Introduction

Work-related physical traumas are a significant public health concern, accounting for about one in every six traumatic injuries in North America [1]. Workers with musculoskeletal (MSK) injuries often endure persistent pain and functional impairment, adversely affecting their quality of life and imposing a significant burden on both workers and employers through lost productivity, high medical costs, and considerable economic losses [2,3]. The complex interaction of physical symptoms, psychological and social factors that influence the severity and persistence of these pain and disability introduce challenge to effective rehabilitation [4–6].

Grounded in the notion that the best approach to managing chronic pain is to prevent it in the first place, one approach to mitigating the impact of chronic post-trauma pain is to optimize acute management [7]. Prognostic tools including questionnaires [8,9] or clinical prediction rules [10,11] are intended to identify the ‘high risk’ patient for targeted early intervention. Several such tools have been published, including the STarT Back [12], Orebro Musculoskeletal Pain Screening Questionnaire (OMPSQ, [13]) and Traumatic Injuries Distress Scale (TIDS, [14]) all previously demonstrating prognostic accuracy for predicting chronic pain or disability in different samples. Other, often shorter tools have also shown varying levels of prognostic validity, such as the Numeric Pain Rating Scale (NPRS, [15]), self-rated general health [16], or measures of depressive symptom severity [17].

The use of risk stratification tools has been facilitated by the publication of threshold or ‘cut’ scores, being the scale score that best discriminates between high and low risk groups. As an example, a prior meta-analysis exploring the prognostic validity of the NPRS in acute whiplash associated disorder (WAD) found that a score of 6/10 or greater provides the best discriminative accuracy between high and low risk groups [18]. We have previously reported that a TIDS score <3/24 is optimal for classifying people with acute MSK injuries into a low-risk category, while a score >11/24 is most accurate for identifying the high-risk category [19]. Importantly, to our knowledge few, if any, of these have been explored for differential function between people of different intersectional identities. Using the NPRS as an example, an abundance of evidence indicates that women [20], people of colour [21], or people of lower socioeconomic status [22] tend to rate higher mean pain intensity for a given stimulus or clinical condition when compared to men, white, and/or wealthier cohorts. If this is consistent, then the rote application of a single risk score to all people, regardless of personal identities or social pressures, may lead to inaccurate estimates of risk in people with acute pain or injury.

As complex computer-assisted models for prognosis and risk-targeted personalized treatment algorithms become more common, it is important to consider what personal or social identity variables should be included in such models. The purpose of this study is to examine differences in baseline values, 8-week prospective outcomes, and prognostic accuracy of three outcome variables, specifically when analyzed across social or demographic subgroups.

Materials and methods

This hypothesis-generating, cross-sectional and short prospective (8-week follow-up) study is part of a larger project (clinicaltrials.gov ID 06016647) conducted in collaboration with a network of community-based physiotherapy clinics providing Workplace Safety and Insurance Board (WSIB)-funded protocol-driven care in Ontario, Canada. It is reported according to the STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) statement [23]. It was conducted in accordance with principles of ethical research and received approval from the Western University Research Ethics Board (REB) prior to initiating recruitment. All participants provided written electronic (digital) consent prior to providing any data.

Social determinants

The social identity variables collected were sex at birth, felt gender, expressed gender, age in years, ethnoracial identity (collected according to the recommendations of the Canadian Institute for Health Information [27]), highest educational level, and annual household income. Table 1 presents the questions asked and response options provided for each variable. An additional variable, the Intersectional Discrimination Index – Major (InDI-M) [28] was collected as an omnibus indication of experiences of unfair treatment or violence as a “result of who you are” (e.g., sex, age, skin colour). It is a 13-item self-report questionnaire that includes experiences such as ‘has a healthcare provider ever refused you care?’ or ‘have you been threatened with a physical or sexual attack?’. Responses are 0 (never), 1 (once) or 2 (more than once) yielding a total score from 0 (no experiences of discrimination) to 26 (many repeated experiences of discrimination). The original development paper [28] supported known-groups validity, sufficient test-retest reliability and evidence of construct validity. We have previously identified a two-factor structure for the InDI-M reflecting two distinct dimensions of discrimination: systemic inequity, and interpersonal violence [29].

