Methods

2.1. Study Design and Participants

This prospective longitudinal observational study was conducted at Aalborg University Hospital and Capio Private Hospital, Aalborg, Denmark. Patients diagnosed with knee OA and scheduled for unilateral TKA were recruited between March 2024 and January 2025. Eligible participants were ≥ 18 years old, ambulatory without walking aids, and able to use a smartphone compatible with the study's data transfer app. The exclusion criteria were severe contralateral knee OA (Kellgren‐Lawrence grade IV), recent lower limb or spinal surgery, frailty (Clinical Frailty Scale score ≥ 5), and any condition that interfered with continuous sensor use. Patients with symptomatic contralateral knee OA of lower grades were not excluded unless they were scheduled for contralateral TKA within the study period.

The participants were followed up from 2 weeks preoperatively to 90 days postoperatively, during which wearable sensors continuously collected gait data. PROMs were recorded once preoperatively and weekly postoperatively. Data collection was completed for each participant after 3 months of follow‐up, marking the end of the observation period.

Ethical approval was obtained from the North Denmark Regional Ethics Committee (N‐20230035) and all participants provided written informed consent before enrollment. The study protocol is registered and publicly available at ClinicalTrials.gov (ID: NCT05992064). Reporting followed the STROBE guidelines for observational studies [20].

2.2. Patient‐Reported Outcomes

The EuroQol 5‐Dimension 3‐Level (EQ‐5D‐3L) [21] questionnaire was used to assess health‐related quality of life throughout the study period. The EQ‐5D‐3L and accompanying VAS have been shown to be valid and reliable instruments for assessing health status in patients undergoing TKA and are widely used in both clinical and research settings [22, 23, 24]. The EQ‐5D‐3L includes five dimensions of mobility, self‐care, usual activities, pain/discomfort, and anxiety/depression, each rated on three levels, as well as a VAS capturing participants' self‐rated health, where 0 represents the worst and 100 the best imaginable health state. Baseline EQ‐5D‐3L assessments were completed during the preoperative visit at least 2 weeks before surgery. After the operation, the participants received invitations via short message service (SMS) to complete the weekly EQ‐5D‐3L electronically using the REDCap platform hosted by the North Jutland region, Denmark. EQ‐5D‐3L index scores were computed using a custom algorithm based on Danish population‐specific value sets [25]. Both the index and VAS scores were calculated at baseline (preoperatively) and weekly for 90 days postoperatively to monitor longitudinal changes in perceived health status.

2.3. Wearable Sensor Data

The participants were equipped with a waterproof wearable accelerometer (SENS Motion, SENS Innovation ApS), which has been previously validated for clinical use in patients with gait impairment [26, 27]. The sensors were positioned by trained research staff on the lateral distal thigh of the surgical limb and aligned with the longitudinal axis of the femur using standardized anatomical landmarks (Figure 1). This location was chosen because the SENS Motion proprietary activity classification and gait detection algorithm is specifically optimized for this placement, leveraging its characteristic movement patterns during walking. The placement was marked on the skin with a semi‐permanent marker to ensure consistent reattachment. The participants were instructed to reapply the sensor to the same marked location if they were detached.

Figure 1.

Illustration of wearable sensor placement on the lateral aspect of the thigh.

Monitoring began 2 weeks preoperatively and continued for 90 days postoperatively. The sensors recorded raw triaxial linear acceleration (along vertical, anteroposterior, and mediolateral axes) at 12.5 Hz. Participants were instructed to wear the device continuously, including during sleep and water‐based activities, except on the day of surgery, when they were advised to remove the sensor and reattach it afterward using the provided materials. If the sensor detached at any point, the participants simply reattached it to the same location; however, no interruptions to the monitoring protocol were reported.

Data were synchronized daily to a secure cloud platform via an application installed on the participants' smartphones. The application used for data transfer was developed and maintained by SENS Innovation ApS, the manufacturer of the wearable sensors. The app automatically synchronized raw triaxial acceleration data from the sensor to a secure, encrypted cloud server using GDPR‐compliant protocols. No data were stored on the participant's phone beyond temporary buffering, and access to the cloud platform was restricted to authorized study personnel. The research team monitored the uploads and contacted the participants if the data transmission had lapsed for more than 3 days. The app did not perform any filtering or preprocessing of the data; all signal processing, gait detection, and feature extraction were conducted by the authors using custom Python scripts after data retrieval. None of the authors was involved in the development or sponsorship of the data transfer app.

