Journal Logo


Framework for improving outcome prediction for acute to chronic low back pain transitions

George, Steven Z.a,*; Lentz, Trevor A.a; Beneciuk, Jason M.b,c; Bhavsar, Nrupen A.d; Mundt, Jennifer M.e; Boissoneault, Jefff

Author Information
doi: 10.1097/PR9.0000000000000809
  • Open


1. Introduction

Chronic pain occurs more frequently than other conditions already widely accepted as public health priorities, with an overall prevalence higher than diabetes, cardiovascular disease, and cancer combined.45 Exact estimates vary based on case definitions, but the prevalence for chronic pain has been reported to be as high as 110 million people in the United States.45 The economic impact of chronic pain is accordingly large, with direct and indirect costs totaling $650 billion.45 In the United States41,65 and Canada,24 the ongoing opioid crisis is further evidence of chronic pain's societal impact. Low back pain (LBP) is the largest subset of chronic pain conditions,45 and rates are increasing. For example, the prevalence of chronic back pain in North Carolina increased from 3.9% in 1992 to 10.2% in 2006.27 Since 1990, the global prevalence of LBP has increased by 17.3%, and it continues to be a leading cause of global years lived with disability.34,35

Accordingly, LBP pain is one of the most common reasons to seek health care.27,45 Clinical practice guidelines19,73 and the Federal Pain Research Strategy (United States)31 highlight priorities for addressing the discord between increasing health care utilization and growing societal impact of LBP.34,35 Limiting the transition of acute pain to chronic LBP is a top research priority cited in these clinical practice guidelines19,73 and the Federal Pain Research Strategy (United States).31 Improving prediction accuracy for transition to chronic LBP is a vital precursor to development effective treatment strategies that limit this transition. For example, the Federal Pain Research Strategy (United States) has highlighted the importance of optimizing screening tools for predicting the development of persistent pain conditions. Implementation of systematic approaches with high predictive accuracy is likely necessary before health care systems can efficiently manage acute LBP by preventing development of chronic LBP conditions.59

In this review, we first describe variability in predictors, outcome measures, analytical approaches, and predictive accuracy from previous studies investigating the acute to chronic LBP transition. The variability in these factors were identified by mapping review, which is used to characterize the quantity and quality of a body of literature for the purposes of making recommendations for future research.38 We then describe a standardized predictive framework for the transition from acute to chronic LBP that aims to address the variability identified from the mapping review by improving methods for predicting the development of chronic LBP. Application of this predictive framework will be considered with the goal of implementing in real-world settings in mind by leveraging novel data sources, including the electronic health record (EHR), to improve prediction of chronic LBP.

2. Mapping review

In comparison with other review options (eg, scoping or systematic review), mapping reviews search based on time and scope constraints.38 Therefore, mapping reviews are not meant to offer an exhaustive, comprehensive, and/or definitive review of a topic. Instead, they are used to characterize a body of literature by identifying key study design elements, with the overall goal of providing informed direction for future primary or secondary research.38 Mapping review results are typically presented in tabular format at the individual study level (ie, no attempt at pooling) and without formal quality assessment.38 A mapping review was included in this review to provide structure to identifying sources of variability from previous predictive studies of acute to chronic LBP transition. The sources of variability selected to highlight from the mapping review were in areas relevant to developing a predictive framework and included: (1) individual predictors; (2) outcome measures; (3) analytical approach; and (4) prediction accuracy.

2.1. Search strategy and study selection

PubMed and Google Scholar searches were conducted in July 2018 using combinations of the following terms: back pain, LBP, acute to chronic, persistent pain, chronicity, prospective, longitudinal, long-term, prognosis, prognostic, predict, outcome, and transition. Potential articles were then identified by independently screening titles and abstracts. Full articles were evaluated by 2 coauthors (J.M.M. and J.B.), who reached consensus on inclusion through discussion. In cases when consensus could not be reached, a third author (S.Z.G.) provided an independent assessment of the article.

Studies were selected for the mapping review based on the following criteria: (1) study population consisting of individuals with acute or subacute LBP (<3 months), (2) follow-up period of at least 12 months, (3) examined predictors of LBP outcomes (rather than only measuring the likelihood of having certain outcomes), (4) not a clinical trial, and (5) used clinically feasible measures for predictors and outcome (eg, excluding structural/functional MRI studies). We did identify 2 neuroimaging studies that characterized brain-derived markers for predicting chronicity of LBP.2,68 However, their focus was mechanistic and not readily applied in contemporary clinical settings. Thus, although such approaches may become clinically feasible in the future, they were excluded from this mapping review. Twenty articles (representing 19 cohorts) meeting these criteria were identified, and key characteristics (ie, sample size, follow-up period, primary outcome, accuracy estimates, and base rates of recovery) are summarized in Table 1.7–10,12,18,22,23,30,33,39,40,43,53,54,56,64,75,77,81

Table 1
Table 1:
Study characteristics and accuracy of predicting low back pain outcomes.
Table 1-A
Table 1-A:
Study characteristics and accuracy of predicting low back pain outcomes.

2.2. Individual predictors

Predictor variables included in each cohort, whether they contributed statistically to the outcome of interest or not, are summarized in Table 2. These studies included predictors falling broadly into demographic, pain, general health, psychosocial, and occupational domains. The number and variety of predictors examined for LBP outcomes is substantial as is the lack of consistency across studies. We acknowledge that the inconsistency of predictors is likely due to researchers' goals for each analysis. Table 2 also highlights areas that have been underrepresented as predictors, including comorbid conditions (2/20 articles) or health behaviors such as physical activity (7/20 articles). Another weakness of the current literature was that some relevant health behaviors were virtually unexplored (eg, alcohol use, drug use, and sleep disturbance) in the 20 articles included in this review.

Table 2
Table 2:
Predictors of chronic low back pain examined in longitudinal studies of at least 12-month duration.
Table 2-A
Table 2-A:
Predictors of chronic low back pain examined in longitudinal studies of at least 12-month duration.

2.3. Outcome measures

Outcomes examined in each study are summarized in Table 3. Most studies examined multiple outcome domains related to defining chronic LBP as an outcome. Functional disability was the most commonly reported outcome (9/20 articles), followed by work status or pain-related work absence (8/20 articles). Measures of pain presence (a dichotomous measure) or pain intensity (a continuous measure) were used as outcomes in 6/20 articles. Most studies defined and reported rates of recovery in terms of these pain outcomes.

