Exploring the Impact of Behavioral Engagement Indicators on Pain Outcomes in AI-Enabled Musculoskeletal Rehabilitation: A Real-World Investigation

Abstract

Purpose

This study aimed to examine whether behavioral engagement indicators reflecting timing, routine formation, and continuity predict last pain scores of an artificial intelligence (AI)–enabled rehabilitation program in individuals with chronic musculoskeletal (MSK) spinal pain.

Methods

This ad-hoc analysis of a real-world clinical registry (June 19, 2020 to November 8, 2024) across the United States (N=19,866) included adults with chronic MSK spinal pain undergoing AI-assisted, fully-remote rehabilitation delivered under physical therapists oversight and combining exercise with real-time biofeedback asynchronously monitored, education, and behavior change (lasting up to 12 weeks). Primary outcome was the last pain score, reported through an 11-point Numeric Pain Rating Scale. Behavioral engagement candidate predictors (derived from session log data) comprised early-phase (first three weeks) – treatment initiation (days until first session), early routine formation (consistent weekday session completion), and early continuity (no missed weeks) – and full-period indicators – weekday routine consistency and existence of one-week interruption of sessions (≥7 consecutive days without sessions). Structural equation modeling assessed associations with pain outcomes, adjusted for confounders. Moderation analyses evaluated subgroup differences for baseline pain severity, age, race/ethnicity, and BMI strata.

Results

All early-phase behavioral engagement indicators – timely treatment initiation, routine formation, and engagement continuity – were associated with lower last pain scores, albeit at very small effect sizes (Cohen’s f²=0.01). Across the full intervention period, greater consistency in weekday session completion was associated with lower pain (β=−0.21, P<0.001), whereas experiencing a full-week interruption predicted higher pain (β=0.14, P<0.001), with small effect sizes (Cohen’s f²=0.02). No moderation was observed for any of the tested subgroups, except for patients with obesity, where a one-week interruption had a stronger effect on pain. Models explained variance was 25%.

Conclusion

Behavioral engagement indicators capturing treatment initiation, routine formation, and continuity were significantly associated with pain outcomes after AI-enabled MSK rehabilitation, albeit at small magnitude. Behavioral engagement may represent potentially modifiable targets for intervention, requiring further validation.

Introduction

Chronic spinal pain is a highly prevalent and persistent condition¹ recognized as a global public health problem,² affecting multiple domains of daily life, including physical and emotional functioning, social interaction, and work ability. Exercise-based rehabilitation is a cornerstone of care for chronic spinal pain^3–6 with demonstrated broad benefits for pain management and quality of life.⁷

Adherence to rehabilitation is a critical determinant of treatment effectiveness, with exercise dosage associated with treatment outcomes such as pain and function.^8,9 Existing predictive research in musculoskeletal rehabilitation has predominantly assessed exercise-dosage metrics, such as the total number of sessions, treatment duration, or exercise volume.^8–15 Although these metrics quantify how much treatment a patient completes, they do not fully capture how engagement develops and is sustained over time. Behavioral science has studied what motivates engagement patterns. In the specific case of physical activity, stable routines, consistency, and continuity facilitate habit formation and long-term engagement with exercise.^16,17 However, whether these findings translate to clinical outcomes in musculoskeletal (MSK) rehabilitation remains unknown.

The emergence of remote alternative rehabilitation delivery modes makes behavioral patterns particularly relevant, as it creates further patient autonomy and requires greater self-regulation^18,19 in determining when sessions are initiated, organized, and sustained over time.²⁰ Remote rehabilitation has demonstrated comparable clinical outcomes to in-person care,^21–23 while also generating objective, time-stamped data that provide an opportunity to examine engagement beyond traditional measures. However, engagement behaviors specifically associated with recovery remain largely unexplored and their identification may inform strategies to optimize treatment outcomes.¹⁰

To address this evidence gap, the present study aimed to evaluate whether behavioral engagement indicators, including timing, routine formation, and continuity, predict pain outcomes following an artificial intelligence (AI)-enabled rehabilitation program in patients with chronic spinal pain. We hypothesized that these behavioral engagement indicators would be significant predictors of pain outcomes following the intervention.

