Predicting pain outcomes after digital care in chronic spinal pain: the roles of disability, work impairment, and occupation in a secondary analysis of a prospective clinical study

Abstract

Objective

To evaluate whether baseline disability, work impairment, and job occupation predicted post-pain levels of a digital care program (DCP) for chronic spinal pain.

Design

Ad hoc analysis of a real-world clinical registry of patients undergoing a DCP.

Setting

DCP delivered remotely across the United States.

Subjects

Adults with chronic spinal pain (N = 13 330) enrolled in a DCP through employer-sponsored health plans.

Methods

Predictors included baseline disability (Oswestry Disability Index or Neck Disability Index), work impairment (Work Productivity and Activity Impairment questionnaire—WPAI), and occupation (job type group). Primary outcome was the last pain score reported during the intervention (11-point Numeric Pain Rating Scale). Structural equation modeling was used, adjusted for demographic and clinical covariates. Moderation analysis assessed whether effects varied by pain location (neck vs low back).

Results

Baseline disability and occupation significantly predicted post-treatment pain. Greater disability was associated with higher last pain scores (β = 0.30, SE 0.02, P<.001). Business-related occupations were non-significantly different from trade, transportation, and utilities, but showed higher last pain score than those in goods-producing (β=−0.18, SE 0.07, P=.015) and healthcare/education (β=−0.14, SE 0.04, P=.001) jobs. WPAI Overall and WPAI Activity were not significant predictors after adjustment. Predictor effects were consistent across spinal locations. Final model explained 21.3% of variance.

Conclusions

Baseline disability and occupation were predictors of outcomes post-digital rehabilitation for chronic spinal pain, while work impairment was non-significant. Integrating these factors into routine screening may enhance predictive accuracy, patient communication, and facilitate personalized care pathways. These results encourage confirmatory studies to reinforce these findings.

Trial registration

ClinicalTrials.gov, NCT05417685. Registered on June 14, 2022; https://clinicaltrials.gov/study/NCT05417685.

Using structural equation modeling in a large real-world cohort (N=13,330), this study identified baseline disability and occupation as significant predictors of last reported pain score of a fully-remote, digital program for chronic spinal pain. Work impairment did not show independent predictive value, and findings were consistent across spinal regions. These results open new research avenues on the integration of these factors into routine screening to support recovery expectations and facilitate personalized care pathways.

Introduction

Chronic spinal pain, the leading driver of disability and work productivity loss, affects over 835 million people worldwide, with nearly 90 million years lived with disability.¹ In the United States (US) alone, healthcare expenditures for spinal pain exceed $245 billion annually,² with indirect costs related to productivity loss estimated to be even higher.³

First-line care for chronic spinal pain involves multimodal rehabilitation combining exercise, education, and behavioral change.^4–6 However, responses to treatment responses vary widely among patients with the same diagnosis,⁷ underscoring the need for research on predictors of outcomes that can support prognosis establishment and lead to more personalized treatment pathways.⁸ Wider availability of care delivery options may further contribute to treatment response variability. Digital interventions have emerged as an accessible, scalable solution to expand the reach of evidence-based care⁹ with comparable effectiveness to in-person treatment across several musculoskeletal conditions,^10,11 including spinal pain.^12,13 Its convenience addresses known barriers in terms of time, location, and provider availability¹⁴ fostering adherence,^10,11 while enabling data-driven care personalization at scale.

However, there are marked differences between remote and in-person settings, namely flexible scheduling, patient autonomy, asynchronous completion of sessions, and remote interaction with clinicians, that can modify adherence patterns, behavioral change, and self-management. These differences imply that predictive knowledge from traditional care may not be directly transferable to digital care, where predictive research is still in its infancy.¹⁵ Therefore, we have recently focused our research efforts on identifying predictors of pain outcomes on remote MSK care. First studies have identified psychological factors as predictors and mediators of pain outcomes.¹⁶ Building on that foundational work, we now aim to explore other core domains relevant to recovery.¹⁷

Physical functioning and work ability are among the most affected aspects in spinal pain.¹ Baseline disability levels have been consistently associated with poorer outcomes following in-person physical therapy.^18–21 Work-related factors (eg, workplace psychosocial stress, job strain) are also extensively studied as contributors to the onset and chronicity of musculoskeletal conditions,^22,23 but their predictive value for treatment response remains underexplored. Similarly, while specific job demands are recognized risk factors, occupation type by itself has yet to be evaluated as a potential predictor of recovery success.