Table 3
Table 3:
Low back pain outcomes examined in longitudinal studies of at least 12-month duration.

2.4. Analytical approach

The review included 10 multivariate linear regression models for continuous outcomes. For categorical outcomes, there were 7 multivariate logistic regression models reported.

2.5. Prediction accuracy

Key study characteristics and factors that determined prediction accuracy for a given study (ie, variance accounted for, recovery base rate, recovery criterion, and classification summary) are summarized in Table 1. Only 1 of the linear regression models explained more than 50% variance in continuous outcomes. For the categorical outcomes, 5 of the 7 multivariate models reported classification accuracy higher than base rates. Base rates were calculated from the proportion of patients with acute or subacute back pain reporting they achieved recovery criterion at follow-up. The range for improvement of classification rates over base rate was 4% to 30%, with only one model reported improvement greater than 10% over the base rate of transition.

3. Current state of acute to chronic low back pain prediction

As expected, the mapping review identified considerable variety in individual measures used to predict the development of chronic LBP (Table 2). There are so many specific measures (or measurement tools) available for predictive modelling identified in the review that it is unlikely standardizing individual predictor variables will be feasible. However, the mapping review did identify opportunities for ensuring representation of each relevant predictor domain. This seems to be an important consideration for predictive frameworks to consider as many of the individual measures used across different studies for a given predictor domain are likely to be highly correlated (ie, different depressive symptom measures). Emphasizing consistency in predictor domain representation in predictive models may improve capabilities to compare model performance or pool data in future analyses. All predictive models included in the mapping review incorporated baseline predictive measures. There may be justification for including select time-varying elements as models that incorporate both static (eg, age, sex, and socioeconomic status) and selected time-varying elements (eg, changes in pain, disability, or psychological distress.6,32) may improve prediction accuracy.

The studies included in the mapping review used a wide variety of outcome measures. This does not necessarily indicate a problematic lack of standardization between studies because choice of outcome measures depends on specific research goals. A study focused on the impact of back pain on inability to work should include return to work as the primary outcome measure. However, such a focus prevents progress in identifying predictive factors that generalize to other outcomes relevant to the development of chronic LBP. For example, a model designed specifically for accurate prediction of pain intensity may not be well suited for predicting disability or patient satisfaction.48 Instead of having a given predictive approach linked to one outcome, there is an opportunity to test a standardized predictive framework across multiple outcome measures that are representative of chronic LBP. Testing the same predictive model for accuracy across multiple outcomes will avoid over specification of predictive approaches (ie, needing separate predictive models for each outcome of interest), and in the process of being simpler to implement, a single predictive framework may have a broader impact for informing clinical decision making.

Given that most of the approaches included in the mapping review incorporated linear or logistic regression, there is an opportunity to explore if novel analytical approaches have the potential to improve prediction accuracy. In particular, novel approaches that leverage machine learning and artificial intelligence methods that identify patterns in care and account for the emergent and dynamic nature of chronic LBP pain development may improve identification of those at risk. Finally, and perhaps the most obvious indicator that new predictive approaches need to be considered, during the 21-year period covered by this mapping review, there was no trend of improved predictive accuracy.

4. Proposed framework for improving prediction of transition from acute to chronic low back pain

The proposed framework for improving prediction of acute to chronic LBP is described in Figure 1 and presented in more detail in the subsequent sections.

Figure 1.
Figure 1.:
Predictive framework for predicting transition from acute to chronic low back pain.

4.1. Standard predictor domains

Previous predictive studies in LBP have considerable variability in demographic, pain, health, psychosocial, and occupational domains. One problem with the use of this many domains is that any one particular study very rarely had representation from each predictor domain. In response, we are including standardized predictor domains in the proposed framework. Standardized predictor domains will allow for better direct comparison of predictive accuracy and also allow for models to be tested for accuracy across multiple outcomes representative of development of chronic LBP.

Although domain standardization is emphasized in this framework, the specific measures used remain an important issue to consider. One important issue is to ensure including measures representing both modifiable and are considered direct treatment targets (eg, baseline pain intensity), as well as nonmodifiable factors for whom treatments can be tailored (eg, age, since care approaches may differ for younger vs older patients). A robust predictive framework will include a mix of modifiable and nonmodifiable factors with the goal of maximizing potential of predictive accuracy for chronic LBP development. Another important consideration is that the specific measures must be pragmatic for capture using electronic health record, and not greatly increase patient or provider burden. Therefore, to improve likelihood for successful implementation, it is recommended that a minimum set of variables be used to represent each domain.

4.2. Minimum variable set for each predictor domain

Representing the demographic domain are variables capturing individual characteristics and social determinants of health. Individual characteristics in the demographic domain are commonly captured in studies predicting MSK outcomes often including age, sex, gender, race, and ethnicity. However, social determinants of health are not often included, and emerging evidence suggests that valuable insights can be gained from this predictor. For example, having Medicaid coverage was an independent predictor of poorer LBP outcomes in a cohort study when compared with a validated LBP screening tool,50 whereas lower education and income levels decreased the positive effects of psychologically informed stratified care in a randomized trial.5

Recommended specific measures for the pain domain are variables representing clinical history (eg, duration of symptoms and history of previous conditions) and the pain experience (eg, anatomical location, pain severity, and pain impact). The pain domain has been frequently included in previous studies, often as measures of pain intensity or duration. However, to improve predictive accuracy, it may be necessary to have broader representation of this domain. For example, a recent study has indicated that multiple sites of pain can be predictive of poorer LBP outcomes.74 Adequate representation of the pain domain beyond intensity and duration is necessary to allow for clear determination of which characteristics of the pain experience have strong and consistent temporal associations with the development of chronic LBP.

The health status domain is not commonly represented in longitudinal studies predicting MSK pain outcomes. Therefore, it is important to adequately represent this domain in future predictive studies. Recommended specific measures for representing this domain are variables for health-related quality of life (eg, functional status and mental health) and disease burden (eg, comorbidity) measures. Quality of life measures are well established in the study of LBP; however, comorbidity measures have not been commonly used. Comorbidity represents an emerging area of interest for the prediction of chronic LBP. Measuring comorbidity number is the current standard, as demonstrated in a recent cohort analysis indicating lower comorbidity was protective of having persistent pain 12 months after seeking physical therapy care for a variety of musculoskeletal pain conditions, including LBP.6 Approaches that systematically consider comorbidity in addition to number will allow for broader consideration of disease impact and may lead to better accuracy for prediction of LBP outcomes.