Materials and Methods

Study Design

This is an ad-hoc analysis of data from an ongoing real-world, clinical registry investigation of patients with chronic spinal pain who underwent an AI-enabled care program. Participants were prospectively enrolled under a study approved by the New England IRB (No. 120190313) and registered at ClinicalTrials.gov on September 17, 2019 (NCT04092946); or a subsequent study approved by the Advarra IRB (Pro00063337) and registered at ClinicalTrials.gov on June 14, 2022 (NCT05417685). Data included in the present analysis were collected between June 19, 2020, and November 8, 2024. The present study adhered to the Declaration of Helsinki and TRIPOD (Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis) guidelines.²⁴ According to the PROGRESS staged continuum for developing prognostic research,²⁵ this study corresponds to an initial stage focused on examining associations between candidate predictors and pain outcomes.

Study Population

Participants were adults (≥18 years) residing across all states of the US, beneficiaries of employer-sponsored health plans covering the AI-enabled care program, and reported chronic spinal MSK pain (ie, persistent or recurrent neck or low back pain with ≥3 months duration).²⁶ Exclusion criteria included: (1) medical contra-indications to perform ≥20 minutes of light-to-moderate intensity exercise; (2) active cancer treatment; (3) unscreened clinical red flags (eg, rapidly progressing neurological deficits); (4) conditions limiting the ability to understand simple and complex motor commands, or independently engage in home-based exercise; and (5) missing baseline assessment or all reassessments.

All participants provided electronic informed consent before enrollment.

Intervention

The AI-enabled care program, detailed elsewhere,^27,28 consisted of fully-remote rehabilitation combining exercise, education, and cognitive behavioral therapy (CBT), based on established clinical guidelines for MSK rehabilitation^3–5 (detailed in Supplementary Table 1).

Patients enrolled in the program through a dedicated digital platform and completed a baseline assessment collecting demographic and clinical information, being assigned to a licensed physical therapist. Following enrollment, each patient had an initial video call with the physical therapist who conducted a clinical assessment and prescribed a personalized treatment plan. A US Food and Drug Administration–listed Class II medical device, including a tablet interface, displayed session content and provided real-time audiovisual biofeedback on movement performance by counting repetitions and identifying movement errors and kinematic deviations, using either sensors (initially employed in an earlier implementation period using wearable motion trackers with inertial measurement units placed on the patient’s trunk with straps) or camera-based (markerless full-body tracking using the embedded camera of a dedicated tablet) motion-tracking technology. Tracking modality reflected technological implementations rather than patient-level clinical criteria. Both sensors and tablet are proprietary to Sword Health. Session data were stored in a cloud-based portal and accessed by the physical therapist, enabling remote, asynchronous monitoring, and intervention adjustments according to the patient’s clinical needs. An AI-assistive tool, described previously,²⁹ supported clinical monitoring by streamlining workflows, flagging patients requiring attention in the portal and proposing potential program adjustments, while the physical therapist retained responsibility for all final clinical and prescription decisions.

The frequency and duration of exercise sessions were individualized by the assigned physical therapist according to each patient’s clinical condition, progress, and rehabilitation goals. Patients were typically recommended to complete at least three sessions per week, although this was not a mandatory minimum and could be adjusted throughout the program. The intervention typically lasted up to 12 weeks.

Education and self-paced CBT were delivered via a smartphone app (Sword Health, US), comprising pre-recorded written/audio materials and interactive modules, developed by a multidisciplinary team including psychiatrists and psychologists based on prior research.^30,31 Ongoing communication between the patient and physical therapist was supported through multiple channels, such as chat or video calls.

Outcome

The primary study outcome was the last pain score reassessment completed, analyzed through the 11-point Numeric Pain Rating Scale (NPRS). This validated and widely used measure captures average pain felt over the last 7 days (0: no pain to 10: worst pain).^32,33 Being consistently identified by patients as one of the most important outcomes in MSK conditions, pain constitutes a core domain in chronic pain research.^33,34 Reassessments were collected at 5-, 9-, 15-, 27-, and 30-sessions, reflecting the fully-remote and asynchronous nature of the intervention, in which patients completed sessions at varying cadences. Although discharge typically occurred within 12 weeks, its timing depended on each patient’s clinical progression, as determined by the assigned physical therapist, and patients could discontinue before discharge. Session-based reassessments, therefore, allowed greater standardization of treatment exposure than fixed timepoints.