2.4. Gait Detection and Processing

Walking segments were identified using the previously validated proprietary SENS Motion® algorithm [28], which labels upright walking bouts based on activity classification synchronized with acceleration data. This algorithm was specifically developed and validated for use with the 12.5 Hz sampling frequency of the sensor, leveraging the movement characteristics of the thigh segment to enable reliable detection of walking at low frequency [28]. To ensure the analysis of sustained and representative gait activity, only walking bouts exceeding 60 s were considered. Extended walking sessions have been demonstrated to yield more stable gait characteristics and decreased variability compared to shorter sessions, which may be affected by environmental constraints and transient movement patterns [29]. The three longest valid bouts per day were selected to capture consistent, higher‐quality walking data and reduce the impact of intra‐day variability in walking behavior.

Within each bout, continuous walking segments were isolated by identifying step sequences using a peak detection algorithm applied to the vector magnitude of triaxial acceleration signals. This composite signal, calculated as the square root of the sum of squares of the vertical, anteroposterior, and mediolateral axes, provides a stable estimate of total body acceleration and enables reliable estimation of vertical acceleration and gravity vector orientation [30]. Peak detection applied to the vector magnitude signal has been shown to accurately identify steps and gait cycles, even when using low‐frequency accelerometry data, as validated against high‐resolution reference systems in individuals with and without knee OA [31]. Stride cycles from continuous gait bouts were identified using custom Python scripts based on detected foot‐contact and foot‐off events according to a previously published algorithm [32]. These cycles were then time‐normalized by resampling the data.

2.5. Gait Temporal Parameter Analysis

Key temporal gait parameters, specifically stride time (duration of the gait cycle) and duty factor (proportion of the gait cycle spent in the stance phase), were derived from the individual stride cycles. These variables were selected since they are among the most commonly reported features in gait analysis and reflect clinically relevant aspects of gait timing, rhythm, and load distribution, which are often impaired following TKA [33, 34]. Percentage changes in stride time and duty factor were evaluated relative to the baseline (average preoperative values), and recovery patterns were examined based on their trajectories over the 90‐day postoperative period.

2.6. Statistical Parametric Mapping Analysis

Statistical parametric mapping (SPM) was applied to evaluate time‐dependent changes in the gait cycle profiles of acceleration across the pre‐ and postoperative periods. SPM is a statistical approach that uses random field theory to detect significant differences in continuous time‐series data, such as normalized gait waveforms, while controlling for multiple comparisons across the trajectory [35]. SPM was performed on vector magnitude acceleration signals, which are orientation invariant. Acceleration signals were low‐pass filtered at 6 Hz using a fourth‐order Butterworth filter before the analysis. Time‐normalized gait cycles derived from vector magnitude acceleration signals were grouped into four periods: preoperative (PreOP), 0–30 postoperative days (M1), 31–60 postoperative days (M2), and 61–90 postoperative days (M3). For each participant, the average gait waveform within each period was computed from all valid strides extracted from the three longest walking bouts per day.

2.7. Gait Frequency Components

Fast Fourier Transform (FFT) magnitudes were squared to obtain the power of frequencies [36] for each gait cycle from the filtered acceleration signals along the vertical, anteroposterior, and mediolateral axes and from the vector magnitude. Only the power of the first five frequencies from each axis and vector magnitude were included in the analysis, reflecting the most dominant harmonic components of human walking and the limitations given by the 12.5 Hz sampling rate of the wearable sensors [37]. Changes in these frequency components were evaluated relative to the baseline (average preoperative values), and recovery patterns were examined using the frequency components from the vector magnitude using their trajectories over the 90‐day postoperative period. Responsiveness analyses were conducted using the frequency components from the vertical, anteroposterior, and mediolateral axes.

2.8. Assessing Responsiveness to Recovery

The frequency domain features of gait (power of frequency components) were used to evaluate whether changes in the gait acceleration profiles were responsive to perceived health improvements during the early postoperative recovery period.