The psychosocial domain has been highly studied in prediction of LBP outcomes. This domain includes the cognitive, affective, and behavioral aspects, and collectively, the psychosocial domain has been used to determine the overall level of distress associated with LBP. Many different individual psychosocial measures have been studied, and they can be broadly categorized into negative affect (eg, depressive symptoms and anxiety) and coping styles (eg, fear avoidance, pain catastrophizing, and self-efficacy). Psychosocial measures consistently predict LBP outcomes in cohort studies.6,26,32 However, head to head comparisons of commonly used screening tools indicate statistical similarity, making recommendation of a specific measurement approach difficult because there is no superior single measure.3,48,49 Instead, it seems important to ensure the measures used capture negative mood and coping styles, and both negative (eg, fear avoidance and catastrophizing) and positive (eg, self-efficacy and acceptance) coping are measured to represent this domain.58

The final domain to consider in this predictive framework is the individual context domain. As per the mapping review, specific measures recommended for this domain have included occupational factors (eg, job satisfaction and perceived work stress). By contrast, for nonoccupational cohorts, this domain has not been well represented. Therefore, it will be necessary to represent this domain with specific measures that capture the perceptions of receiving care, including patient expectations and treatment preferences. For example, a validated prediction tool for the development of chronic LBP included one item on the expectation of having persistent pain in its final 5-item version.79 Beyond that example, the individual context domain has been largely unexplored in the transition from acute to chronic LBP prediction studies. Including this a standardized domain in future studies could be an important way to improve prediction accuracy.55

4.3. Time-varying factors

Traditionally, prediction of LBP outcomes has included static, baseline determinants of risk. The primary limitation with this approach is that it does not account for any time-varying factors of the care episode that may indicate change in the initial risk status.21,42 Static risk determination may be an acceptable strategy for certain nonmodifiable factors; however, it inherently limits the impact modifiable, time-varying factors have on outcome prediction. Without accounting for such changes in modifiable factors, predictive models cannot distinguish between an outcome driven by an overall poor prognosis vs an initial poor treatment response. Models that account for this distinction by including static and time-varying factors are important for advancing LBP outcome prediction.

Recent evidence from LBP studies have demonstrated that predictive approaches allowing for early changes that occur when receiving health care can improve predictive accuracy for treatment outcomes.4,32,83 This process has been described as “treatment monitoring,” and in LBP, it often accounts for changes in the psychosocial domain to improve on baseline risk determination.4,32,83 However, since not all studies will involve care seeking cohorts or monitoring could continue following the end of formal treatment, the term “longitudinal monitoring” will be used in this framework to describe the capture of time-varying factors. Psychosocial measures are the most obvious choice for longitudinal monitoring, given the current state of the literature. There are likely other time-varying measures that can be used for longitudinal monitoring; however, these have not been clearly identified as this line of research is still emerging. Accounting for longitudinal monitoring in predictive models does increase the burden of data collection, as capturing multiple, patient-level time points are required. However, collection of these factors need not be comprehensive, should be driven by empirical evidence, and may be amenable to use of mobile applications or wearable technology.16 For instance, several treatment mediators57,66,67 for LBP outcomes have already been identified, and many of these are from the psychosocial (eg, fear avoidance and self-efficacy) or pain (eg, pain intensity) domains. Therefore, only these variables would be included in predictive models until other time-varying factors are confirmed through external validation studies.

4.4. Outcomes

There are multiple definitions of chronic pain in the literature, yet no one definition is widely enough accepted to be considered as a standard.80 Recent International Classification of Disease recommendations provide diagnostic codes from chronic pain as a primary condition70 and also codes for secondary conditions such as musculoskeletal71 or postoperative pain.76 These diagnostic codes will be very helpful in identifying those that already have chronic pain, but these codes do not directly address which outcome measures should be used for predicting acute to chronic LBP transitions. The lack of standard definitions for what constitutes chronic LBP means there is a need for different perspectives on which specific measures should be used to define chronic LBP. At a minimum, the patient, provider, and payer perspectives should be considered because there is an expectation that the most robust definitions of chronic LBP will be a convergence of these perspectives. This means that prediction models will need to be flexible to allow for the prediction of different definitions of chronic LBP and not any one measure alone.59 This creates the need for the accuracy of a given predictive model to be simultaneously tested across multiple outcome measures, a different approach than was identified in the mapping review (ie, most studies had single primary outcome). Outcomes that can be used to accommodate multiple definitions of chronic LBP are a priority in this predictive framework and include (1) incident chronic pain state, (2) quality of life, (3) health care utilization, and (4) health care costs. The capture of these outcomes, while not exhaustive, would provide enough information to meet multiple definitions of chronic LBP and thereby providing better support to subsequent clinical, policy, and public health actions.

5. Implementation of predictive framework for predicting low back pain outcomes

5.1. Application example

The National Institutes of Health (United States) Pain Consortium convened a Research Task Force for chronic LBP in 2009 to 2010. The charge of the Research Task Force was to review definitions, diagnostic criteria, and outcome measures for clinical research, develop a draft set of standards for research on chronic LBP, and engage the research community and government agencies in developing research standards. The Research Task Force disseminated their recommendations, which included a minimal data to support research standards.17 This minimal data set can be used as an application example for this predictive framework, while fully acknowledging there is a much larger pool of potential measures available. Following the Research Task Force, recommendations has the advantages of including measures already vetted and endorsed by an expert, multidisciplinary panel, and emphasizing a pragmatic approach by including a standard data set that may make this framework easier to implement in real-world settings.

The Research Task Force minimal data set includes 40 items; many derived from previously validated questionnaires like the Start Back Screening Tool and Patient Reported Outcome Measurement Information System domains.17 The minimal data set was designed to be broad enough to capture domains applicable to stakeholder groups including patients, providers, and policy makers. In Table 4, we have listed each of the Research Task Force's items and indicated how they could be represented as predictor or outcome domains in this predictive framework. There is coverage of these domains by minimal data set items, consistent with the Research Task Force's charge. Therefore, practitioners or researchers looking to adopt this predictive framework could use the Research Task Force's minimal data set as a starting point for implementation in their setting.