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Personal identity and socioeconomic variables collected for this study, and response options presented for each.

https://doi.org/10.1371/journal.pone.0346697.t001

Statistical analyses

As a hypothesis-generating study, analyses were not corrected for multiple comparisons.

Results

Totally, 1,956 injured workers presented to the clinical network for rehabilitation, of which 320 (16.4% of eligible population) people consented to participate. Of those, 203 (63.4% of consenting) provided adequately complete data for this analysis. The final sample was 59% female, mean age 43.9 years (range 20–73), and 61.6% identified as white only. The mean time from injury to completion of the intake survey was 49 days (range 2–90). At 8-week follow-up, 150 provided responses to at least one of the two recovery questions, of which 8 also indicated a new injury in the intervening 8 weeks and were removed. This left n = 142 (70.0% retention) for the AUC analyses using the self-reported recovery data, and we were able to match n = 190 (93.4% retention) with the CBI Health Group Inc. administrative database for the administrative recovery analysis. Comparing completers to non-completers, the only significant differences were for sex and race, with completers more likely to be female (65% vs. 45%) and to identify as white (67.4% vs. 48.4%) compared with non-completers. There were no other significant between-group differences.

Differences in pain, quality of life, and distress by social determinants

Older participants (>41 years) reported significantly higher mean pain intensity (mean = 5.5, 95%CI 5.0 to 5.9) compared to younger participants (mean = 4.3, 95%CI 3.9 to 4.7; p < 0.01). Those who reported higher experiences of intersectional discrimination (InDI-M > 3) scored significantly higher on the TIDS (mean = 11.1, 95%CI 10.0 to 12.1) compared to those who reported fewer experiences (mean = 9.2, 95%CI 8.0 to 10.4; p < 0.05). There were no other statistically significant group-based differences at baseline (Table 2).

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Mean scores and comparisons of difference in each of the three predictor variables by level of sociodemographic identity variable.

https://doi.org/10.1371/journal.pone.0346697.t002

Sociodemographic group-based differences in recovery outcomes

Recovery rates differed substantially based on the definition used. When asked directly, only 21.7% (30/138) responded ‘yes’ to the question about being fully recovered. When respondents were asked about their work status in relation to that held pre-injury, 47.9% (68/142) indicated they had returned to full pre-injury job duties. When querying the CBI Health Group Inc. administrative database, 54.7% (104/190) were considered fully back to work and discharged from further rehabilitation based on clinician or funder opinion. Table 3 presents the proportions of recovery status both overall and stratified by sociodemographic subgroups. Chi square analysis revealed only one difference in proportions that reached Bonferroni-corrected significance: those without a university education were less likely to be coded as fully recovered in the administrative data compared to those with a university undergraduate degree or higher (χ² = 7.60, p = 0.006). There were no other differences in proportions recovered across any of the outcomes or sociodemographic categories.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Proportions of recovery status both overall and when stratified by sociodemographic subgroups.

https://doi.org/10.1371/journal.pone.0346697.t003

Risk stratification accuracy

Tables 4–6 present the results of the AUC analyses. For self-rated recovery as the outcome, all three predictor tools showed significant main effect to discriminate between those who did and did not describe themselves as recovered 8 weeks later (NPRS: AUC = 0.66, 95%CI 0.56 to 0.76; TIDS: AUC = 0.75, 95%CI 0.64 to 0.86; EQ-5D-5L: AUC = 0.78, 95%CI 0.70 to 0.87). AUC of both the TIDS and EQ-5D-5L was significantly greater than the NPRS for this outcome. Main effects for discriminating between those who reported being fully returned to pre-injury work status compared to those not fully returned were significant for only the TIDS (AUC = 0.68, 95%CI 0.59 to 0.76) and EQ-5D-5L (AUC = 0.73, 0.65 to 0.82) but not the NPRS (AUC = 0.59, 95%CI 0.49 to 0.68). When administrative data on recovery status at completion of the PoC was the outcome, all three tools functioned almost identically and significantly better than chance (NPRS: AUC = 0.65, 95%CI 0.57 to 0.73; TIDS: AUC = 0.64, 95%CI 0.56 to 0.72; EQ-5D-5L: AUC = 0.64, 95%CI 0.57 to 0.72).