To quantify short‐term changes, the weekly differences in the power of the first five frequencies were calculated relative to the preceding week. Responsiveness was anchored to weekly changes in patient‐reported EQ‐5D‐3L outcomes. Participants were classified as EQ‐5D responders if their EQ‐5D index score increased from that of the previous week (ΔEQ‐5D Index ≥ 0.04). For VAS, responders were defined similarly by a positive weekly change (ΔEQ‐5D VAS ≥ 4). These thresholds correspond to the lower bounds of the 95% confidence intervals for minimal clinically important differences reported by Yapp et al. [38]. Since these MCIDs were derived from cumulative changes over 6 to 12 months, we used the lower bounds as conservative estimates appropriate for detecting smaller, yet meaningful, week‐to‐week changes in recovery trajectories during the early postoperative period.

2.9. Statistical Methods

Continuous variables were summarized using means and standard deviations or medians with interquartile ranges depending on the distribution. Categorical variables were presented as counts and percentages. Missing data were assessed separately for sensor‐derived outcomes and PROMs. No data imputation was performed. All available data were included in the descriptive summaries and visualizations. Only participants with complete weekly PROMs were included in the study for analyzes requiring complete repeated measures (e.g., responsiveness models).

For each participant, percentage changes from baseline were calculated for temporal gait parameters (stride time and duty factor) and for the first five frequency components of the vector magnitude signal. Baseline was defined as the average of all valid preoperative measurements during the 2 weeks before surgery. To evaluate the recovery trajectories, generalized additive mixed models (GAMMs) were used to visualize smoothed trends in daily means over time while accounting for repeated measures. Linear mixed‐effects models (LMMs) were employed to quantify the significance of longitudinal changes in both temporal and frequency‐domain gait features.

SPM was applied to the vector magnitude signals to assess the temporal changes in gait acceleration patterns. For each participant, the gait trajectories were averaged within each of the four predefined periods (PreOP, M1, M2, and M3) to ensure balanced contributions. One‐way repeated‐measures analysis of variance (ANOVA) was conducted using the SPM1D framework to test for statistically significant differences across the gait cycle (α = 0.05). Post‐hoc paired t‐tests between the preoperative period and each postoperative interval were used to identify specific temporal regions of significant deviation. As multiple comparisons were performed across postoperative time points, p‐values were adjusted using the Bonferroni method to control for type I error across the set of pairwise tests.

Logistic regression with LASSO (Least Absolute Shrinkage and Selection Operator) [39] regularization was used to evaluate whether weekly changes in frequency‐domain gait features could classify responder status. This regularization technique performs both variable selection and model fitting, making it well suited for high‐dimensional and correlated data. LASSO regression has been increasingly applied in clinical gait research to identify discriminative features and construct interpretable prediction models [40, 41, 42, 43, 44]. Model tuning was performed using 10‐fold cross‐validation to identify the optimal regularization parameter (λ), and model performance was assessed using the area under the receiver operating characteristic curve (AUC). While the data structure included repeated measures, we attempted to fit mixed‐effects models with participant‐level random intercepts; however, these models failed to capture subject‐level variability due to singular fit. LASSO was therefore retained as the primary method due to its robustness and generalizability in this context.

The sample size was based on detecting a moderate within‐subject effect (Cohen's d ≈ 0.5), a conventional benchmark for clinically relevant changes in longitudinal rehabilitation studies. With α = 0.05% and 80% power, at least 30 participants were required, allowing for up to 20% attrition over 90 days. While no formal power calculation was conducted for the GAMMs or statistical parametric mapping, this sample size aligns with previous sensor‐based observational studies of recovery after joint arthroplasty [45, 46, 47].

All statistical analyzes were conducted using R software (version 4.2.2). SPM analysis was conducted in Python using the SPM1D package (version 0.4.18).

3. Results

3.1. Participant Characteristics

A total of 31 participants were enrolled in the study, including 17 females (55%) and 14 males (45%). The median age was 68.3 years (range: 44.7–86.6), with female participants having a median age of 69.0 years (44.7–82.9) and males 65.4 years (55.9–86.6).

The median stature was 170.0 cm (156–191 cm), weight was 85.0 kg (60–122 kg), and body mass index (BMI) was 28.4 kg/m² (21.8–39.1). Females had a median stature of 165.0 cm, weight of 78.0 kg, and BMI of 28.0 kg/m²; the corresponding values for males were 181.0 cm, 98.0 kg, and 29.0 kg/m².