Candidate Predictors

Behavioral engagement indicators were computed to capture distinct aspects of how patients initiated and participated in the AI-enabled program, including initiation timing, routine formation, and continuity in session completion, informed by behavioral research highlighting their relevance on health-related behaviors (further details in Table 1).^{16,17,35–40} Indicators were derived from objectively logged device data, enabling precise measurement and reducing social desirability bias from self-report.

Table 1.

Description of Behavior Engagement Indicators Used as Candidate Predictors

Timing	Metric	Definition	Supporting Evidence
Early-phase (first 3 weeks)	Initiation timing	Number of days from program enrollment to completion of the first rehabilitation session.	[38–40]
	Early routine formation	Binary indicator of early weekday routine formation: 1 if the participant completed ≥1 session on the same weekday in at least two of the first three treatment weeks, 0 otherwise	[16,17,36]
	Early continuity	Binary indicator of early continuous engagement: 1 if the participant completed ≥1 session in each of the first three weeks (no missed weeks), 0 otherwise	[37]
Full intervention period	Weekday routine	Proportion of treatment weeks that included ≥1 session on the modal weekday (defined as the weekday with the greatest number of completed sessions across the full intervention period).	[16,17,36]
	One-week interruption	Binary indicator of the presence of an interruption: 1 if the participant had ≥7 consecutive days without performing any sessions during the intervention period, 0 otherwise	[35,37]

Indicators were computed for the first 3 weeks (early phase) and the full intervention period to capture both early and sustained engagement patterns impact on pain outcome. The 3-week window was selected based on prior evidence indicating that significant clinical improvement is expected and is commonly reassessed during this period,^41,42 and that early improvements predict longer-term outcomes.^43,44 Early participation has also been linked with subsequent adherence in digital rehabilitation contexts.⁴⁵ This window also enables recurring weekly patterns, including routine formation and continuity, to be meaningfully captured while preserving the timeliness of these indicators.

Confounders

Analyses were adjusted on potential confounders based on prior evidence^46–49 and the disjunctive cause criterion,⁵⁰ including: age, gender (men; non-men [women and non-binary]), body mass index (BMI), race/ethnicity (Asian, African American or Black, Hispanic or Latino, non-Hispanic white, and Other), education (Graduate, Bachelor’s degree, Some college, and High school diploma or less), socioeconomic status (measured by social deprivation index – SDI – and categorized in fifths: C1 (0–20), C2 (20–40), C3 (40–60), C4 (60–80), and C5 (80–100)), analgesic intake (yes; no), treatment duration and baseline pain, depression (PHQ-9), anxiety (GAD-7), and fear-avoidance beliefs (FABQ-PA).²⁷ Treatment duration was included to account for variation in the time elapsed between baseline and the last completed pain reassessment across participants.

Sample Size

Sample size estimations followed guidelines indicating at least 20 participants per estimated parameter.⁵¹ The planned model included up to 230 free parameters, encompassing regression paths, variances, covariances, intercepts, and residual variance. Based on this criterion, the minimum required sample size was approximately 4600 participants. The study sample included over 10,000 participants, substantially exceeding this threshold, and supporting stable parameter estimation and adequate statistical power.

Statistical Analysis

Descriptive statistics were used to summarize the cohort’s baseline characteristics. Continuous variables are reported as means and standard deviations, and categorical variables as counts and percentages.

Inverse probability weighting (IPW) was applied to minimize selection bias due to restricting the study sample to participants completing at least 1 reassessment (required ≥5 sessions). Inclusion probabilities were estimated via logistic regression including baseline demographic and clinical variables described in the “Confounders” subsection. Stabilized weights were derived (mean≈1.0; range = 0.7–2.9) and applied to subsequent analyses. Balance was evaluated by comparing the weighted study sample with the unweighted full starter population using standardized mean differences (SMDs) for continuous variables and absolute differences in category proportions for categorical variables, alongside visual balance plots. All covariates demonstrated excellent balance after weighting (absolute SMD <0.02 and absolute differences in category proportions <1 percentage points; Supplementary Figure 1).

Structural equation modeling (SEM) was used to examine associations between behavioral engagement indicators and last pain score. Two separate SEMs were modeled for early-phase and full treatment period predictors (see “Candidate Predictors” subsection) to avoid multicollinearity among temporally overlapping variables. Both models were adjusted for demographic, clinical, and psychological covariates (see “Confounders” subsection). Missing data were handled using full information maximum likelihood, allowing all participants data to enter the model.^52,53 Continuous variables entered in the model were standardized using z-scores, thereby beta coefficients correspond to the average change in the final pain score per each one-SD increase in the candidate predictor.