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Discriminative accuracy of NPRS, TIDS, and EQ-5D-5L for predicting self-rated recovery status.

https://doi.org/10.1371/journal.pone.0346697.t004

Download:

• PNG
  larger image
• TIFF
  original image

Table 5. Discriminative accuracy of NPRS, TIDS, and EQ-5D-5L for predicting self-rated work status.

https://doi.org/10.1371/journal.pone.0346697.t005

Download:

• PNG
  larger image
• TIFF
  original image

Table 6. Discriminative accuracy of NPRS, TIDS, and EQ-5D-5L for predicting administrative discharge status¹.

https://doi.org/10.1371/journal.pone.0346697.t006

When split by subgroups no comparison included any single cell with <10 participants. When the outcome was self-rated recovery, the main effects for both the NPRS and TIDS were only significantly better than chance for younger participants but not older, females but not males, those without a university education but not those with, and those with lower but not higher annual household income (Table 4). The NPRS also showed significant discriminative accuracy only for participants who identified as white but not for non-white, and with a lower InDI-M score but not a higher score. Post-hoc Z-tests indicated none of the AUC differences reached between-group statistical significance despite the significant main effects. The EQ-5D-5L functioned significantly better than chance across all subgroup analyses.

When the outcome was self-rated work status (fully returned / not fully returned), the NPRS was marginally better than chance accuracy for younger participants, those identifying as white, and those with lower household income, while showing no discriminative accuracy in older participants, those identifying as non-white, and those with higher household income (Table 5). Again, post-hocs indicated the magnitude of AUC was not significantly different between any of the subgroups. The TIDS showed significant main effects across all analyses except sex where both main effects and post-hocs indicated AUC was significant in females (AUC = 0.79, 95%CI 0.69 to 0.88) but not males (AUC = 0.47, 95%CI 0.30 to 0.63) with the difference reaching significant (p < 0.01). Again, the EQ5D-5L showed similar and significant AUC across all subgroup analyses.

When administrative discharge status was the outcome (Table 6), the NPRS showed significant discriminative accuracy in only the younger, male, white, less educated, and in both InDI-M groups, but no significant effect in older, female, non-white, higher-educated, and both income category groups. Post-hoc analysis indicated significantly better function in the younger (AUC = 0.72, 95%CI 0.61 to 0.83) compared to older (AUC = 0.54, 95%CI 0.42 to 0.67) groups (p < 0.01). The TIDS showed consistent and significant discriminative accuracy across all subgroup analyses save for older participants in which it failed to function better than chance accuracy (AUC = 0.56, 95%CI 0.44 to 0.68). The EQ5D-5L demonstrated better than chance accuracy across most subgroup analyses but failed to reach significance in the older (AUC = 0.59, 95%CI 0.47 to 0.71), female (AUC = 0.60, 95%CI 0.49 to 0.71) and lower income (AUC = 0.61, 95%CI 0.49 to 0.73) groups.

Discussion

The purpose of this study was to contribute to the growing discourse around personalized rehabilitation and prognosis in acute injury by exploring differences in baseline mean values, 8-week prospective outcomes, and discriminative (prognostic) accuracy across 3 relevant predictor and outcome variables when disaggregated by meaningful social identity or demographic subgroups. The results of these analyses provide direction for the development of future prognostic tools in adults (workers) with acute MSK injuries by providing evidence for the sociodemographic variables that can be expected to result in differential prognostic functioning.

