|Home | About | Journals | Submit | Contact Us | Français|
Primary care clinicians need to identify candidates for early interventions to prevent patients with acute pain from developing chronic pain.
We conducted a 2-year prospective cohort study of risk factors for the progression to chronic pain and developed and internally validated a clinical decision rule (CDR) that stratifies patients into low, medium and high-risk groups for chronic pain.
Prospective cohort study in primary care.
Patients with acute low back pain (LBP; ≤30 days duration)
Self-reported perceived non-recovery and chronic pain.
Patients were surveyed at baseline, 6 months and 2 years. We conducted bivariate and multivariate regression analyses of demographic, clinical and psychosocial variables for chronic pain outcomes, developed a CDR and assessed its performance by calculating the bootstrapped areas under the receiver operating characteristic curve (AUC) and likelihood ratios. This study was supported by NIH/NCCAM grants K23 AT002298, R21 AT004467, NIH/NCCAM K24 AT007827, the Research Evaluation and Allocation Committee (REAC) of the University of California San Francisco, and the Mount Zion Health Fund, San Francisco. The funding agencies played no role in design and conduct of the study; collection, management, analysis, and interpretation of the data; and preparation, review, or approval of the manuscript. The authors report no conflict of interests.
605 patients enrolled. 13% had chronic pain at 6 months, 19% at 2 years. An eight-item CDR was most parsimonious for classifying patients into three risk levels. Bootstrapped AUC was 0.76 (0.70–0.82) for the 6-month CDR. Each 10-point score increase (60-point range) was associated with an odds ratio of 11.1 (10.8–11.4) for developing chronic pain. Using a <5% probability of chronic pain as the cutoff for low risk and a >40% probability for high risk, likelihood ratios were 0.26 (0.14–0.48) and 4.4 (3.0–6.3) for these groups, respectively.
A CDR was developed that may help primary care clinicians classify patients with strictly defined acute LBP into low, moderate and high-risk groups for developing chronic pain and performed acceptably in 1,000 bootstrapped replications. Validation in a separate sample is needed.
Although most patients presenting with an episode of acute low back pain (LBP) in primary care will recover in six to eight weeks with or without medical intervention,1,2 those who subsequently develop chronic pain suffer considerably,3 often are difficult to treat, and account for most LBP-related health expenses.4 Primary care clinicians need decision support to identify candidates for early interventions for secondary prevention of chronic pain. Previous studies have identified risk factors for chronic pain, and have attempted to develop clinical decision rules for the primary care setting.5,6 The most important are the STarT-Back developed in the UK7,8 and the Chronic Pain Risk Screener (CPRS) developed in the US.9 The STarT-BACK and several instruments developed in Europe (Örebro Musculoskeletal Pain Screening Questionnaire (ÖMPSQ)10,11, Kiel Pain Inventory and Avoidance-Endurance Questionnaire, 12,13 and Heidelberger Kurz-Fragebogen (HKF).14) have not been evaluated in the US. Other limitations of the latter instruments are that they were not developed or validated in primary care patients and used delayed return-to-work as chronic pain outcomes, which only captures a subset of patients taking sick leave.
Both the STarT-BACK and CPRS have been well validated in patients shortly following an index visit at a primary care office.15 However, these index visit patients included patients with a wide range of LBP duration; less than half suffered from acute LBP. Because patients who suffer LBP for more than 3 months already have a much worse prognosis, instruments that work for this population may not perform as well in patients with acute LBP. Hence, clinicians need a tool that only addresses the prognosis of patients with truly acute LBP.1
We therefore conducted a prospective cohort study to investigate the prognosis of patients with strictly defined acute LBP16, and whether we can identify early risk factors that can help primary care clinicians determine a more accurate prognosis. If available such risk stratification would be feasible for primary care clinics and could potentially support physicians in treatment allocation decisions. We included questionnaire items representative of all risk factors known at the time of the cohort’s inception and set out to develop a novel clinical decision rule (CDR).
The Prognosis of Pain (POP) study was a 2-year longitudinal telephone survey of 18–70 year old members of Kaiser Permanente, Northern California, the largest integrated health plan in its region with 2.4 million adult members at the time. Acute LBP was defined as back pain between the rib cage and buttocks of less than one month that was severe enough to seek medical care and was not preceded by any other episodes of LBP in the past year. The 1-month criterion for acuteness of pain was chosen in part for pragmatic reasons, as we found that the time from scheduling a doctor’s visit to being seen might be more than two weeks from the date of first pain onset. Patients were included if they spoke English and had no fever, history of cancer, chronic inflammatory disease, previous spine surgery, fibromyalgia, chronic pain conditions, disabling psychiatric diseases, or ongoing prescriptions for narcotics prior to the LBP episode. Patients with sciatica (i.e. LBP radiating below the knee), were not excluded.