Table 4
Table 4:
Application of predictive framework with the National Institutes of Health chronic low back pain research task force recommendations for a minimal data set.

There are, of course, caveats to consider when using the Research Task Force's minimal data set within this predictive framework. First, there are some measures considered as both predictors and outcomes in Table 4. This may be entirely appropriate given the research question (eg, knowing the present pain state to predict a future pain state); however, care must be taken when interpreting models that include the same measures as both predictors (ie, independent variables) and outcomes (ie, dependent variables). One potential way to address this issue is to compare prediction characteristics of models with and without the baseline dependent variable included to inform the impact on prediction accuracy. Second, Table 4 shows there are certain domains that may be better represented with additional measures beyond the Research Task Force's recommendations. Although several items in the minimal data set could be used as time-varying factors, it may be better to have full-length questionnaires representing this domain as they are more sensitive to change (eg, instead of using one item from the Pain Catastrophizing Scale would use the entire questionnaire for longitudinal monitoring). This line of research (ie, time-varying factors and associated psychometric properties) is still emerging, and we have cite several examples for those interested in more details.4,32,83 The context domain was only represented by 3 items (Table 4), and these items were specific to those with LBP that was work-related or having legal involvement. These are important factors to capture, but researchers interested in other contextual issues (eg, treatment expectations and health care system characteristics) would need to include additional measures. Collectively, these caveats provide examples of how the Research Task Force's recommendations can be adapted to better address specific research questions while applying the proposed predictive framework.

5.2. Analytical considerations

There is growing interest in the use of machine learning methods to improve the health of patients by identifying latent patterns in data that can aid in prediction.25 Machine learning is a subset of artificial intelligence that aims to train computers (ie, machines) to improve the performance of tasks such as prediction through supervised, semisupervised, and unsupervised approaches. Machine learning–based approaches can address some of the limitations of traditional regression approaches, including nonlinearities, heterogeneity of effects (ie, interactions), and numerous, complex predictor variables.36 They can also be used to address missing data, which are common in studies conducted using administrative and health record data. To be sure, data are only collected when patients interact with health systems, and previous work has shown that this interaction can impact inference72 and risk prediction.82 Machine learning–based imputation methods, such as multilayer perceptron, k-nearest neighbor, and self-organization maps have been shown to outperform traditional statistical approaches for risk prediction.47,82 However, these approaches are not commonly used in LBP research to address missing data. Common machine learning methods used in pain research include classification for clinical diagnosis, structure detection to identify clusters of patients, and knowledge discovery to discover patterns in clinical data. Specifically, it has been proposed that areas where machine learning–based approaches may impact pain research the most include phenotyping and classifying pain,11,63 predicting outcomes of interventions to address pain,84 and differentiating pain from other physiological signals.46

Although there is a need to consider machine learning approaches in future predictive modeling, there is also the need to ensure they are adequately tested against traditional approaches. For example, in risk prediction models for mortality for patients receiving hemodialysis, there was no advantage of more complex approaches (eg, machine learning) when a range of statistical approaches were considered.37 In a systematic review of clinical prediction models across a variety of practice areas and including studies with a wide range of sample sizes, machine learning (ie, classification trees, random forests, artificial neural networks, and support vector machines) was compared with logistic regression.14 For the 71 studies included in this review, there was no evidence of better prediction performance for the machine learning approaches. Therefore, while consideration of advance analytical approaches is necessary to determine whether their use results in better predictive accuracy than identified in our mapping review, it cannot be assumed that machine learning approaches will always outperform traditional approaches for improved accuracy.

Another important aspect of prediction models for chronic LBP outcomes—whether derived through traditional or novel analytical approaches—is the need for validation across multiple health systems. Models developed in a single health system may not have the same prediction characteristics as models that are validated across multiple health systems. This is of particular importance because there is increasing interest and need to leverage real-world data for predicting LBP outcomes, yet most approaches are developed and validated in a single health system using a cohort approach. This can be problematic for generalization of predictive models because previous research in other clinical areas has shown that patients recruited through cohort studies and clinical trials are not necessarily representative of real-world patients.52 As the volume and velocity of real-world data increases, great care must be taken to validate results generated from a single health system to determine how useful that predictive framework will be in other health systems. Available data resources in the United States that might be helpful for establishing the generalizability of findings within specific health systems include the American Physical Therapy Association's Outcomes Registry, commercial payer databases (eg, Optum Labs, MarketScan), and publicly available population-based data sets (eg, Medical Expenditures Panel Survey, and National Health and Nutrition Examination Survey). For example, retrospective cohort data from commercially insured US adults have been used to identify how early exposure to nonpharmacological providers limits short- and long-term opioid use for those with LBP51 and establishing the risk of serious infection among users of biologics for psoriasis and psoriatic arthritis.62

6. How this predictive framework will advance research and practice

This review presented a framework that could be used in future studies to improve predictive modeling approaches for studying the transition from acute to chronic LBP pain. The predictive approach proposed in Figure 1 promotes standardization of predictor domains and multiple outcome measures that represent chronic LBP. This framework provides the opportunity to develop and test models in a structured manner to determine whether improvements in predictive accuracy occur. The framework was developed to be used as a companion to recommendations for improving methodological reporting of predictive studies, for example, the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) statement.15 Awareness of the TRIPOD statement will ensure those studying the transition from acute to chronic LBP report the essential methodological elements of predictive studies, which are often missing when the models are published.44 The emphasis on standardizing predictive domains provides flexibility for this framework to be adapted for widespread testing in the electronic health record. This predictive framework provides directions for approaches that can be adopted and integrated into the electronic health record, so that it evolves past being an administrative platform to a critical component of learning health systems. Widespread testing will result in validation and refinement to prediction models that improve accuracy and the development of clinical decision aids that could be used to support treatments that better limit the transition from acute to chronic LBP.

Improving the prediction accuracy of the transition to chronic LBP is strongly aligned with recent clinical practice guidelines19,73 and the Federal Pain Research Strategy (United States).31 Refined prediction can enhance value of care by identifying individuals appropriate for condensed care episodes or alternative delivery options that are less resource intensive (eg, telehealth options) based on their risk of developing chronic LBP. For instance, patients in a care pathway designed to accommodate a low risk to transition to chronic LBP may be appropriate for exposure to a variety of nonpharmacological treatments. This low-risk pathway would also have strict criteria for escalation to care options that have higher risk and no guarantee of additional benefit (eg, injections, opioids, or surgery) because the overall prognosis is generally favorable. By contrast, patients in a high risk to transition to chronic LBP pathway would be more closely monitored with pain and quality of life measures, so that timely and appropriate systematic decisions could be made for care escalation.