Eighteen participants (58.1%) had at least one medical comorbidity. The most frequent comorbidity was hypertension (six participants, 33.3%), followed by other conditions (three participants, 16.7%), heart disease, and diabetes (two participants each, 11.1%). Single cases (5.6% each) of depression, previous thyroid cancer, or asthma were also observed. Nineteen participants (61.3%) had a history of surgery, with 10 (52.6%) involving the ipsilateral side and 9 (47.4%) involving the contralateral side. Anatomically, these included 16 knee surgeries (84.2%) and single lumbosacral, thigh, and hip surgeries (5.3% each). Regarding the procedure type, 13 (68.4%) were soft tissue procedures, 5 (26.3%) were arthroplasties, and 1 (5.3%) was an osteotomy or osteosynthesis procedure.

3.2. Missing Data

The sensor data completeness was high, with less than 1% of the daily recordings missing during the 12‐week follow‐up period. In contrast, the EQ‐5D‐3L descriptive and VAS dimensions had a mean missing rate of 16.2% across all participant‐weeks. The pattern of missing recordings was random, with no evidence of systematic dropout over time. A subset of 17 participants completed all weekly PROMs and was included in the responsiveness analyzes.

3.3. Patient‐Reported Outcomes

Following surgery, both the EQ‐5D‐3L index and VAS scores declined, reaching their lowest values within the first two postoperative weeks (Figure 2). The VAS scores showed a statistically significant decline at week 1 (p  <  0.01) and a significant improvement by week 12 (p <  0.05) compared to the preoperative levels. Although the EQ‐5D‐3L index scores followed a similar recovery trajectory, no statistically significant changes were observed, although the initial decline approached statistical significance. By day 90, both PROM measures surpassed the preoperative levels, indicating an overall improvement in self‐reported health status.

Figure 2.

Postoperative trajectory of EQ‐5D‐3L scores after total knee arthroplasty (TKA). This figure shows the weekly mean scores for the EQ‐5D‐3L Index (top panel) and EQ‐5D Visual Analog Scale VAS (bottom panel) over the 90‐day postoperative period. The dashed red line on day 0 indicates the day of surgery. Error bars represent 95% confidence interval.

3.4. Gait Acceleration Analysis

SPM revealed distinct time‐dependent alterations in gait dynamics during the postoperative period. The analysis of the vector magnitude signal demonstrated statistically significant differences across the gait cycle, with the most prominent deviations occurring during the mid‐stance and terminal swing phases (F > 5.724, p < 0.05; Figure 3).

Figure 3.

Statistical parametric mapping (SPM) analysis of the vector magnitude of the gait cycle across perioperative time points. The top panel displays the mean vector magnitude acceleration profiles (with 95% confidence intervals) across the normalized gait cycle (0%–100%) at four time points: preoperative (PreOP) and postoperative months 1 (M1), 2 (M2), and 3 (M3). The bottom panel shows the SPM{F} curve from a repeated‐measures ANOVA, identifying significant differences in the vector magnitude profiles across the time points. The shaded region indicates where the F‐statistic exceeded the critical threshold (F* = 5.724, α = 0.05), indicating significant temporal changes in the gait acceleration during recovery.

Paired comparisons between preoperative and postoperative data showed marked reductions in vector magnitude 1 month after surgery, particularly within the early stance and mid‐swing intervals (t > 4.244, p < 0.02; Figure 4A). After 2 months, although gait deviations persisted, the magnitude of the change was attenuated and localized primarily in the loading response phase (t > 4.228, p < 0.02; Figure 4B). At 3 months postoperatively, gait patterns largely returned to baseline, with no suprathreshold clusters exceeding the critical t‐statistic (t > 4.215, p > 0.05; Figure 4C), indicating substantial normalization of vertical loading characteristics by the end of the 90‐day follow‐up.

Figure 4.