Effective predictors should be pragmatic, clinically relevant, and feasible to measure without placing undue burden on patients or clinicians.⁸ Unlike complex or costly biomarkers, disability, and work-related factors (such as work impairment or occupation type) could be seamlessly integrated into clinical workflows to inform triage, manage expectations, and tailor treatment in a sustainable way. Addressing this evidence gap would represent an important foundational step and yield insights with clinical utility, particularly in digital settings, where data-driven care personalization may support outcomes optimization.

The present study aimed to assess whether disability, work impairment, and occupation predict the post-pain outcomes of a digital care program (DCP) among patients with chronic spinal pain. This DCP safety, feasibility, and effectiveness compared to in-person physical therapy have been previously demonstrated through a randomized controlled trial in chronic low back pain.¹² We hypothesized that the aforementioned factors would significantly predict pain outcomes following a DCP.

Methods

Study design

Ad hoc analysis of an ongoing real-world, prospective clinical registry study of patients with musculoskeletal pain treated with a DCP, under a study approved by Advarra Institutional Review Board (Pro00063337) and registered on ClinicalTrials.gov (NCT05417685) on June 14, 2022. This study was conducted in accordance with the Declaration of Helsinki, and the TRIPOD (Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis) reporting guidelines. Consistent with the staged continuum for developing, validating, and implementing prognostic research described by PROGRESS,²⁴ this study corresponds to an initial stage focused on examining associations between candidate baseline predictors and pain outcomes.

Study population

Adults (≥18 years) from any US state, beneficiaries of employer health plans, and reporting chronic spinal musculoskeletal pain (ie, persistent or recurrent neck or low back pain with ≥3 months duration)²⁵ were eligible to participate in the study. All participants provided electronic informed consent.

Exclusion criteria comprised: (1) health conditions (eg, cardiac, respiratory) or other contra-indications precluding performing ≥20 min of light-to-moderate intensity exercise; (2) actively receiving cancer treatment; (3) uncleared clinical red flags (eg, rapidly progressive loss of strength and/or numbness in the arms/legs or unexplained change in bowel or urinary function in the previous 2 weeks); (4) inability to follow simple and complex motor commands, or to perform independently home-based exercise; and (5) patients who did not complete baseline assessment or any reassessment throughout the study.

Intervention

This intervention has been described in detail elsewhere.²⁶ Briefly, this multimodal DCP included exercise, education, and cognitive behavioral therapy (CBT), according to clinical guidelines.^4–6 Each participant was assigned to a physical therapist, who oversaw their progress throughout the program. During onboarding, the assigned physical therapist performed a remote clinical assessment to confirm the absence of potential red flags and gather relevant clinical information, after which a personalized intervention was prescribed. The exercise plan was performed independently by the patient at their own convenience, using an FDA-listed class II medical device composed of a dedicated tablet with a mobile app, a motion tracking system and a cloud-based portal (Digital Therapist, United States). Patients received visual and audio real-time biofeedback during the exercise sessions displayed in the tablet. Data from these sessions was stored in the portal, being accessed by the physical therapist to asynchronously monitor and adjust the intervention accordingly.

The education and CBT components were designed according to current clinical guidelines^4–6 and consisted of educational written materials, audio content, and interactive audio-guided CBT modules, made available to patients through an accompanying smartphone app (Sword Health). Main topics included MSK pain pathophysiology, active coping skills, and fear-avoidance behaviors. The CBT program was based on principles of mindfulness, acceptance and commitment therapy and empathy-focused therapy.²⁷

Communication between patients and physical therapists was ensured throughout the intervention, and occurred either through the chat within the smartphone app, videocall, or phonecall, according to patients’ preferences.

Data collection

Electronic, self-reported surveys were completed at baseline (including both clinical and sociodemographic characteristics), and at specific sessions: 9-, 15-, 27-, and 30-sessions (between June 19, 2020, and November 4, 2024). The session-based reassessment schedule standardizes treatment dosage and reduces bias from uneven session frequency and variable discharge timing—factors that are less reliably captured with time-based assessments—and it aligns with the real-world clinical nature of the program while supporting the study’s predictive focus. Although discharge typically occurred within 12 weeks, the exact timing was dependent on each patient’s clinical evolution as determined by the assigned physical therapist, and patients could discontinue participation before discharge.