Early and accurate prediction of the development of chronic LBP will allow for efficient distribution of health care resources at the initial point of care. For LBP, this initial point of care is extremely important because it can have dramatic effects on downstream pain-related outcomes, health care utilization, and costs.28,29 In this manner, the updated predictive framework would facilitate delivering value by aligning effective care with utilization and cost resources, consistent with the Institute for Healthcare Improvement Triple Aim Initiative.59 More efficient resource allocation can only be accomplished by more accurate identification of individuals that are going to subsequently resolve their acute pain condition vs those that are going to progress into a chronic condition.

Existing predictive approaches for chronic LBP outcomes did not incorporate time-varying, modifiable factors to refine outcome prediction. The proposed predictive framework adds time-varying factors through longitudinal monitoring, consistent with the literature citing the importance of treatment mediators57,66,67 and the type of monitoring that has already been previously described for LBP.4,32,83 This addition is burdensome, and in that it adds another data collection point beyond baseline, but it is a necessary step to prepare for moving toward dynamic modeling of LBP pain outcomes. There is already evidence supporting dynamic models to predict outcome of other chronic nervous system diseases, including incident Alzheimer's and Huntington's Disease.60,61 Implementation of this predictive framework will enable development of similar approaches to predict incident chronic LBP. Dynamic predictive models allow for learning health systems in which multiple time points can be used to more accurately determine risk status, with care options adjusted in real time. For example, a short-term decrease in pain after exercise therapy that is indicative of long-term recovery from back pain may result in a real-time decision to delay spine surgery or avoid use of prescription opioids.

Finally, improved accuracy in outcome prediction could reduce uncertainty surrounding optimal LBP management strategies. This uncertainty is driven by multiple pharmacological and nonpharmacological treatment options that have very similar treatment effects.13,78 This lack of treatment superiority for any given treatment option clouds clinical decision-making. It is likely that this lack of definitive treatment superiority contributes to the unwarranted variability in health care delivery observed for LBP. A validated prediction framework could ultimately reduce this uncertainty (and the associated care variability) moot by providing accurate long-term estimation of developing a chronic pain state. Interestingly, empirically based approaches for predicting outcomes have been used in other areas of medicine, including watchful waiting in prostate cancer20 or shared decision-making for left ventricular assist device in heart failure.1,69 For LBP, there is potential for predictive models with increased accuracy to advance care decision-making in similar ways, either by routine monitoring of patients to assure an initial prognosis remains favorable or by using the likelihood of chronic LBP development to inform the length and intensity of a care plan. Importantly, we present a framework for enhanced prediction, but this framework is not intended to result in the development of static models. As additional potential predictors are identified, particularly those amenable to large scale application in real-world settings (eg, neuroimaging of brain structure or function becomes more common), we expect predictive models to continue to grow and evolve, as they work to meet a goal of optimized predictive accuracy.

7. Conclusion

Predictive approaches for the transition from acute to chronic LBP pain need to improve to meet practice and research priorities from clinical guidelines19,73 and the Federal Pain Research Strategy (United States).31 This review presented a predictive framework that improves upon previous approaches by standardizing predictor domains and encouraging use of multiple outcome measures to represent chronic LBP. There is potential that this predictive framework could lead to improvements in predictive accuracy that has not occurred naturally over time. However, empirical testing of this framework is necessary to determine whether it actually improves predictive accuracy, and whether advanced analytical approaches outperform traditional statistical approaches.


The authors have no conflicts of interest to declare.

All authors contributed substantially to the manuscript, including a review of the final version before being submitted for peer review.

Some of this content was presented by S.Z. George and T.A. Lentz at the 2017 North Carolina Physical Therapy Association Annual Meeting.

The Duke Clinical Research Institute's Communication team assisted with the graphic design of Figure 1.