Model fit was evaluated using multiple indices, including the comparative fit index (CFI), root mean square error of approximation (RMSEA), and the standardized root mean square residual (SRMR). Model fit was considered acceptable based on conventional thresholds (CFI≥0.80, RMSEA ≤0.08, SRMR ≤0.08).⁵⁴ Explained variance (R²) for the pain outcome was also reported.

Moderation was tested with multigroup SEM to assess whether associations between behavioral engagement indicators and last pain score differed across baseline pain severity, race/ethnicity, BMI, and age categories. Baseline pain severity was categorized as mild (NPRS ≤3), moderate (NPRS 4–6), or severe (NPRS ≥7), consistent with previously established thresholds.⁵⁵ For each moderator, an unconstrained model allowing engagement–outcome paths to vary across groups was compared with a constrained model in which these paths were held equal across groups. Overall moderation was assessed using chi-square difference tests. When overall tests suggested potential group differences, pairwise comparisons of path coefficients were evaluated using Wald tests.

All analyses incorporated inverse probability weights and robust (sandwich) standard errors. Statistical significance was evaluated using two-sided tests with a P-value <0.05. Analyses were conducted in R (version 2024.09.0+375) using the lavaan package (version 0.6.19).

Results

Overall Cohort Characteristics

The study cohort included 19,866 participants with 48.4 years on average (SD 11.6) and a gender-wise balanced distribution (women: 55.3%, N=10,993). Mean BMI was 29.2 kg/m² (SD 6.7), with 72.1% (N=14,256) of participants classified as having overweight or obesity (Table 2). Most participants were non-Hispanic White (68.8%, N=12,778), had at least some college education (89.4%, N=17,127), lived in urban areas (86.2%, N=15,728), and were employed (91.5%, N=16,151). Social deprivation was broadly distributed across fifths, with 29.4% (N=5817) in the least deprived category and 9.9% (N=1960) in the most deprived.

Table 2.

Overall Cohort Characteristics (N=19,866)

Characteristics	Overall Cohort (N=19,866)
Demographic characteristics
Age (years), mean (SD)	48.4 (11.6)
Age categories (years), N (%):
<25	209 (1.1)
25–40	5328 (26.8)
41–60	10,875 (54.7)
>60	3454 (17.4)
Gender, N (%):
Woman	10,993 (55.3)
Man	8654 (43.6)
Non-binary or other	94 (0.5)
Prefer not to answer or not available	125 (0.6)
BMI (kg/m2), mean (SD)a	29.2 (6.7)
BMI categories (kg/m2), N (%)a:
Underweight (<18.5)	191 (1.0)
Normal (18.5–25)	5317 (26.9)
Overweight (≥25–30)	6962 (35.2)
Obesity (≥30–40)	5872 (29.7)
Severe Obesity (≥40)	1422 (7.2)
Race/ethnicity, N (%)b:
Asian	1762 (9.5)
Black and African American	1508 (8.2)
Hispanic or Latino	1974 (10.6)
Non-Hispanic white	12,778 (68.8)
Native	74 (0.4)
Multi-racial	340 (1.8)
Other	139 (0.7)
Education level, N (%)c:
Less than high school diploma	179 (0.9)
High school diploma	1843 (9.6)
Some college	5145 (26.9)
Bachelor’s degree	7405 (38.7)
Graduate degree	4577 (23.9)
Geographic location, N (%)d:
Urban	15,728 (86.2)
Rural	2520 (13.8)
Employment status, N (%)e:
Full-time job	15,063 (85.4)
Part-time job	1088 (6.2)
Retired	631 (3.6)
Not employed	865 (4.9)
Social deprivation index, N (%)f:
C1 (0–20)	5817 (29.4)
C2 (20–40)	4847 (24.5)
C3 (40–60)	4098 (20.7)
C4 (60–80)	3080 (15.6)
C5 (80–100)	1960 (9.9)
Clinical data
Symptomatic anatomical area, N (%):
Low back	14,943 (75.2)
Neck	4923 (24.8)
Pain Intensity, mean (SD)	4.8 (1.9)
Analgesic intake, N (%)g:	4648 (26.1)
Mental health, mean (SD):
FABQ-PAh	8.5 (6.1)
PHQ-9i	2.9 (4.7)
PHQ-9 ≥5j	9.5 (4.3)
GAD-7i	3.7 (4.7)
GAD-7 ≥5k	8.8 (3.9)
Work productivity, mean (SD)
WPAI overall >0l	31.1 (21.7)
WPAI work >0m	28.4 (18.6)
WPAI time >0n	22.3 (26.0)
WPAI activity >0o	34.9 (22.4)