Outcome

The primary outcome was the last pain score reported during the DCP (ie, the last reassessment completed by each participant), measured through the 11-point Numeric Pain Rating Scale [NPRS; scoring 0- (no pain) to 10- (worst imaginable pain)] within a 7-day recall period.²⁸ The last observed pain score used corresponded to 9-sessions in N = 5601 subjects (mean treatment duration 7.07 weeks, SD 2.79), 15-sessions in N = 2720 subjects (8.85 weeks, SD 2.65), 27-sessions in N = 4070 subjects (10.50 weeks, SD 2.05), and 30-sessions in N = 939 subjects (11.70 weeks, SD 0.85). Pain intensity is a core outcome domain in chronic pain trials, being reported as one of the most important attributes reported by patients with chronic pain²⁹ and a central contributor to limitations in other domains (eg, disability, work impairment). NPRS is among the most widely recommended measures for this purpose.³⁰

Predictors and clinically relevant variables

Candidates for predictors of pain were chosen according to their clinical relevance, theoretical plausibility, and feasibility for use in real-world scenarios, comprising disability, work impairment, and occupation (Table 1 and Table S1). These variables reflect core clinical domains commonly affected by chronic spinal pain,¹ being primary targets of rehabilitation.¹⁷ The potential predictive value of disability and work factors is supported by available evidence from in-person rehabilitation contexts.^18,31–33 Additionally, these factors meet criteria for clinical utility, as they are easy to collect and may offer important predictive insights that can support early triage and guide care planning.

Table 1.

Description of predictor variables and respective measures included in the predictive model.

Variable	Definition and measures
Disability	• Oswestry Disability Index (ODI) for patients with low back pain35 • Neck Disability Index (NDI) for patients with neck pain37 Each questionnaire contains 10 items rated on a 6-point Likert scale, with total scores ranging from 0% (no disability) to 100% (maximum disability).
Work impairment	Work Productivity and Activity Impairment questionnaire (WPAI) General Health v2.0,38 considering a 7-day recall period. Two subscales were analyzed: • WPAI overall (combining presenteeism and absenteeism) • WPAI Activity (capturing non work-related activities impairment). Each subscale ranges from 0% (no impairment) to 100% (complete impairment).
Occupation	Classified according to the Standard Occupational Classification (SOC) system from the US Census Bureau—Occupational Employment and Wage Statistics (OEWS) program, as previously reported.54 SOC codes were combined into broader occupational groups for analysis: Goods-producing and manual labor (GP); trade, transportation, and utilities (TTU); healthcare, education, and social services (HE); business, financial, information, and related services (BFIS). The specific SOC codes comprising each occupational group are listed in Table S1.

Following best practices for health-related predictive modeling,³⁴ models were adjusted for variables previously identified as clinically relevant covariates^16,31,33,36 to mitigate confounding bias and improve model accuracy. Covariates included: Age, gender (binary: Men; non-men-women and non-binary), body mass index (BMI), race/ethnicity (Asian, Black or African American, Hispanic or Latino, non-Hispanic white, and Other), education levels (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), and baseline pain, fear-avoidance beliefs (Fear-Avoidance Beliefs Questionnaire subscale for physical activity—FABQ-PA), depression (Patient Health 9-item Questionnaire—PHQ-9), and anxiety (Generalized Anxiety Disorder 7-item scale—GAD-7).

Sample size

Following the recommended guideline of at least 20 participants per estimated parameter to reduce bias in calculated coefficients,³⁹ the planned structure equation model (SEM) with 246 parameters (6 predictors, 19 covariates, 25 variances, 1 residual variance, 26 intercepts, 169 covariances) required approximately 7540 participants. With over 10 000 participants enrolled, the sample size substantially exceeds this minimum requirement, ensuring model stability and sufficient power for valid predictive analyses.

Statistical analysis

Cohort baseline characteristics were summarized through descriptive statistics. Initial data exploration included examining scatterplots of candidate predictors against last pain score. Model assumptions (including linearity, homoscedasticity, independence, and residuals normality) were verified. Multicollinearity was evaluated through Pearson’s correlations between variables and Variance Inflation Factor (VIF) values (Supplementary Material 1). Given the high collinearity between WPAI Work and WPAI Overall and the limited variability observed on WPAI Time (86.1% reported zero absenteeism), the final analysis focused solely on WPAI Overall (which combines both work and time subscales) and WPAI Activity. The final analysis VIF values (1.02-3.53) indicated no significant collinearity. Outliers and influential cases were identified through residual plots and Cook’s distance. To assess potential attrition bias, baseline characteristics of participants included in the analysis were compared with those excluded for not completing any reassessment (Table S2).

Structural equation models (SEM) were built to investigate the predictive value of identified predictors on the last pain score. SEM was selected for its flexibility in modeling latent variables and complex relationships (including covariances), robustness to non-normal data, and ability to handle missingness using full information maximum likelihood (FIML).^40,41 To avoid overspecification and ensure parsimony, covariances were estimated based on the correlation matrix (Supplementary Material 1) and theoretical rationale supporting relationships between variables. Several candidate structures were evaluated to achieve an optimal balance between parsimony and adequate model fit. Both unadjusted and adjusted models were developed, with the latter including covariates listed in the “Predictors and confounding variables” section. Missingness ranged from 0% to 5.2% across variables (complete observations for each variable is reported in footnote of Table 2).