Both the NPRS and TIDS have previously shown ability to predict recovery outcomes in acute MSK trauma [18,19] but to our knowledge the finding that the EQ5D-5L index score also functions as a prognostic screening tool in acute MSK trauma is novel. Importantly, in no prior analysis of any of these tools or of other prognostic screening tools has consideration been given to the potential for such tools to function differently, or require different risk cut scores, across different social or demographic strata. While we found little evidence that the outcomes could be simply predicted by sociodemographic variables (university education for predicting administrative outcome being the only exception), or that scores on the three prognostic tools are different (age for the NPRS and experiences of discrimination for the TIDS being the exceptions), we found several examples of potential differential functioning of the prognostic tools between sociodemographic subgroups. Assuming these findings hold in other samples, the most intuitive interpretation is that while there may be little need for different risk cut-scores on prognostic tools, there appears to be a need for deeper consideration about what prognostic tools to use for whom.

It is interesting that of the three tools that we explored, it is the one least aligned with prognosis and with the least prior exploration of prognostic utility (EQ5D-5L) that showed the most consistent accuracy across the three outcomes being predicted, and that its function was largely robust to differences in sociodemographic subgroups. That the NPRS showed the least robust accuracy across outcomes and sociodemographic characteristics is perhaps not surprising given its single-item nature. The TIDS functioned similarly to the EQ-5D-5L though with some notable exceptions. We note that of the three, the TIDS is the only one that was initially developed specifically for the purposes of risk/prognosis screening in acute MSK trauma and can also provide scores on three subscales (uncontrolled pain, negative affect, intrusion/hyperarousal) that are intended to be useful for prioritizing treatment targets in those deemed at higher risk of poor recovery. This added utility may still place it as the most desirable of the three tools explored here, though the function of the EQ5D-5L in this context appears to support further prognostic exploration.

The comparison of mean difference in the three predictors at intake between sociodemographic subgroups reveals interesting results both in terms of the significant differences and the non-significant differences. That the older participants scored on average 1.2 NPRS points higher than the younger participants is supported by considerable prior evidence [31], though does bring into question our own prior meta-analysis findings in which an NPRS score ≥6/10 was best for discriminating between high and low risk of non-recovery after acute neck trauma (whiplash, [18]). Based on existing data at the time, there was insufficient published evidence to suggest the 6/10 threshold should be different for older vs. younger respondents, but the findings here suggest that a single threshold applied equally across all patients may lead to inaccurate prognoses in at least a subset of the population. Despite prior work suggesting females tend to rate higher NPRS than males [32] we did not find that same effect in our data, which is difficult to interpret but may be due to differences in patient populations (acutely injured workers), sampling frames (community-based insurance-funded physical rehabilitation clinics), or sample size. The non-significant differences across levels of education, income, or ethnoracial identity provide some support for use of a single cut-score on these tools, though we acknowledge that the strata used here are quite broad because of low proportions for many identities. This is particularly problematic in the ethnoracial comparisons, where we were forced to collapse the responses into overly broad ‘white’ and ‘non-white’ categories owing to the low frequency of non-white identities in this sample – for purposes of both statistical power and to avoid re-identification of those in smaller categories. While we believe race-based analyses can be valuable when performed with appropriate attention to equity and have included those results here in part as an illustration of the practice, we must acknowledge that the overly-broad categorizations used very likely obscure differential effects for many racial identities. We have provided the full breakdown of ethnoracial identities in Supplemental S1 Table for illustrative purposes. These analyses should not be considered definitive evidence that there is or is not a race-based effect on prognostic tools, rather we have also chosen to present these results in this way in the hopes of opening conversation around equity in research. While calls for sex- and gender-based analyses [33] have contributed to more concerted efforts to recruit women and females into research, there has been less attention paid to recruiting different racialized groups. Perhaps our findings, that did not include special considerations for specifically recruiting non-white participants, will contribute to that conversation.