[1]. Allen LA, McIlvennan CK, Thompson JS, Dunlay SM, LaRue SJ, Lewis EF, Patel CB, Blue L, Fairclough DL, Leister EC, Glasgow RE, Cleveland JC Jr, Phillips C, Baldridge V, Walsh MN, Matlock DD. Effectiveness of an intervention supporting shared decision making for destination therapy left ventricular assist device: the DECIDE-LVAD randomized clinical trial. JAMA Intern Med 2018;178:520–9.
[2]. Baliki MN, Petre B, Torbey S, Herrmann KM, Huang L, Schnitzer TJ, Fields HL, Apkarian AV. Corticostriatal functional connectivity predicts transition to chronic back pain. Nat Neurosci 2012;15:1117–19.
[3]. Beneciuk JM, Bishop MD, Fritz JM, Robinson ME, Asal NR, Nisenzon AN, George SZ. The STarT back screening tool and individual psychological measures: evaluation of prognostic capabilities for low back pain clinical outcomes in outpatient physical therapy settings. Phys Ther 2013;93:321–33.
[4]. Beneciuk JM, Fritz JM, George SZ. The STarT Back Screening Tool for prediction of 6-month clinical outcomes: relevance of change patterns in outpatient physical therapy settings. J Orthop Sports Phys Ther 2014;44:656–64.
[5]. Beneciuk JM, Hill JC, Campbell P, Afolabi E, George SZ, Dunn KM, Foster NE. Identifying treatment effect modifiers in the STarT back trial: a secondary analysis. J Pain 2017;18:54–65.
[6]. Beneciuk JM, Lentz TA, He Y, Wu SS, George SZ. Prediction of persistent musculoskeletal pain at 12 months: a secondary analysis of the optimal screening for prediction of referral and outcome (OSPRO) validation cohort study. Phys Ther 2018;98:290–301.
[7]. Bousema EJ, Verbunt JA, Seelen HA, Vlaeyen JW, Knottnerus JA. Disuse and physical deconditioning in the first year after the onset of back pain. PAIN 2007;130:279–86.
[8]. Burton AK, Tillotson KM, Main CJ, Hollis S. Psychosocial predictors of outcome in acute and subchronic low back trouble. Spine 1995;20:722–8.
[9]. Burton AK, McClune TD, Clarke RD, Main CJ. Long-term follow-up of patients with low back pain attending for manipulative care: outcomes and predictors. Man Ther 2004;9:30–5.
[10]. Campbell P, Foster NE, Thomas E, Dunn KM. Prognostic indicators of low back pain in primary care: five-year prospective study. J Pain 2013;14:873–83.
[11]. Cannistraci CV, Ravasi T, Montevecchi FM, Ideker T, Alessio M. Nonlinear dimension reduction and clustering by minimum curvilinearity unfold neuropathic pain and tissue embryological classes. Bioinformatics 2010;26:i531–539.
[12]. Cherkin DC, Deyo RA, Street JH, Barlow W. Predicting poor outcomes for back pain seen in primary care using patients' own criteria. Spine 1996;21:2900–7.
[13]. Chou R, Deyo R, Friedly J, Skelly A, Hashimoto R, Weimer M, Fu R, Dana T, Kraegel P, Griffin J, Grusing S, Brodt ED. Nonpharmacologic therapies for low back pain: a systematic review for an American College of Physicians Clinical Practice Guideline. Ann Intern Med 2017;166:493–505.
[14]. Christodoulou E, Ma J, Collins GS, Steyerberg EW, Verbakel JY, Van Calster B. A systematic review shows no performance benefit of machine learning over logistic regression for clinical prediction models. J Clin Epidemiol 2019;110:12–22.
[15]. Collins GS, Reitsma JB, Altman DG, Moons KG. Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD): the TRIPOD statement. Ann Intern Med 2015;162:55–63.
[16]. Corbett DB, Simon CB, Manini TM, George SZ, Riley JL III, Fillingim RB. Movement-evoked pain: transforming the way we understand and measure pain. PAIN 2019;160:757–61.
[17]. Deyo RA, Dworkin SF, Amtmann D, Andersson G, Borenstein D, Carragee E, Carrino J, Chou R, Cook K, DeLitto A, Goertz C, Khalsa P, Loeser J, Mackey S, Panagis J, Rainville J, Tosteson T, Turk D, Von Korff M, Weiner DK. Report of the NIH Task Force on research standards for chronic low back pain. J Pain 2014;15:569–85.
[18]. Dionne CE, Koepsell TD, Von Korff M, Deyo RA, Barlow WE, Checkoway H. Predicting long-term functional limitations among back pain patients in primary care settings. J Clin Epidemiol 1997;50:31–43.
[19]. Dowell D, Haegerich TM, Chou R. CDC guideline for prescribing opioids for chronic pain—United States, 2016. JAMA 2016;315:1624–45.
[20]. Drost FH, Rannikko A, Valdagni R, Pickles T, Kakehi Y, Remmers S, van der Poel HG, Bangma CH, Roobol MJ. Can active surveillance really reduce the harms of overdiagnosing prostate cancer? A reflection of real life clinical practice in the PRIAS study. Transl Androl Urol 2018;7:98–105.
[21]. Dunn KM, Croft PR. Repeat assessment improves the prediction of prognosis in patients with low back pain in primary care. PAIN 2006;126:10–15.
[22]. Epping-Jordan JE, Wahlgren DR, Williams RA, Pruitt SD, Slater MA, Patterson TL, Grant I, Webster JS, Atkinson JH. Transition to chronic pain in men with low back pain: predictive relationships among pain intensity, disability, and depressive symptoms. Health Psychol 1998;17:421–7.
[23]. Felicio DC, Diz JBM, Pereira DS, Queiroz BZ, Silva JP, Moreira BS, Oliveira VC, Pereira LSM. Handgrip strength is associated with, but poorly predicts, disability in older women with acute low back pain: a 12-month follow-up study. Maturitas 2017;104:19–23.
[24]. Fischer B, Gooch J, Goldman B, Kurdyak P, Rehm J. Non-medical prescription opioid use, prescription opioid-related harms and public health in Canada: an update 5 years later. Can J Public Health 2014;105:e146–149.
[25]. Fodeh SJ, Finch D, Bouayad L, Luther SL, Ling H, Kerns RD, Brandt C. Classifying clinical notes with pain assessment using machine learning. Med Biol Eng Comput 2018;56:1285–92.
[26]. Foster NE, Thomas E, Bishop A, Dunn KM, Main CJ. Distinctiveness of psychological obstacles to recovery in low back pain patients in primary care. PAIN 2010;148:398–406.
[27]. Freburger JK, Holmes GM, Agans RP, Jackman AM, Darter JD, Wallace AS, Castel LD, Kalsbeek WD, Carey TS. The rising prevalence of chronic low back pain. Arch Intern Med 2009;169:251–8.
[28]. Fritz JM, Brennan GP, Hunter SJ, Magel JS. Initial management decisions after a new consultation for low back pain: implications of the usage of physical therapy for subsequent health care costs and utilization. Arch Phys Med Rehabil 2013;94:808–16.
[29]. Fritz JM, Brennan GP, Hunter SJ. Physical therapy or advanced imaging as first management strategy following a new consultation for low back pain in primary care: associations with future health care utilization and charges. Health Serv Res 2015;50:1927–40.