Table 2.

Baseline characteristics of the study cohort (N = 13 330).

Characteristics	Study cohort (N = 13 330)
Demographic characteristics
Age (years), mean (SD)	48.5 (11.1)
Age categories (years), N (%):
<25	80 (0.6)
25-40	3462 (26.0)
41-60	7624 (57.2)
> 60	2164 (16.2)
Gender, N (%):
Woman	7136 (53.5)
Man	6044 (45.3)
Non-binary or other	64 (0.5)
Prefer not to answer or not available	86 (0.6)
BMI (kg m−2), mean (SD)a	29.1 (6.7)
BMI categories (kg m−2), N (%)a:
Underweight (<18.5)	128 (1.0)
Normal (18.5-25)	3603 (27.1)
Overweight (≥25-30)	4766 (35.9)
Obesity (≥30-40)	3871 (29.2)
Severe obesity (≥40)	908 (6.8)
Race/ethnicity, N (%)b:
Asian	1258 (10.0)
Black and African American	999 (7.9)
Hispanic or Latino	1296 (10.3)
Non-Hispanic white	8698 (68.8)
Native American	53 (0.4)
Multi-racial	241 (1.9)
Other	89 (0.7)
Education level, N (%)c:
Less than high school diploma	74 (0.6)
High school diploma	1120 (8.7)
Some college	3246 (25.2)
Bachelor’s degree	5158 (40.0)
Graduate degree	3282 (25.5)
Employment status, N (%)d:
Full-time job	12 410 (93.2)
Part-time job	899 (6.8)
Occupation, N (%):
Goods-producing and manual labor	741 (5.6)
Trade, transportation, and utilities	1933 (14.5)
Healthcare, education, and social services	3030 (22.7)
Business, financial, information, and related services	7626 (57.2)
Social deprivation index, N (%):
C1 (0–20)	3976 (29.9)
C2 (20–40)	3291 (24.8)
C3 (40–60)	2679 (20.2)
C4 (60–80)	2004 (15.1)
C5 (80–100)	1340 (10.1)
Geographic location, N (%)e:
Urban	10 601 (86.9)
Rural	1593 (13.1)
Clinical data
Pain anatomical area, N (%):
Low back	10 075 (75.6)
Neck	3255 (24.4)
ODI, mean (SD)	20.7 (12.0)
NDI, mean (SD)	24.4 (11.6)
Pain Intensity, mean (SD)	4.7 (1.9)
Analgesic intake, N (%)f:	3385 (25.4)
Mental health, mean (SD):
FABQ-PAg	8.4 (6.0)
PHQ-9h	2.7 (4.6)
GAD-7h	3.6 (4.5)
Work productivity, mean (SD)
WPAI overall	18.3 (22.6)
WPAI work	16.2 (19.8)
WPAI time	3.2 (12.6)
WPAI activity	25.1 (24.2)

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; NDI, Neck Disability Index; ODI, Oswestry Disability Index; WPAI, Work Productivity and Activity Impairment Questionnaire.

^a N = 13 276;

^b N = 12 634;

^c N = 12 880;

^d N = 13 309;

^e N = 12 194;

^f N = 13 326;

^g N = 13 261;

^h N = 13 264.

ODI and NDI scores were combined into a single disability variable for analysis. Both instruments assess the shared underlying construct of spinal pain-related disability and include several overlapping functional domains (eg, personal care, lifting, sleep disturbance). Prior research has proposed unified disability indices applicable across spinal regions,^42,43 supporting its shared measure construct. A sensitivity analysis was run to confirm the appropriateness of merging ODI and NDI, comparing the model performance of 2 multiple-group SEMs (grouped by pain location) with the disability coefficient constrained (ie, precluding variation) versus unconstrained (ie, allowing variation). The chi-square difference test was non-significant (χ(1)=2.08, P=.149), supporting their combination into a single variable, which enhances model parsimony and preserves statistical power.

Candidate predictors and respective confounders (continuous variables) were standardized using z-scores (ie, subtracting the mean and dividing by the standard deviation—SD). For reference, 1 SD corresponded to: ODI/NDI = 12.0, WPAI Activity = 24.2, WPAI Overall = 22.6, age = 11.1, BMI = 6.6, baseline pain = 1.9, baseline FABQ-PA = 6.1, baseline PHQ-9 = 4.6, and baseline GAD-7 = 4.5. For occupation (categorical), the BFIS group was selected as the reference category due to its distinct nature, being predominantly sedentary and involving prolonged sitting, contrasting with greater physical demands or movement variability in the other groups.