The only other significant finding was that those workers who indicated a higher number of prior experiences of intersectional discrimination also rated higher mean trauma-related distress on the TIDS. This study was not designed to explore mechanisms so any such commentary is speculative, though assuming experiences of discrimination are stressors this finding would fit with Cumulative Stress Theory (CST, [34]). CST posits a cumulative effect of stressors over a lifetime, such that subsequent stressors (like work injury) are experienced as more distressing by those who have experienced considerable prior stress. Of note is that the mean TIDS score of 11.1 in the high-discrimination group is at the previously published cut-score of 11 for higher risk of chronicity based on data from other cohorts [19]. This appears to align with prior work showing that the strongest predictors of poor outcome after acute injury are not the magnitude of damage or injury but rather the magnitude of distress experienced [35]. The potential for complex interactions or mutual synergy between prior life stressors and new trauma for predicting future recovery introduces an intriguing direction for further research. One such direction is the effect of integrating trauma-informed/assumed care or other strategies that manage cumulative psychological and social load in supporting recovery after injury, beyond focusing solely on the tissues at fault [36].

The data on recovery rates are also of interest and reveal important considerations for how such outcomes are defined. The concept of recovery as simply ‘a return to pre-injury state’ has long been challenged by scholars and patients alike [e.g., [37]] and discourse on the optimal recovery indicator for quantitative research remains unresolved. Using data from this study, when participants were directly asked whether they believed themselves to be fully recovered, without further delineation of what recovery means to each person, only 21.7% indicated they were fully or mostly recovered. When respondents were directly queried about their work status in relation to pre-injury status, 47.9% indicated they were performing their job duties at the same level. When recovery status was inferred from clinical administrative data, typically being a result of clinician interpretation in partnership with the patient (worker) and insurer, 54.7% were deemed recovered (fully back to work, no additional rehabilitation needed or offered). While recovery rates were similar when work status was explicit within the definition, those rates were considerably lower when recovery was defined by broader personal perspective. One logical interpretation is that people were returning to full work duties before feeling fully recovered – maybe not inherently problematic yet still illustrating the importance of clear and judicious selection of recovery indicators. This appears to have important implications for future prognostic work and will contribute to the ongoing discourse around the nature of recovery and how the concept can best be operationalized.

The results of the differential functioning of the three tools for discriminating between those who would and would not be deemed recovered after an 8-week standardized rehabilitation program require further confirmation though if replicated these findings create several potentially new lines of inquiry. As we anticipate prognostic tools to become increasingly capable especially in the context of machine learning tools, knowing what variables are important to collect and use in risk stratification is important. To start, both the TIDS and EQ-5D-5L were more accurate in discriminating between recovery groups than the NPRS, and both tools were more accurate when the outcomes were self-reported than when drawn from administrative data. That the EQ-5D-5L functioned largely the same regardless of sociodemographic identity suggests it is worth further exploration in independent cohorts as a potential prognostic tool after MSK injury. The TIDS also functioned largely consistent across sociodemographic strata though there were some subgroups (older age, males, university educated, higher income) for which it failed to discriminate between recovery groups at a level greater than chance for some of the outcomes. While potentially a function of lower-than-expected recovery rates and small numbers in some cells of the 2x2 table introducing the potential for imprecision, this suggests that distress as a predictor of outcome may be most suited for use in younger people, females, those without university degrees and those with lower annual income. If replicated, this would indicate that the type of prognostic tool selected for an individual patient should be at least partly influenced by social and demographic identities of the respondent. However, the only comparison in which the AUC for the TIDS was significantly different between strata was seen when self-reported work status was the outcome and the sample was disaggregated by sex. In that analysis, the TIDS accurately discriminated between recovery groups in females but not males. This finding differs from our prior analysis of TIDS functioning by sex in an independent sample of n = 224 acutely injured adults from all-cause trauma (not only work injuries) [19]. While it is difficult to interpret the difference, the prior study focused on self-rated pain and disability scores as the outcomes where herein we see the largest difference when self-rated work status is the outcome. We also acknowledge the large number of analyses for which the significance threshold (p) was not adjusted, raising the possibility that this is an alpha error. Regardless of the mechanism, the findings of the current study have opened the possibility that males may require a different prognostic tool than females, which if true is the first time such a finding has been shown to our knowledge and creates space to question the results of prior work on prognostic tools. Similarly, the EQ-5D-5L results may signal something about the items in that scale or its scoring algorithm that lend to its robustness across demographic strata, from which other prognostic tools may benefit.