Notes: ^aN=19,764; ^bN=18,575; ^cN=19,149; ^dN=18,248; ^eN=17,647; ^fN= 19,802; ^gN=17,838; ^hN=17,765; ⁱN=19,219; ^jN=5250; ^k7086; ^lN=10,210; ^mN=9844; ⁿ2461; ^o14,591.

Abbreviations: BMI, body mass index; GAD-7, Generalized Anxiety Disorder 7-item scale; PHQ-9, Patient Health 9-item Questionnaire; FABQ-PA, Fear-Avoidance Beliefs Questionnaire subscale for physical activity; WPAI, Work Productivity and Activity Impairment Questionnaire.

Clinically, participants experienced a mean baseline pain intensity of 4.8/10 (SD 1.9), with a majority suffering from low back pain (75.2%, N=14,943). Participants completed a mean of 21.0 sessions (SD 13.3) over a mean treatment duration of 8.4 weeks (SD 3.1). The last pain reassessment occurred after 5 sessions in N = 2245 subjects (mean treatment duration 5.93 weeks, SD 2.76), 9 sessions in N = 7187 subjects (mean treatment duration 6.85 weeks, SD 2.73), 15 sessions in N = 3504 subjects (8.71 weeks, SD 2.68), 27 sessions in N = 5727 subjects (10.46 weeks, SD 2.11), and 30 sessions reassessments in N = 1203 subjects (11.67 weeks, SD 0.85).

Associations Between Engagement Indicators and Pain Outcomes

Early-Phase Metrics

Initiation timing, early routine formation, and early continuity were all significantly associated with the last pain score across the first 3 weeks of treatment, although the magnitude of these associations was modest (Table 3). Longer time to complete the first session was associated with higher pain score at program completion (β=0.03, SE=0.01, P=0.014). Completing sessions on a recurring weekday within the first three weeks was a significant predictor of lower pain scores at program completion (β=−0.23, SE=0.04, P<0.001). Similarly, early continuity (defined as not missing any week in the first three treatment weeks) was linked to better post-pain scores (β=−0.21, SE=0.04, P<0.001). Performing sessions in a routine weekday early on is associated with less 0.23-points, on average, on final pain scores, conditional on the remaining variables in the model. Despite statistical significance, effect sizes for early-phase behavioral engagement indicators had a very small incremental effect size (Cohen’s f²=0.01). This model demonstrated acceptable fit and explained 24.4% of the pain outcomes variance.

Table 3.

Predictive Model of Last Pain Score

Model	Variable	N	Coefficient	SE	P-value	Df	R2	BIC	CFI	RMSEA	SRMR
Early-phase indicators	Initiation timing (z-scores)	19,866	0.03	0.01	0.014	125	0.244	715,149.6	0.86	0.047	0.037
	Early routine formation (=yes)		−0.23	0.04	<0.001
	Early continuity (=yes)		−0.21	0.04	<0.001
Full-period indicators	Weekday routine (z-scores)	19,866	−0.21	0.02	<0.001	88	0.249	699,128.8	0.83	0.053	0.038
	One-week interruption (=yes)		0.14	0.04	<0.001

Notes: The reference category for each categorical variable was “no”, specifically for early routine formation, early continuity, and one-week interruption. Significant p-values are denoted in bold.

Abbreviations: BIC, Bayesian information criterion; CFI, Comparative fit index; Df, degrees of freedom; RMSEA, Root Mean Square Error of Approximation index; SE, standard error; SRMR, Standardized Root Mean Square Residual index.

Full-Period Metrics

Across the full intervention period, greater consistency in weekday routine was significantly associated with lower pain scores (β=−0.21, SE=0.02, P<0.001). Conversely, experiencing a one-week interruption (ie, seven consecutive days without sessions) was associated with higher last pain scores (β=0.14, SE=0.04, P<0.001). Incremental effect sizes for full-period behavioral engagement indicators were small (Cohen’s f²=0.02). The model showed acceptable fit and accounted for 24.9% of the variance in last pain score.