Model coefficients (β) represent the average change in the final pain score per each one-SD increase in a given predictor. Goodness of fit was evaluated through comparative fit index (CFI; values close to 1 indicate good fit), root mean square error of approximation index (RMSEA; values <0.08 indicate good fit), and standardized root mean square residual index (SRMR; values <0.05 indicate good fit).⁴⁴ Model’s predictive power was calculated through R², depicting the proportion of variance in post-intervention pain explained by the studied predictors. To verify the contribution of each predictor to model fit, the full SEM (ie, including all variables) was compared with nested versions where each predictor’s coefficient was constrained to zero through chi-square difference tests. An increase in chi-square value above the critical value (eg, Δχ²> 3.84, considering P < .05, df = 1) indicates that constraining the variable worsens model fit, supporting a meaningful contribution.

A sensitivity analysis was conducted to account for variability in intervention exposure, incorporating the total number of completed sessions and the treatment duration (in weeks) as additional covariates to the fully adjusted SEM.

Finally, a moderation analysis explored whether associations between baseline disability, work impairment, and occupation with final pain score differed by pain location. Moderation was first tested through interactions added to SEM. Since adding interaction terms resulted in a poor model fit, we allowed for covariances being freely estimated among predictors. To provide a complementary and more parsimonious assessment of moderation, a multiple-group SEM using the same model structure as the main analysis was also conducted, fitted separately for neck and low back pain.

All statistical analyses were performed in R Studio (version 2024.09.0 + 375), using the R lavaan package (version 0.6.19) with the maximum likelihood (ML) estimator. Statistical significance was set at P < .05 for all tests considering a 2-sided hypothesis test.

Results

Baseline characteristics

The study cohort (N = 13 330) was predominantly middle-aged (mean 48.5 years, SD 11.1), gender-balanced (woman: 53.5%, N = 7136), with 68.8% (N = 8698) identifying as non-Hispanic White, and 71.9% (N = 9545) presenting overweight or obesity (mean BMI: 29.1 kg m⁻², SD 6.7) (Table 2). Most participants had higher education (65.5%), resided in urban areas (86.9%), worked in BFIS occupations (57.2%, N = 7626), and presented low social deprivation (C1 and C2: 54.7%, N = 7267). Three-quarters of patients suffered from low back pain (75.6%, N = 10 075), and one-quarter neck pain (24.4%, N = 3255).

Predictors of pain outcomes

Correlation analyses between variables showed that ODI/NDI had moderate associations with WPAI Overall (r(13 328)=0.498, 95% CI 0.486; 0.510; P<.001) and WPAI Activity (r(13 328)=0.577, 95% CI 0.565; 0.588; P<.001), but all VIF values remained below 3.53, indicating no multicollinearity (Supplementary Material 1). Without adjustment for potential confounders, baseline disability, and impairment in work and non-work-related activities were significant predictors of last pain score, with worse baseline scores being associated with higher post-intervention pain (β = 0.66, SE 0.02, P<.001, β = 0.05, SE 0.02, P=.041, and β = 0.19, SE 0.02, P<.001, respectively; Table S3). Occupation was also associated with post-intervention pain, with individuals from the business, financial, information, and related services (BFIS) group reporting higher pain scores compared to those in the healthcare and education (HE) group (β=−0.13, SE 0.04, P=.001). The unadjusted model reached an explained variance of 15.3%.

After adjusting for demographic and clinical variables, baseline disability (β = 0.30, SE = 0.02, P<.001; Table 3) and occupation remained significant predictors of pain outcomes, contrary to work and non-work-related activities. Greater disability was associated with higher last pain scores. For every 12-point increase in ODI or NDI, post-intervention pain increased by 0.30 points. Individuals in the BFIS occupational group had similar behavior as trade, transportation, and utilities (TTU) group, but were associated with higher last pain levels compared to HE (β=−0.14, SE = 0.04, P=.001) and to goods-producing (GP; β=−0.18, SE = 0.07, P=.015) groups.

Table 3.

Predictive model of last pain score: adjusted analysis.