SPM analysis of paired differences in vector magnitude gait acceleration between preoperative and postoperative time points. Panels A, B, and C show comparisons between the preoperative gait cycle (blue) and the corresponding postoperative months: 1 month (M1, red), 2 months (M2, green), and 3 months (M3, purple), respectively. The top plot in each panel displays the mean vector magnitude acceleration profiles with 95% confidence intervals across the normalized gait cycle (0%–100%). The bottom panels show the corresponding SPM{t} statistics from the paired t‐tests. The shaded regions in the t‐statistic plots denote segments of the gait cycle where the difference between the preoperative and postoperative signals exceeded the critical threshold for statistical significance (α = 0.02).

3.5. Recovery of Gait Frequency Components

Figure 5 displays the percent change from the baseline in the power of the first five Fourier coefficients (FC 1–5) of the vector magnitude acceleration signal. Most coefficients showed an immediate decrease in power following surgery, followed by a gradual recovery trend toward baseline over the 90‐day period. Linear mixed‐effects models revealed statistically significant changes over time for all five coefficients (all p < 0.001). Specifically, FC 3 exhibited the largest transient reduction in the early postoperative phase (intercept = − 46.2%, β = 4.48, p < 0.001), suggesting a marked early disruption in rhythmic gait structure. FC 1 and FC 2 also showed substantial early reductions (intercepts = − 52.2% and − 30.1%, respectively), with significant positive slopes indicating recovery over time (FC 1: β = 5.80, p < 0.001; FC 2: β = 6.53, p < 0.001). FC 4 exhibited a smaller but still significant increase (β = 2.68, p = 0.001), while FC 5 showed the strongest recovery slope (β = 7.07, p < 0.001), despite its initial reduction (intercept = − 42.1%).

Figure 5.

Percentage change from baseline in the power of the first five frequency components (FC1–FC5) of the vector magnitude acceleration signal over the 90‐day postoperative period. Each panel shows individual participant data points and the group‐level recovery trajectory with days since surgery on the x‐axis. The solid line represents the smoothed group mean (with shaded 95% confidence interval), and the dashed line marks zero change from the baseline.

3.6. Gait Temporal Parameters

Figure 6 shows the percent change from baseline for the gait cycle period and duty factor aggregated across all participants.

Figure 6.

Recovery trajectory of stride time and duty factor following TKA. This figure shows the percentage change from baseline in two temporal gait parameters, duty factor (top panel) and stride time (bottom panel), over the 90‐day postoperative period. Gray dots represent daily mean values, whereas blue lines depict generalized additive model (GAM) trends. The dashed red line marks the day of surgery (day 0).

Stride time (gait cycle period) exhibited an initial increase after surgery, reaching its peak within the first 3 weeks, followed by a gradual decline toward baseline. By 90 days, the stride time returned to or slightly below the preoperative level. The LMM showed that the stride time percentage change decreased significantly during the postoperative period (β = −1.28, SE = 0.24, p < 0.001), indicating a progressive return toward baseline.

The duty factor demonstrated an immediate postoperative reduction, with gradual recovery back to baseline values over time. The LMM detected a significant increasing trend across the recovery period (β = 0.81, p = 0.008), consistent with a return toward the baseline stance behavior.

Both parameters displayed moderate interindividual variability captured by participant‐level random effects.

3.7. Predictive Performance of Fourier Coefficients for Responsiveness

The LASSO model trained on changes in power of the frequency components from the anteroposterior, mediolateral, and vertical axes achieved an AUC of 0.80 for predicting weekly improvements in EQ‐5D‐3L Index scores. At the optimal regularization parameter (λ = 0.02), the model retained four directional power features at specific frequency components: changes in power at the first‐frequency power of the vertical and mediolateral axes and the second‐ and fifth‐frequency power of the anteroposterior axis. Selective inference confirmed that all four power features were statistically significant (all p < 0.001).

To predict weekly improvements in EQ‐5D‐3L VAS scores, the LASSO model, trained on weekly changes in the power of the frequency components across the vertical, mediolateral, and anteroposterior axes, achieved an AUC of 0.81 at an optimal λ of 0.035. The model retained four directional power features at specific frequency components: changes in power at the first and third frequencies of the vertical axis, fourth frequency of the mediolateral axis, and fifth frequency of the anteroposterior axis (all p < 0.001).

The ROC curves for both models are presented in Figure 7.

Figure 7.