Moderation Analysis

Evaluations of associations between behavioral engagement indicators and last pain score differed across baseline pain, race/ethnicity, BMI, and age categories are reported at Supplementary Tables 2–5.

BMI category moderated associations between full-period behavioral engagement indicators and pain (Δχ²(2) = 10.09, P=0.006). The impact of one-week interruptions on pain outcomes was stronger among patients with obesity compared to those without obesity (β=0.27, SE=0.06, P<0.001 and β=0.06, SE=0.04, P=0.131, respectively; Supplementary Table 4).

No significant moderation effects were observed for baseline pain (early-phase indicators: P=0.518; full-period indicators: P=0.140), race/ethnicity (early-phase indicators: P=0.248; full-period indicators: P=0.997), or age (early-phase indicators: P=0.924; full-period indicators: P=0.168), indicating that the associations between behavioral engagement indicators and pain outcomes were consistent across these subgroups (Table 4).

Table 4.

Chi-Square Difference Tests on Multiple-Group Models to Evaluate Moderation Effects

Moderator	Early-Phase Model			Full-Period Model
	ΔDf	Δχ2	P-value	ΔDf	Δχ2	P-value
Pain baseline	6	5.20	0.518	4	6.92	0.140
Race/ethnicity	3	4.13	0.248	2	0.01	0.997
BMI	3	3.77	0.287	2	10.09	0.006
Age	6	1.96	0.924	4	6.44	0.168

Notes: Δχ² indicates the chi-square difference between constrained and unconstrained models. Significant p-values are denoted in bold.

Discussion

Main Findings

This large, real-world study underscored behavioral engagement indicators (both at early-phase and full-intervention period) as significant predictors of pain outcomes following AI-enabled MSK rehabilitation. Faster initiation, early weekday routine, and early continuity (no missed weeks) were each associated with lower pain. Across the full-period, maintaining a weekday routine and avoiding a one-week interruption (≥7 consecutive days without sessions) were associated with better final pain scores. Models explained roughly one-quarter of pain outcomes variance (R²≈24–25%), indicating meaningful, though not exhaustive, explanatory value.

The association between a one-week interruption and pain outcomes was stronger among participants with obesity. Relationships between engagement indicators and pain outcomes were consistent across baseline pain severity, race/ethnicity, and age.

Together, these findings support that behavioral engagement indicators, particularly initiation, routine formation and continuity, may provide complementary predictive information of pain outcomes after AI-enabled rehabilitation. Future studies are needed to determine how these signals can guide care personalization, timely outreach, or escalation for patients at greater risk of poorer pain outcomes.

Comparison with Existing Literature

Most prior studies assessing engagement in MSK rehabilitation, both in-person^8–12 and remote,^13–15 have focused on exercise dosage metrics, namely total number of sessions, duration, exercise volume, or adherence rates. Moving beyond adherence volume, our goal was to characterize how participants engaged with the remote rehabilitation intervention and whether engagement behavior variations were prognostically informative regarding last pain score.

Initiation timing was a statistically significant predictor of final pain, with earlier initiation associated with better outcomes. This direction is consistent with in-person MSK literature linking delays in initiating physical therapy to poorer outcomes and higher healthcare utilization and costs.^38–40 However, direct comparisons warrant careful interpretation. While initiation delays in traditional care often reflect structural barriers (eg, referrals, scheduling, transportation, and cost),⁵⁶ in remote settings initiation timing likely reflects broader factors, including perceived value and personal relevance of care, motivation and confidence to engage, and early decisional engagement.⁵⁷ Past literature on remote physical therapy has shown similar clinical outcomes across rural and urban areas,¹³ and socioeconomic strata,²⁸ suggesting that remote delivery can reduce common logistical barriers. Within this context, initiation timing may retain predictive value as an early marker of readiness and momentum rather than access alone.