A computer program screened electronic medical records to identify patients seen the day before for LBP, and a written invitation was sent by mail to join the study. This invitation offered a $20 gift certificate and did not reveal the inclusion criterion of pain duration; it therefore prioritized minimization of false reporting over larger numbers of ineligible respondents. Respondents were interviewed over the phone at baseline and 6 months. For the 2-year follow-up, participants, when reached (maximum of 3 attempts), were given a choice between a phone interview and an internet-based survey using SurveyGizmo (http://www.surveygizmo.com).17 The study was approved by the Institutional Review Boards of the University of California, San Francisco and Kaiser Permanente. Two follow-up survey time points at 6 months and 2 years allowed us to determine consistency of predictors over time. The surveys were conducted between February 2008 and November 2010.
In addition to the typical demographic items (age, sex, ethnicity, foreign born, education and income) we asked about marital status, employment status, heavy or monotonous work, job satisfaction, and smoking. The following clinical parameters were assessed at baseline: duration of current episode; history of prior episodes; pain-free interval before current episode; pain location(s); sciatica; pain intensity by 11-point numeric rating scale (NRS) as average, worst, and most tolerable pain or average bothersomeness; McGill Pain Questionnaire;18 Roland-Morris Disability Questionnaire (RMDQ);19 and days on sick leave and of reduced daily activities. The complete 24-item ÖMPSQ,11 10-item HKF14 and the 4-item Perceived Stress Scale (PSS-4)20 were included. Additional psychological predictor variables were selected from validated instruments according to strong factor loadings and face validity (Table 1).
To avoid overextensive participant burden from lengthy questionnaires we limited the survey to selected items expected to perform reasonably well when a reduced item set was needed.21 In addition to the psychological items in ÖMPSQ and HKF, we included another fear-avoidance beliefs item from the Fear-Avoidance Beliefs Questionnaire (FABQ),22 another catastrophizing item from the Coping Strategies Questionnaire (CSQ)21 and two from the Pain Catastrophizing Scale (PCS)23. As additional coping style items, we included ignoring and positive distracting using single items from the CSQ, seeking instrumental and emotional support using four items from CPCI24–26 and KPI 27–29 and denial of stress using two items from Brief-COPE.30 The 2-item version of the self-efficacy for pain subscale21 from Arthritis Self-Efficacy Scale was also included. Anxiety and depression were assessed by multiple ÖMPSQ and HKF items, and positive affect by one item from the CES-D31 (Table 1).
No gold standard or international consensus exists regarding the outcome definition for chronic LBP in cohort studies. Following recommendations from expert LBP epidemiologists in the Netherlands,32 we applied a previously published primary outcome measure that combines a lack of perceived recovery (less than “much improved” on a 6-point Likert Perceived Recovery Scale)33 with current pain intensity of 3 or more on 0–10 Numeric Rating Scale.32,34 Its accuracy was assessed for this population sample in a prior study.34 In an exploratory fashion we also used a Grade 2 or higher chronic pain level according to the validated Graded Chronic Pain Scale (GCPS) by von Korff (this instrument yields a 4-grade-level chronic pain score as a function of pain intensity and pain-related disability for the past 6 months), but only for 2-year follow-up analyses, as it includes recall of the acute phase LBP at onset.9
All analyses were conducted using Stata.35 We proceeded in a series of analytic steps from bivariate to multivariate analysis of predictor variables, data and consensus-driven decisions for scoring and cut-off, and the calculation of areas under the receiver operating characteristic curve (AUC) and likelihood ratios for the resulting CDR.
Step 1 was raw bivariate logistic regression analysis of predictor variables for our primary outcome at 6 months and 2 years. We used individual items assessed at baseline as predictor variables. In addition, we compared the odds ratios for predictors of the primary 2-year outcome with those of the 2-year GCPS. We selected variables that were associated with our outcome at p <0.1 consistently at both time points for inclusion in Step 2.
Step 2 used the variables selected in Step 1 in multivariate logistic regression models and proceeded to stepwise backward elimination of the least supported variables. In addition, we analyzed the correlations between predictor variables and eliminated collinear redundant variables (r >.35) to arrive at the most parsimonious sets, separately for the 6-month and the 2-year predictions.
Step 3 used the beta-coefficients of each of the final multivariable regression models (one predicting 6-month outcomes, one predicting 2-year outcomes), multiplied by 10 and rounded to the closest half integer, to create two scoring rules. The performance of each scoring rule was assessed by bootstrapped analysis with 1,000 replications and calculation of the AUROCC and associated 95% confidence intervals (CIs).
Step 4 determined the number of patients who were positive or negative on our outcome measures for each point score. We then empirically identified reasonable cutoffs for low, moderate and high-risk groups based on discussion with content experts. A minimal risk group was defined as patients that have a less than 5% chance of developing chronic pain at 6 months and less than 10% at 2 years. Conversely, if a patient has a risk of 40% or more of developing chronic pain at either 6 months or 2 years, a clinician likely would consider closer follow-up office appointments and potentially more intensive early interventions to prevent chronic pain, and these were defined as a high-risk group. In the middle range, less intensive interventions may suffice until further follow-up assessments. By simple inspection of the table with the two patient groups being either positive or negative on the outcome for each score point, we determined the cut-off for the summary score and assessed the proportion of patients in each risk group.