Variable	N	Coefficient	SE	95% CI	P-value	R 2	BIC	CFI	RMSEA	SRMR
ODI/NDI (z-scores)	13 330	0.30	0.02	0.26; 0.35	<.001	0.213	491945.5	0.83	0.063	0.058
WPAI overall (z-scores)		0.00	0.02	−0.04; 0.05	.903
WPAI activity (z-scores)		0.05	0.02	−0.04; 0.05	.831
Occupation (=GP)		−0.18	0.07	−0.32; −0.03	.015
Occupation (=TTU)		−0.03	0.05	−0.12; 0.06	.529
Occupation (=HE)		−0.14	0.04	−0.22; −0.06	.001
Age (z-scores)		−0.01	0.02	−0.05; 0.02	.391
Gender (=man)		−0.11	0.03	−0.17; −0.04	.002
BMI (z-scores)		0.07	0.02	0.03; 0.10	<.001
Race (=Asian)		0.02	0.06	−0.09; 0.13	.726
Race (=Black)		−0.08	0.06	−0.21; 0.04	.190
Race (=Hispanic)		−0.06	0.06	−0.17; 0.05	.286
Race (=Other)		0.02	0.10	−0.17; 0.20	.873
Education (=Graduate)		−0.09	0.07	−0.22; 0.04	.157
Education (=Bachelor’s)		−0.08	0.06	−0.20; 0.04	.183
Education (=Some college)		−0.07	0.06	−0.19; 0.06	.298
SDI (=C2)		0.06	0.04	−0.02; 0.15	.163
SDI (=C3)		0.03	0.05	−0.06; 0.12	.517
SDI (=C4)		0.14	0.05	0.04; 0.24	.006
SDI (=C5)		0.11	0.06	0.00; 0.23	.057
Analgesics intake (=yes)		0.06	0.04	−0.02; 0.14	.117
Baseline pain (z-scores)		0.74	0.02	0.71; 0.78	<.001
Baseline FABQ-PA (z-scores)		0.07	0.02	0.04; 0.11	<.001
Baseline PHQ-9 (z-scores)		0.09	0.02	0.04; 0.13	<.001
Baseline GAD-7 (z-scores)		0.04	0.02	−0.01; 0.08	.082

Abbreviations: BFIS, Business, Financial, Information, and related services; BIC, Bayesian information criterion; BMI, body mass index; CFI, Comparative fit 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; GP, Goods-producing and manual labor; HE, Healthcare, Education and Social services; NDI, Neck Disability Index; ODI, Oswestry Disability Index; RMSEA, Root Mean Square Error of Approximation index; SDI, Social deprivation index; SE, standard error; SRMR, Standardized Root Mean Square Residual index; TTU, Trade, Transportation, and Utilities; WPAI, Work Productivity and Activity Impairment Questionnaire.

The model has 131 degrees of freedom. The reference category for each categorical variable was: BFIS for occupation, non-man for gender (ie, women and non-binary), non-Hispanic white for race/ethnicity, high school diploma or less for education, C1 (ie, lowest social deprivation) for SDI, and “no” for analgesics intake. Significant P-values are denoted in bold.

Chi-square difference testing confirmed that the identified predictors (ie, disability and occupation) contributed significantly to model fit, since its removal led to a significant deterioration in model fit (Δχ²>3.84; Table S4). In contrast, WPAI Overall and WPAI Activity constraints did not significantly worsen model fit, indicating that these variables did not contribute to the model.

Among all variables introduced in the model, baseline pain (β = 0.74, SE = 0.02, P<.001), fear-avoidance beliefs (β = 0.07, SE 0.02, P<.001), and depression symptoms (β = 0.09, SE = 0.02, P<.001) were also significantly associated with the pain outcome, with worse scores indicating higher last pain scores. The adjusted model incorporating all confounders reached an explained variance of 21.3%.

Sensitivity analysis

SEM-adjusted analysis considering treatment duration and total sessions was consistent with the main model (Table S5). The direction and significance of all predictors remained unchanged, with only decimal-level differences in the coefficient detected.

Moderation analysis

Consistent with previous evaluation, interaction-term models showed a consistent predictive effect of baseline disability and occupation on the last pain score across neck and low back pain, whereas work and non-work activities impairment were not predictors (Table S6). One interaction was observed between WPAI Activity and pain location, which was statistically significant, albeit to a very small effect (β=−0.02; Table S6).

The multiple-group SEM confirmed these findings, suggested by the observed similar predictor coefficients across neck and low back pain, supporting the absence of meaningful moderation (Table S7).

Discussion

Despite the expansion of digital care, predictive research has remained largely unexplored and confined to in-person rehabilitation. However, remote rehabilitation differs markedly from traditional settings, particularly regarding patient autonomy and adherence, asynchronous care delivery, and flexible structures. These unique, distinct features may alter the influence of known predictors, urging the need for dedicated evidence in digital contexts. This large, real-world study bridges this gap by examining baseline disability and job occupation factors associated with post-pain outcomes of a fully remote digital care.