[30]. Gatchel RJ, Polatin PB, Mayer TG. The dominant role of psychosocial risk factors in the development of chronic low back pain disability. Spine 1995;20:2702–9.
[31]. Gatchel RJ, Reuben DB, Dagenais S, Turk DC, Chou R, Hershey A, Hicks G, Licciardone JC, Horn SD. Research agenda for the prevention of pain and its impact: report of the work group on the prevention of acute and chronic pain of the federal pain research strategy. J Pain 2018;19:837–51.
[32]. George SZ, Beneciuk JM, Lentz TA, Wu SS, Dai Y, Bialosky JE, Zeppieri G Jr. Optimal screening for prediction of referral and outcome (OSPRO) for musculoskeletal pain conditions: results from the validation cohort. J Orthop Sports Phys Ther 2018;48:460–75.
[33]. Gheldof EL, Vinck J, Vlaeyen JW, Hidding A, Crombez G. Development of and recovery from short- and long-term low back pain in occupational settings: a prospective cohort study. Eur J Pain 2007;11:841–54.
[34]. Global, regional, and national disability-adjusted life-years (DALYs) for 315 diseases and injuries and healthy life expectancy (HALE), 1990-2015: a systematic analysis for the Global Burden of Disease Study 2015. Lancet 2016;388:1603–58.
[35]. Global, regional, and national incidence, prevalence, and years lived with disability for 310 diseases and injuries, 1990-2015: a systematic analysis for the Global Burden of Disease Study 2015. Lancet 2016;388:1545–602.
[36]. Goldstein BA, Navar AM, Carter RE. Moving beyond regression techniques in cardiovascular risk prediction: applying machine learning to address analytic challenges. Eur Heart J 2017;38:1805–14.
[37]. Goldstein BA, Pomann GM, Winkelmayer WC, Pencina MJ. A comparison of risk prediction methods using repeated observations: an application to electronic health records for hemodialysis. Stat Med 2017;36:2750–63.
[38]. Grant MJ, Booth A. A typology of reviews: an analysis of 14 review types and associated methodologies. Health Info Libr J 2009;26:91–108.
[39]. Grotle M, Brox JI, Glomsrod B, Lonn JH, Vollestad NK. Prognostic factors in first-time care seekers due to acute low back pain. Eur J Pain 2007;11:290–8.
[40]. Haldorsen EM, Indahl A, Ursin H. Patients with low back pain not returning to work. A 12-month follow-up study. Spine 1998;23:1202–7; discussion 1208.
[41]. Han B, Compton WM, Jones CM, Cai R. Nonmedical prescription opioid use and use disorders among adults aged 18 through 64 Years in the United States, 2003-2013. JAMA 2015;314:1468–78.
[42]. Hayden JA, Dunn KM, van der Windt DA, Shaw WS. What is the prognosis of back pain? Best Pract Res Clin Rheumatol 2010;24:167–79.
[43]. Henschke N, Maher CG, Refshauge KM, Herbert RD, Cumming RG, Bleasel J, York J, Das A, McAuley JH. Prognosis in patients with recent onset low back pain in Australian primary care: inception cohort study. BMJ 2008;337:a171.
[44]. Heus P, Damen J, Pajouheshnia R, Scholten R, Reitsma JB, Collins GS, Altman DG, Moons KGM, Hooft L. Poor reporting of multivariable prediction model studies: towards a targeted implementation strategy of the TRIPOD statement. BMC Med 2018;16:120.
[45]. Institute of Medicine Committee on Advancing Pain Research C, Education. The National Academies Collection: reports funded by National Institutes of Health. Relieving pain in America: a blueprint for transforming prevention, care, education, and research. Washington: National Academies Press (US) National Academy of Sciences, 2011.
[46]. Jang EH, Park BJ, Park MS, Kim SH, Sohn JH. Analysis of physiological signals for recognition of boredom, pain, and surprise emotions. J Physiol Anthropol 2015;34:25.
[47]. Jerez JM, Molina I, Garcia-Laencina PJ, Alba E, Ribelles N, Martin M, Franco L. Missing data imputation using statistical and machine learning methods in a real breast cancer problem. Artif Intell Med 2010;50:105–15.
[48]. Karran EL, McAuley JH, Traeger AC, Hillier SL, Grabherr L, Russek LN, Moseley GL. Can screening instruments accurately determine poor outcome risk in adults with recent onset low back pain? A systematic review and meta-analysis. BMC Med 2017;15:13.
[49]. Karran EL, Traeger AC, McAuley JH, Hillier SL, Yau YH, Moseley GL. The value of prognostic screening for patients with low back pain in secondary care. J Pain 2017;18:673–86.
[50]. Katzan IL, Thompson NR, George SZ, Passek S, Frost F, Stilphen M. The use of STarT back screening tool to predict functional disability outcomes in patients receiving physical therapy for low back pain. Spine J 2018;19:645–54.
[51]. Kazis LE, Ameli O, Rothendler J, Garrity B, Cabral H, McDonough C, Carey K, Stein M, Sanghavi D, Elton D, Fritz J, Saper R. Observational retrospective study of the association of initial healthcare provider for new-onset low back pain with early and long-term opioid use. BMJ Open 2019;9:e028633.
[52]. Kennedy-Martin T, Curtis S, Faries D, Robinson S, Johnston J. A literature review on the representativeness of randomized controlled trial samples and implications for the external validity of trial results. Trials 2015;16:495.
[53]. Klenerman L, Slade PD, Stanley IM, Pennie B, Reilly JP, Atchison LE, Troup JD, Rose MJ. The prediction of chronicity in patients with an acute attack of low back pain in a general practice setting. Spine 1995;20:478–84.
[54]. Koleck M, Mazaux JM, Rascle N, Bruchon-Schweitzer M. Psycho-social factors and coping strategies as predictors of chronic evolution and quality of life in patients with low back pain: a prospective study. Eur J Pain 2006;10:1–11.
[55]. Kongsted A, Vach W, Axo M, Bech RN, Hestbaek L. Expectation of recovery from low back pain: a longitudinal cohort study investigating patient characteristics related to expectations and the association between expectations and 3-month outcome. Spine 2014;39:81–90.
[56]. Law RK, Lee EW, Law SW, Chan BK, Chen PP, Szeto GP. The predictive validity of OMPQ on the rehabilitation outcomes for patients with acute and subacute non-specific LBP in a Chinese population. J Occup Rehabil 2013;23:361–70.
[57]. Lee H, Hubscher M, Moseley GL, Kamper SJ, Traeger AC, Mansell G, McAuley JH. How does pain lead to disability? A systematic review and meta-analysis of mediation studies in people with back and neck pain. PAIN 2015;156:988–97.