There are important limitations to consider. This work was conducted within a single Canadian province, with rehabilitation provided by clinics within a single large network, and under the funding models and guided by protocols of a single rehabilitation funder. While these helped strengthen some aspects of internal validity, they also provide questions on generalizability that would benefit from work outside of these jurisdictions. As mentioned earlier, the sociodemographic variables were dichotomized into groups that are arguably overly broad and may not hold intuitive meaning. Our selection of thresholds on variables other than sex was largely driven by the data, using median splits as a meaningful distinction that also permitted interpretable AUC analyses. Arguments could easily be made for choosing different thresholds, for creating more categories (though sample sizes would get quite small in some cells) or for leaving age and InDI-M score in their raw ordinal or continuous forms. Some of these decisions were a priori while others were a posteriori when the lower-than-expected rates of recovery were identified. We acknowledge that dichotomizing variables can obscure some (non)-linear effects, and expect that with larger samples there will be opportunities to explore finer gradations in differential effects across levels of constructs like age, income, or education. We also note that sample size likely played a role in some of the findings being just shy of Z-test significance (e.g., NPRS AUC when self-reported work status was the outcome). A larger sample would likely have reduced the width of the confidence intervals and may have led to more significant findings. We also highlight the exploratory nature of the analyses of mean baseline values of the three predictors, for which we did not correct the alpha error rate based on multiple analyses. This could have led to some means being significantly different by chance, though the only finding that this could have affected was the TIDS x InDI-M categories, for which a familywise Bonferroni corrected p value would have made that difference not statistically significant. The NPRS x Age comparison would remain significant, and there were no other significant findings, rendering this a mild concern. The almost fully self-reported nature of the data introduces both strengths and limitations, a strength in that the results are patient-centred (reported directly and independently by the patient) without risk of observer bias, though limited as these were completed without oversight meaning we cannot ensure whether adequate attention was paid or that the questions were properly interpreted. This is a common limitation of survey-based research, and the ability to also mine the administrative database, which was created through opinions of clinicians, funders, and administrators, was a useful and different source of information for these analyses.

Conclusions

The purpose of this hypothesis-generating study was to explore the ways that different sociodemographic identities influence either the scores expected on three common clinical self-report tools, or the ability of those tools to predict outcomes after an 8-week standardized rehabilitation program. The results indicate that when using NPRS as a potential risk/prognosis screening tool after acute work injuries consideration should be given to using different cut scores for older and younger respondents, and not at all depending on the outcome being predicted. Scores on the TIDS, a tool specifically intended for risk/prognosis in MSK trauma, were only different across levels of prior experiences of intersectional discrimination which supports Cumulative Stress Theory but is unlikely to be easily implemented in practice. Recovery rates differed considerably based on the definition used, from 21.7% to 54.7%, highlighting the importance of careful selection of recovery indicators in prognostic research. Finally, both the TIDS and EQ5D-5L showed consistent and largely similar ability to discriminate between those who would and would not eventually report recovery, though findings provided some evidence of differential functioning of the TIDS by sociodemographic categories that may suggest a different prognostic/risk screening tool be considered for some groups, such as males. We are unaware of prior work on any prognostic screening tool for MSK pain that has explored these differential effects. Several new lines of inquiry have been proposed as a result of these findings.

Supporting information

Acknowledgments

We thank the staff and clients at CBI Health Inc. in Ontario for their contributions to the success of this study.