Disability emerged as a significant predictor of final pain levels, with greater baseline disability levels being associated with higher last pain scores. Our results are consistent with the existing evidence base within in-person rehabilitation studies, albeit the diverse instruments and outcome definitions used.^19,21,33 Studies using ODI^20,21 have shown that greater disability predicts smaller improvements in pain (ie, pre-post change)²⁰ and lower odds of treatment response (ie, at least 30% reduction in both ODI and pain scores),²¹ and those using the Roland Morris Disability Questionnaire found an association with higher post-intervention pain levels.³³ Within digital care, only one prior study—focused on developing a prediction tool for pain response following the same DCP—pointed toward the potential predictive role of disability.⁴⁵ Present findings extend that work, reinforcing disability as a predictor in a fully-remote, asynchronous rehabilitation. Importantly, disability is easily measured through ODI and NDI—validated, low-burden instruments practical for clinical use, consistent with recommendations for predictive research.⁸

Conversely, work impairment (WPAI Overall), alongside non-work activity impairment (WPAI Activity), were not significant predictors after confounders adjustment, suggesting that variation in final pain levels was largely explained by the remaining factors. To our knowledge, this is the first study examining work-related factors as predictors in a digital rehabilitation context, and prior research in in-person settings is limited to one study linking work ability to post-treatment physical quality of life.¹⁸ Several mechanisms may explain our findings. Work impairment may only exert an indirect influence, mediated by other clinical (such as psychological factors) or unmeasured job-related characteristics (eg, job control, workplace support).²³ Indeed, work impairment has been associated with depression symptoms.⁴⁶ However, the absence of historical, pre-baseline data in this study hampers the establishment of the temporal direction of these associations. Another possibility is that stronger predictors in the model may have overshadowed any predictive contribution of WPAI due to conceptual overlap. Although WPAI Activity and ODI/NDI were moderately correlated, multicollinearity was not problematic (all VIF < 3.53), indicating that its non-significance was not attributable to statistical instability. This pattern aligns with the multidimensional structure of ODI/NDI, which include items featuring examples of daily activities, and a noteworthy item on pain intensity level, while WPAI activity relies on a single question about non-work-related impairment. Our findings underscore the importance of ODI/NDI as a more robust predictor of pain outcomes. Future studies are needed to explore interactions between WPAI and specific occupational characteristics or psychosocial factors on shaping recovery.

Occupation showed a significant predictive contribution, with patients working in business occupations (BFIS) reporting higher final pain levels than those in goods-producing (GP), and healthcare and education (HE). Although GP and HE occupations often involve physically demanding or repetitive tasks—recognized risk factors for musculoskeletal pain development, including spinal pain,²²—evidence on their influence on outcomes after in-person interventions is mixed.^32,47 A prior systematic review⁴⁷ has identified heavy physical demands as a predictor of poorer disability outcomes, but not of pain, suggesting that the impact of specific job characteristics may differ by outcome domain. Accordingly, whether occupational groups also predict additional domains (eg, disability) should be examined in future studies. On the contrary, the BFIS group—which was similar to the transportation, trade, and utility (TTU) group—is predominantly characterized by prolonged sitting and limited movement variability, which may contribute to higher last pain scores. Recent evidence in primary-care physical therapy has shown that limited ability to change posture during work is associated with higher odds of chronic pain, aligning with the patterns common in BFIS roles.⁴⁸ While more granular job descriptors (eg, physical demands, workplace support) could provide deeper mechanistic insight, such details are not always collected in routine clinical workflows in a standardized way, as it requires additional clinician time and may be difficult to implement consistently at scale. In this context, broader occupational grouping is easier to collect and may function as a useful proxy for work-related exposure patterns with meaningful predictive value, as suggested in this study.

The contribution of each candidate predictor was further supported by confirmatory analyses through chi-square difference testing, which pointed toward the predictive value of all candidate variables except WPAI Overall and WPAI Activity, strengthening our findings. Additionally, the effects of all candidate predictors were consistent across neck and low back pain, indicating that pain location was not a moderator and that their influence may be driven by systemic mechanisms. This aligns with emerging evidence that phenotypes—combinations of clinical, psychological, and behavioral features—hold greater value over diagnosis alone in determining outcomes.⁷ Disability and occupation may therefore represent potentially generalizable predictors across spinal conditions. Previous studies have found generic predictors across musculoskeletal conditions,³¹ and, particularly digital care, research has identified psychological factors¹⁶ (aligned with present findings in which fear-avoidance beliefs and depression contributed significantly). Future studies should analyze the predictive role of disability and occupation across other musculoskeletal conditions.