Receiver operating characteristic (ROC) curves predicting weekly improvements in EQ‐5D‐3L outcomes based on gait features. The ROC curves illustrate the predictive performance of the changes in power of the frequency components of acceleration signals for classifying weekly improvements in patient‐reported outcomes. The blue curve represents predictions for the EQ‐5D‐3L Index, and the orange curve represents the VAS score.

4. Discussion

This exploratory study aimed to evaluate whether sensor‐derived gait acceleration features can reflect early functional changes during recovery after TKA, and to assess their potential clinical utility. We found significant postoperative disruptions in gait acceleration profiles, especially during the mid‐stance and terminal swing phases, with near‐complete recovery by the end of the 3 months follow‐up. Fourier‐derived gait features exhibited clear sensitivity to week‐by‐week improvements in patient‐reported health status, as compared to EQ‐5D scores.

Previous studies have predominantly assessed TKA recovery through intermittent subjective assessments, such as patient‐ and clinician‐reported outcome measures and periodic clinical gait evaluations [48, 49, 50]. It is well established that gait disturbances are common following TKA and tend to improve over time [51]; however, most existing studies rely on cross‐sectional or episodic measurements that often fail to capture subtle day‐to‐day changes in gait quality and function. While some recent work has begun to explore wearable sensor data [52], these efforts have largely focused on basic metrics such as step counts or total activity levels [7, 53, 54]. Our study builds on the well‐known recovery pattern of gait after TKA by using continuous stride‐level acceleration data and frequency‐based analysis to examine how gait changes during the recovery period. To our knowledge, this is the first study to demonstrate that these gait features are not only sensitive to short‐term postoperative changes, but also closely track weekly improvements in patient‐reported outcomes. However, while these findings demonstrate strong temporal responsiveness, claims regarding clinical precision or relevance should be interpreted cautiously, as these outcomes were not directly assessed in this study. Moreover, as this was an exploratory study, it was not powered for all individual statistical comparisons.

The postoperative trajectory of EQ‐5D‐3L scores in this study provided meaningful insights into patient‐perceived recovery after TKA. Both the EQ‐5D index and VAS scores declined sharply during the first postoperative week, reflecting the immediate physiological and psychological effects of surgery. This initial decline was followed by steady improvement, with scores surpassing preoperative levels approximately 30 days after surgery and continuing to increase throughout the 90‐day follow‐up period. These trends suggest that, by 3 months postoperatively, patients perceived a higher QoL and overall health status than at baseline, although the differences were not statistically significant. A cohort study of fast‐track TKA patients reported similar improvements, with VAS scores exceeding baseline levels as early as 6 weeks postoperatively [55]. A larger multicenter study also demonstrated substantial gains across all EQ‐5D dimensions, although the follow‐up extended well beyond 3 months [56]. These findings reaffirm EQ‐5D‐3L as a responsive and clinically meaningful tool for monitoring recovery. Rather than replacing PROMs, wearable gait metrics may serve as complementary measures, offering continuous, objective data that can enhance recovery tracking when integrated with established subjective assessments. However, further research is required to determine whether these metrics provide sufficient added value to justify their broad clinical implementation.

SPM analyzes showed reduced vector magnitude acceleration during the first 2 months after surgery, particularly during the mid‐stance and terminal swing phases, reflecting impaired limb loading and altered swing mechanics, consistent with previous reports of asymmetry, reduced stability, and compensatory unloading in early recovery after TKA [57, 58, 59]. By month three, these abnormalities had largely resolved, with no suprathreshold differences from baseline, in line with studies showing partial recovery of gait parameters within 3–6 months  [50, 60]. Temporal gait analysis also indicated a transient increase in stride time [19] and a reduction in duty factor [61] postoperatively, followed by partial recovery over 90 days. We recognize that gait parameters may not fully return to preoperative values within this period; however, our aim was to assess early recovery trends and not complete normalization. These early changes, reflecting compensatory adaptations like reduced stance time and cautious foot placement [62], may still offer clinical value for tracking progress, identifying deviations, and guiding rehabilitation. Despite measurable improvements by 90 days, subtle stride abnormalities persisted, supporting the need for continued monitoring beyond the early postoperative phase [51].