A central finding was the association between weekday routine formation, both early in treatment and across the intervention period, and lower final pain score. This aligns with behavioral science evidence that early consistency and stable routines promote habit formation, behavioral automaticity, and sustained engagement in physical activity.^16,17 Related work in post-stroke neurorehabilitation has shown that early scheduling consistency predicts longer-term perseverance and greater exercise exposure.³⁶ Our results extend these insights to remote MSK rehabilitation, indicating that a predictable session cadence may represent a meaningful behavioral signature of engagement linked to better pain outcomes. Potential explanations include reduced cognitive effort to initiate sessions, greater resilience to motivation fluctuations, and easier integration of rehabilitation activities within daily life,^35,58 which may be particularly relevant in self-directed, remote care. Noteworthy, these mechanisms were not directly evaluated in the present study and should be investigated in future research.

One-week interruptions (both early on and across the full intervention) were associated with higher final pain scores, consistent with evidence that interruptions can disrupt emerging routines, undermine habit formation, and increase the effort required to re-engage.³⁷ Within rehabilitation contexts, interruptions may also reflect difficulties sustaining regular exposure and achieving the prescribed therapy dosage necessary for benefit. However, reverse directionality is also plausible: higher pain may hinder participation and contribute to more frequent interruptions. As temporal or causal sequences cannot be established, these indicators should be interpreted as predictors rather than causal mechanisms. Future work should explore exercise dosage (eg, total sessions or volume) or other factors (eg, self-efficacy) as mediators linking engagement patterns to pain, to better characterize digital engagement phenotypes and advance predictive knowledge.

Although statistically significant, incremental effect sizes for behavioral engagement were small (Cohen’s f² = 0.01–0.02). As engagement indicators were evaluated to determine their distinct association with pain outcomes after accounting for a wide set of confounders, these estimates reflect their additional predictive contribution beyond that provided by demographic, clinical, and psychosocial domains. Given the multifactorial nature of chronic pain prognosis, individual predictors may have limited stand-alone effects while still providing complementary prognostic information when considered alongside other prognostic domains. The magnitude of the individual coefficients should therefore not be interpreted as evidence of clinically meaningful pain improvement attributable to engagement indicators in isolation. Their practical value may lie less in large stand-alone effects and more in providing behaviorally interpretable and potentially modifiable signals to complement traditional predictors, although further confirmatory research is needed to determine their clinical utility. Overall explained variance (R² ≈ 24–25%) was comparable to prior predictive models in remote MSK care^27,59 and in-person rehabilitation studies.^{11,12,60–63}

BMI moderated the association between full-week interruptions and pain outcomes, with stronger adverse associations among participants with obesity. This aligns with evidence that individuals with obesity may experience additional barriers to sustained engagement in exercise-based interventions,^64,65 including higher symptom burden, fatigue, altered physiological capacity affecting exercise adaptation, mobility limitations, and psychosocial factors. These may make re-engagement after interruptions more difficult and, in turn, adversely affect recovery. In contrast, the lack of moderation by baseline pain, race/ethnicity, and age suggests that herein engagement–outcome associations were broadly consistent.

The AI-enabled care model included several features that may have supported patient engagement and therefore merit consideration. Automated motion tracking and real-time audiovisual biofeedback may facilitate patient autonomy during exercise sessions by providing immediate guidance on repetitions and movement performance. Additionally, the AI-assisted tool, which supported physical therapists in patient prioritization and monitoring, may have helped streamline clinical review and treatment adjustment. These mechanisms could plausibly support initiation, routine formation, and continuity. However, the engagement indicators analyzed herein were derived from device logs and were not prospectively used by the AI system to personalize care. Moreover, as all participants received the AI-enabled intervention, the findings reflect predictive associations observed within a fully-remote, AI-enabled rehabilitation intervention, and not effects attributable to AI itself.

These results provide initial evidence that objective engagement behaviors – early initiation, routine consistency, and continuity – are measurable and associated with pain outcomes in remote MSK rehabilitation. These indicators may represent candidate signals for identifying patients at risk of poorer outcomes and could inform proactive care tailoring. Future pragmatic trials or adaptive intervention designs should test whether modifying engagement behaviors improves outcomes through targeted strategies, such as by strengthening onboarding support for delayed initiation, planning prompts for weak routine formation, or outreach and problem-solving when interruptions emerge.

Strengths and Limitations

This study leveraged a large, nationwide real-world cohort with diverse representation (eg, race/ethnicity, education) aligned with the US demographic profile,⁶⁶ increasing generalizability. Objective engagement measurement from device logs reduces recall and social desirability bias. Assessing early- and full-period metrics allowed distinguished engagement patterns emerging early versus over time. Application of SEM adjusted for a comprehensive set of covariates, alongside IPW, increased findings robustness.