This study provides first evidence of the predictive value of disability and occupation post-pain outcomes of a remote intervention. Confirmatory studies designed to assess external validation and explore causal mechanisms are needed since the present findings reflect associations with end-of-program pain, not treatment effects of the DCP. These insights open new avenues for scalable predictive research in digital rehabilitation by identifying easily measurable factors that may support clinicians with recovery expectations setting, risk stratification, and treatment tailoring. They also create opportunities for future intervention studies to explore whether targeted strategies modify recovery trajectories. For instance, patients with high disability may benefit from a greater emphasis on exercise and graded activity,⁴⁹ while workplace-oriented strategies (such as regular active breaks^50,51 and workstations adjustments⁵²) might be relevant for business-related jobs and drive positive effects on pain. From a public-health perspective, advancing predictive knowledge may optimize resource allocation and support equitable access to high-quality musculoskeletal care.

This study contains key strengths, filling an important gap by examining predictors of pain outcomes in a digitally delivered rehabilitation—a context where predictive research is extremely limited. The use of real-world data from a large (N = 13 330) and diverse cohort, aligned to the U.S. population with MSK conditions,⁵³ represents a key strength. The application of robust statistical methodologies (SEM) enabled the acknowledgement of complex pathways, adjustment of estimations to mitigate confounding, and the handling of missing data. The breadth of variables enhanced statistical power, and the easy collection of the studied predictors strengthens its relevance and facilitates clinical translation. Finally, the study adhered to best-practice methodological guidelines for predictive modeling.⁸

Nevertheless, there are limitations that need consideration. First, despite adjustment for a comprehensive set of covariates, residual confounding from unmeasured variables (eg, health literacy, pain phenotype, comorbidities, including additional psychological factors) cannot be entirely excluded. Second, additional work factors (namely physical demands, job satisfaction, worker’s compensation, workplace support) were unavailable in this study and are still underexplored,⁴⁷ warranting further investigation. Third, occupation was categorized into broad sectors for analytical feasibility, which may conceal nuanced differences between jobs included in the same category. A more granular classification or based on more specific biomechanical aspects may provide a deeper understanding of the occupation’s potential role in recovery. Moreover, the absence of data on employment duration within the reported occupation hindered the assessment of potential exposure effects. Fourth, the study did not control for concurrent treatments other than baseline analgesic use. Lastly, the cohort consisted of insured individuals from the US, and external generalizability to uninsured populations or different healthcare contexts requires confirmation.

Consistent with the PROGRESS prognostic framework,²⁴ these findings represent an initial step in identifying predictors within digital rehabilitation. Further research should explore additional aforementioned candidate predictors and incorporate mediators that would allow to disentangle the complex interplay between factors. Although pain was prioritized as the outcome herein, as a core domain, it captures a single dimension of the multifactorial chronic pain experience and expanding analyses to other domains (eg, functional improvement or patient global impression of change) would enrich current knowledge. Future studies should consider decision-tree methods to translate findings into hierarchical frameworks to support decision-making. Moreover, evaluating whether specific intervention adjustments modify the influence of predictors is a critical next step. Finally, external validation in independent cohorts to establish causal importance of the identified predictors should also be pursued.

Conclusions

In this large-scale study of fully remote digital care for chronic spinal pain, baseline disability and occupation emerged as predictors associated with last reported pain levels, whereas work impairment was not a significant predictor. These associations were consistent across neck and low back pain. These findings emphasize the potential value for routinely screening of disability and occupation to support recovery expectations setting and communication, and identify individuals who may require greater clinical attention or targeted support. Future confirmatory studies are needed to further validate these predictors and determine how they can best inform personalized care strategies.

Author contributions

Conceptualization: D.J., D.P., F.C., and H.F.; Data curation: D.J.; Formal analysis: D.J. and X.T.; Interpretation of data: All; Writing—original draft preparation: D.J.; Writing—reviewing and editing: All. All authors have read and agreed to the published version of the manuscript.

Supplementary material

Supplementary material is available at Pain Medicine online.

Funding

This work was supported by Sword Health Inc. and also developed within the scope of project number 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.

Conflicts of interest

D.J. and F.C. are employees at Sword Health. X.T. receives scientific advisor honorarium from Sword Health.

Ethics approval

This study was approved by the Advarra IRB (Pro00063337) and was conducted in accordance with the Declaration of Helsinki.

Data availability

Data in this study can be made available on request from the corresponding author. Data are not publicly available due to privacy restrictions.