Our responsiveness analysis demonstrated that specific spectral components of gait acceleration were highly responsive to short‐term patient‐perceived changes in health status. Weekly changes in Fourier‐derived gait features were strongly associated with week‐by‐week improvements in EQ‐5D outcomes, with predictive models achieving AUCs of 0.80 and 0.81 for the EQ‐5D index and VAS scores, respectively. We selected the EQ‐5D‐3L as the primary patient‐reported anchor because it is a validated, widely adopted measure in TKA populations and captures key health domains that are directly influenced by changes in gait and function [22, 63]. Its brevity and suitability for frequent remote administration made it particularly feasible for weekly follow‐ups, supporting a high‐resolution analysis of recovery dynamics. In contrast, longer joint‐specific PROMs such as Knee injury and Osteoarthritis Outcome Score or Oxford Knee Score, would likely impose a higher respondent burden and increase the risk of missing data during weekly remote follow‐up.

A major strength of this study is the inclusion of continuously collected gait data over several weeks, which provides high‐resolution insight into individual recovery trajectories during early rehabilitation. This prospective longitudinal design combined with continuous gait monitoring in free‐living environments enhances ecological validity and enables the detection of short‐term functional changes during recovery. However, this study had several limitations. The relatively small sample size may limit the generalizability of our results, and larger studies are required to confirm these findings in a broader patient population. While all participants contributed gait data, only 17 had complete records for the week‐to‐week responsiveness analysis due to missing EQ‐5D responses. Despite efforts to minimize this through automated reminders and follow‐up, nonresponse remains a challenge. This highlights a broader limitation of the PROM‐based follow‐up, which may be subject to participant fatigue or disengagement over time. While sensor adherence was high, with less than 1 percent of daily recordings missing, minor data gaps occurred but appeared random and unlikely to introduce systematic bias. Additionally, the relatively low sampling frequency (12.5 Hz) may have affected the precision of certain temporal gait parameters and frequency components. However, the practical strength of this system is that a low‐burden, lightweight sensor capable of long‐term use and simplified data handling is still able to capture responsive and clinically meaningful changes in gait during the recovery period in everyday settings.

The findings of this study highlight the potential value of wearable technology for enhancing postoperative care and rehabilitation strategies after TKA. Continuous gait monitoring using wearable accelerometers provides clinicians with objective high‐resolution data on functional recovery, potentially enabling personalized and timely intervention strategies. By identifying real‐time deviations from expected recovery trajectories, healthcare professionals can implement proactive strategies to tackle emerging challenges or suboptimal progress, potentially improving long‐term outcomes. Extending monitoring beyond the initial 90‐day period may offer valuable insights into sustained gait adaptations and late‐phase recovery. Moreover, incorporating these objective gait measurements into clinical decision support tools and platforms that provide feedback to patients could enhance rehabilitation strategies, boost patient involvement, and ultimately lead to greater patient satisfaction and better functional outcomes. However, as an exploratory study, our analysis does not fully establish clinical utility, and additional work is required to validate these metrics in diverse patient populations and assess their feasibility, cost‐effectiveness, and clinical impact.

5. Conclusion

Wearable accelerometer‐based gait analysis offers a promising approach for tracking functional recovery following TKA. This exploratory study demonstrated that continuously collected gait features were responsive to short‐term changes in the patient‐reported health status. The ability to monitor gait recovery on a day‐to‐day and week‐to‐week basis provides high‐resolution insights into postoperative function that may complement traditional outcome measures. These findings support the potential role of wearable sensor systems in postoperative care and rehabilitation. Future research should focus on validating these metrics in larger, diverse cohorts and determining their added clinical value in guiding recovery and decision‐making.

Author Contributions

All authors (AG, PDC, AK, JR, RDG, SK, and OR) conceptualized and designed the study. The methodology was developed by AG, PDC, AK, JR, and RDG. PDC led the data collection. Data analysis was performed by AG. The manuscript was drafted by AG and revised and edited by all co‐authors. Supervision was provided by SK, OR, AK, and JR.

Supporting information

N‐20230035 Endelig godkendelse.

Ghaffari A., Clasen P. D., Kappel A., et al., “Monitoring Gait Recovery After Total Knee Arthroplasty Using Wearable Sensors: Responsiveness of Gait Accelerations,” Journal of Orthopaedic Research 43 (2025): 2165‐2177. 10.1002/jor.70058.