Limitations include the observational design (precluding causal inference), potential residual confounding from unmeasured variables (eg, additional contextual factors, pain characteristics and phenotypes, and care interactions), and heterogeneity in discharge timing and the last completed pain reassessment milestone (despite adjustment for treatment duration and IPW correction). Engagement indicators were necessarily incomplete; additional features (eg, time-of-day routine, contextual barriers, symptom-contingent use) may also be relevant. These findings should be interpreted as predictive associations rather than causal mechanisms of pain improvement, and confirmatory studies are needed for external validation and test causality. Finally, other core outcomes, such as disability and quality of life, were not assessed and could further expand knowledge in this field.

Conclusions

In this large, real-world cohort of adults with chronic spinal pain undergoing AI-enabled, remote MSK rehabilitation, behavioral engagement indicators were significantly associated with pain outcomes, albeit at a small effect magnitude. Earlier initiation, early routine formation, and early continuity were associated with lower last pain. Weekday routine consistency and avoiding one-week interruptions across the full program predicted better pain. Associations were consistent across baseline pain, race/ethnicity, and age, while the adverse effect of interruptions was stronger among participants with obesity.

Behavioral engagement patterns may represent potentially modifiable targets for intervention in remote care and may help identify patients who could benefit from targeted support. However, confirmatory studies are needed to further validate these findings and determine how they may inform personalized care.

Funding Statement

This work was supported by Sword Health Inc. and also developed within the scope of project n° 62—“Responsible AI”, financed by European Funds, namely the Recovery and Resilience Plan—“Componente 5: Agendas Mobilizadoras para a Inovação Empresarial”, included in the NextGenerationEU funding program.

Declaration of Generative AI and AI-Assisted Technologies in the Writing Process

During the preparation of this work the author(s) used OpenAI ChatGPT (GPT-4o) in order to support copyediting and grammar checking. The tool was not used to generate study data, analyses, results, interpretations, or references. All AI-assisted content was reviewed and edited by the authors. The authors confirm the originality and accuracy of all manuscript content and take full responsibility for the integrity of the entire work, including the accuracy and verification of all references. The authors also confirm that the terms of use applicable to GPT-4o were checked and consider its use suitable for publication.

Abbreviations

AI, Artificial intelligence; BMI, Body Mass Index; CBT, Cognitive Behavioral Therapy; CFI, Comparative Fit Index; FABQ-PA, Fear-Avoidance Beliefs Questionnaire subscale for physical activity; GAD-7, Generalized Anxiety Disorder 7-item scale; IPW, Inverse Probability Weighting; MSK, Musculoskeletal; NPRS, Numerical Pain Rating Scale; PHQ-9, Patient Health 9-item Questionnaire; PROGRESS, PROGnosis RESearch Strategy; RMSEA, Root Mean Square Error of Approximation; SDI, Social Deprivation Index; SEM, Structural Equation Modeling; SMD, Standardized Mean Difference; SRMR, Standardized Root Mean Square Residual; TRIPOD, Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis; US, United States.

Data Sharing Statement

Data in this study can be made available on request from the corresponding authors Fabíola Costa or Dora Janela. Data are not publicly available due to privacy restrictions.

Ethics Approval and Informed Consent

This study was approved by Advarra Institutional Review Board (Pro00063337). The trial was registered at ClinicalTrials.gov (NCT05417685) prior to patient enrollment and was conducted in accordance with the Declaration of Helsinki. Informed consent was obtained from all participants before participation in the study.

Author Contributions

Conceptualization: D.J., D.P., F.C., and H.F.; Data curation: D.J.; Formal analysis: D.J. and T.X.; Interpretation of data: all authors; Writing—original draft preparation: D.J.; Writing—review and editing: all authors. All authors made a significant contribution to the work reported, whether that is in the conception, study design, execution, acquisition of data, analysis and interpretation, or in all these areas; took part in drafting, revising, or critically reviewing the article; gave final approval of the version to be published; have agreed on the journal to which the article has been submitted; and agree to be accountable for all aspects of the work.

Disclosure

D. Janela and F. Costa are employees at Sword Health. X. Tong receives scientific advisor honorarium from Sword Health.