[58]. Lentz TA, Beneciuk JM, Bialosky JE, Zeppieri G Jr, Dai Y, Wu SS, George SZ. Development of a yellow flag assessment tool for orthopaedic physical therapists: results from the optimal screening for prediction of referral and outcome (OSPRO) cohort. J Orthop Sports Phys Ther 2016;46:327–43.
[59]. Lentz TA, Harman JS, Marlow NM, George SZ. Application of a value model for the prevention and management of chronic musculoskeletal pain by physical therapists. Phys Ther 2017;97:354–64.
[60]. Li K, Chan W, Doody RS, Quinn J, Luo S. Prediction of Conversion to Alzheimer's disease with longitudinal measures and time-to-event data. J Alzheimers Dis 2017;58:361–71.
[61]. Li K, Furr-Stimming E, Paulsen JS, Luo S. Dynamic prediction of motor diagnosis in Huntington's disease using a joint modeling approach. J Huntingtons Dis 2017;6:127–37.
[62]. Li X, Andersen KM, Chang HY, Curtis JR, Alexander GC. Comparative risk of serious infections among real-world users of biologics for psoriasis or psoriatic arthritis. Ann Rheum Dis 2019. doi: 10.1136/annrheumdis-2019-216102. [Epub ahead of print].
[63]. Lotsch J, Ultsch A. Machine learning in pain research. PAIN 2018;159:623–30.
[64]. Machado GC, Ferreira PH, Maher CG, Latimer J, Steffens D, Koes BW, Li Q, Ferreira ML. Transient physical and psychosocial activities increase the risk of nonpersistent and persistent low back pain: a case-crossover study with 12 months follow-up. Spine J 2016;16:1445–52.
[65]. Manchikanti L, Helm S II, Fellows B, Janata JW, Pampati V, Grider JS, Boswell MV. Opioid epidemic in the United States. Pain Physician 2012;15(3 suppl):Es9–38.
[66]. Mansell G, Hill JC, Main C, Vowles KE, van der Windt D. Exploring what factors mediate treatment effect: example of the STarT back study high-risk intervention. J Pain 2016;17:1237–45.
[67]. Mansell G, Hill JC, Main CJ, Von Korff M, van der Windt D. Mediators of treatment effect in the back in action trial: using latent growth modeling to take change over time into account. Clin J Pain 2017;33:811–19.
[68]. Mansour AR, Baliki MN, Huang L, Torbey S, Herrmann KM, Schnitzer TJ, Apkarian AV. Brain white matter structural properties predict transition to chronic pain. PAIN 2013;154:2160–8.
[69]. Matlock DD, McGuire WC, Magid M, Allen L. Decision making in advanced heart failure: bench, bedside, practice, and policy. Heart Fail Rev 2017;22:559–64.
[70]. Nicholas M, Vlaeyen JWS, Rief W, Barke A, Aziz Q, Benoliel R, Cohen M, Evers S, Giamberardino MA, Goebel A, Korwisi B, Perrot S, Svensson P, Wang SJ, Treede RD. The IASP classification of chronic pain for ICD-11: chronic primary pain. PAIN 2019;160:28–37.
[71]. Perrot S, Cohen M, Barke A, Korwisi B, Rief W, Treede RD. The IASP classification of chronic pain for ICD-11: chronic secondary musculoskeletal pain. PAIN 2019;160:77–82.
[72]. Phelan M, Bhavsar NA, Goldstein BA. Illustrating informed presence bias in electronic health records data: how patient interactions with a health system can impact inference. EGEMS (Wash DC) 2017;5:22.
[73]. Qaseem A, Wilt TJ, McLean RM, Forciea MA. Noninvasive treatments for acute, subacute, and chronic low back pain: a Clinical Practice Guideline from the American College of Physicians. Ann Intern Med 2017;166:514–530.
[74]. Rundell SD, Patel KV, Krook MA, Heagerty PJ, Suri P, Friedly JL, Turner JA, Deyo RA, Bauer Z, Nerenz DR, Avins AL, Nedeljkovic SS, Jarvik JG. Multisite pain is associated with long-term patient-reported outcomes in older adults with persistent back pain. Pain Med 2019. doi: 10.1093/pm/pny270. [Epub ahead of print].
[75]. Schiottz-Christensen B, Nielsen GL, Hansen VK, Schodt T, Sorensen HT, Olesen F. Long-term prognosis of acute low back pain in patients seen in general practice: a 1-year prospective follow-up study. Fam Pract 1999;16:223–32.
[76]. Schug SA, Lavand'homme P, Barke A, Korwisi B, Rief W, Treede RD. The IASP classification of chronic pain for ICD-11: chronic postsurgical or posttraumatic pain. PAIN 2019;160:45–52.
[77]. Sieben JM, Vlaeyen JW, Portegijs PJ, Verbunt JA, van Riet-Rutgers S, Kester AD, Von Korff M, Arntz A, Knottnerus JA. A longitudinal study on the predictive validity of the fear-avoidance model in low back pain. PAIN 2005;117:162–70.
[78]. Skelly AC, Chou R, Dettori JR, Turner JA, Friedly JL, Rundell SD, Fu R, Brodt ED, Wasson N, Winter C, Ferguson AJR. AHRQ comparative effectiveness reviews. Noninvasive nonpharmacological treatment for chronic pain: a systematic review. Rockville: Agency for Healthcare Research and Quality (US), 2018.
[79]. Traeger AC, Henschke N, Hubscher M, Williams CM, Kamper SJ, Maher CG, Moseley GL, McAuley JH. Estimating the risk of chronic pain: development and validation of a prognostic model (PICKUP) for patients with acute low back pain. PLoS Med 2016;13:e1002019.
[80]. Treede RD, Rief W, Barke A, Aziz Q, Bennett MI, Benoliel R, Cohen M, Evers S, Finnerup NB, First MB, Giamberardino MA, Kaasa S, Korwisi B, Kosek E, Lavand'homme P, Nicholas M, Perrot S, Scholz J, Schug S, Smith BH, Svensson P, Vlaeyen JWS, Wang SJ. Chronic pain as a symptom or a disease: the IASP classification of chronic pain for the International Classification of Diseases (ICD-11). PAIN 2019;160:19–27.
[81]. Truchon M, Schmouth ME, Cote D, Fillion L, Rossignol M, Durand MJ. Absenteeism screening questionnaire (ASQ): a new tool for predicting long-term absenteeism among workers with low back pain. J Occup Rehabil 2012;22:27–50.
[82]. Wang LE, Shaw PA, Mathelier HM, Kimmel SE, French B. Evaluating risk-prediction models using data from electronic health records. Ann Appl Stat 2016;10:286–304.
[83]. Wideman TH, Hill JC, Main CJ, Lewis M, Sullivan MJ, Hay EM. Comparing the responsiveness of a brief, multidimensional risk screening tool for back pain to its unidimensional reference standards: the whole is greater than the sum of its parts. PAIN 2012;153:2182–91.
[84]. Zhang G, Liang Z, Yin J, Fu W, Li GZ. A similarity based learning framework for interim analysis of outcome prediction of acupuncture for neck pain. Int J Data Min Bioinform 2013;8:381–95.

Chronic pain; Outcome prediction; Pain research

Copyright © 2020 The Author(s). Published by Wolters Kluwer Health, Inc. on behalf of The International Association for the Study of Pain.