Step 5: We calculated likelihood ratios with 95% confidence intervals (CI’s) for the classification into these three risk groups by our CDR, according to the formula by Simel and colleagues.36
The Prognosis of Pain (POP) study enrolled 605 eligible members of Kaiser Permanente, Northern California (KPNC) from February 2008 to March 2009 (Figure 1). This represents 25% of the 2,454 respondents to invitations mailed to 42,650 patients who were seen for any kind of LBP in clinics of the health plan during the twelve months of recruitment. Overall, 521 participants (86%) responded at 6 months and 443 (73%) at the 2-year follow-up. The average age was 50.5 (±12.6) years, 56% were female, 65% Caucasian-American, 18% foreign born, 61% had a college degree, and 59% were employed full-time (for further details, see 16) The sample represented the socioeconomic and ethnic diversity of primary care patients in the San Francisco Bay Area.37 These patients sought medical care for pain of considerable intensity (average in past week 5.6 ±1.8; 2.6 ±1.8 when most tolerable; 8.6 ±1.4 when worst; 11-point NRS), bothersomeness (6.5 ±2.3) and disability (mean Roland-Morris score 15.8 ±4.7). The median duration of pain at baseline interview was 14 days; 8% had been on sick leave; and 27% had some sciatic pain to below the knee during this episode, 10% at the time of the interview. The final sample included 510 patients with complete 6-month follow-up data and 443 patients with complete 2-year data. Using our primary combined outcome criterion,34 13% of the patients (95% confidence interval [CI], 10%–16%) experienced persistent or recurrent pain at 6 months and 19% (CI, 15%–22%) at 2 years after pain onset.16 Numerous patients who self-reported as much improved at 6 months felt worse at 2 years (details in 16). Participants lost to follow-up were slightly younger and included slightly more females but did not differ in those variables that were included in the CDR.
All variable had <2% missing responses and were used without substitution. The following 12 variables had odds ratios at significance levels of p<0.1 for our primary outcome at both time points, five were protective and seven were predictive of chronic pain. Protective were: completed college, ability to walk for 1 hour, ability to sleep tonight, coping by TV or music, and self-efficacy in ability to decrease pain; predictive were additional pain in upper back, higher level of least pain since onset, smoking, catastrophizing (2 items), expectancy of chronicity and the need to holding onto something when getting off the sofa. Additional variables that satisfied at least one of our outcome criteria and the GCPS were perceived stress, coping by ignoring, coping by prayer, belief that activity worsens pain, anxiety or tension, RMDQ items 2, 5, 18, 22, yoga at baseline, McGill overall pain intensity, worst pain since onset, sciatica since onset, African-American ethnicity, and being separated or widowed. No significant bivariate associations consistent across at least 2 outcome measures were found for age, sex, income level, born outside the US, duration of pain, sciatica at time of interview, average pain intensity since onset, pain level willing to tolerate, cut-down activity days, days in bed, days lost from work, retirement, job satisfaction, other RMDQ items, positive affect, enjoyment, positive thinking, depression, ability to do light work or household chores for 1 hour, heavy or monotonous work, multiple other pain-avoidance and catastrophizing items, coping by seeking a friend or talking with family member, staying active, detachment, reinterpretation, challenge appraisal, asking for instrumental or emotional support, other perceived stress or stress denial items.
After backward elimination, eight variables remained for the 6-month prediction model and eight slightly different variables for the 2-year model. They are listed in Table 2 and discussed below. When using only these eight variables for each model, the regression models explained 16% (6-month) and 10% (2-year) of the respective outcome variance. Using Grade 2 or higher of the GCPS as outcome (instead of our primary outcome combining perceived recovery with pain intensity) at 2 years provided similar results (not shown) for included parameters, AUC, and explained variance.
Table 3 shows the beta-coefficients and odds ratios for each model. Multiplying the beta coefficients by 10 and rounding to the closest half integer created the 6-month and 2-year scoring rules. This method gives differential weights to individual predictors according to their beta-coefficients in the multivariate model. Note that multipliers for dichotomous items are based on values of 0 or 1, whereas continuous variables are multiplied by values between 0 and 10. The 6-month scoring rule ranged from −25 to 34 points. Applied to our sample, the AUC was 0.78 (95% CI 0.72–0.84; Figure 1). Applying the bootstrap procedure for 1,000 replications, a 10-point increase in the 60-point score was associated with a 11.1 odds ratio (95% CI 10.8–11.4; p <0.001) for having chronic pain 6 months after baseline. After bootstrapping, the AUC was slightly lower: 0.76 (95% CI 0.70–0.82).
Using the same process for creating a scoring rule for the eight strongest predictor variables at the 2-year outcome, we obtained summary scores between −18 and 28.5. Applied to our sample at the 2-year follow-up, the AUC was 0.70 (95% CI 0.64–0.76; bootstrapped 0.69; 0.62–0.75). Applying the bootstrap procedure for 1,000 replications, a 10-point increase in the score was associated with a 11.1 odds ratio (95% CI 10.7–11.5; p <0.001) for having chronic pain 2 years after baseline.
After inspection of the outcomes table for each rule, we identified optimal score cut-offs for creating the three clinically-useful risk groups at 6 months and 2 years and assessed the proportion of patients in each risk group. The results are shown in Table 4.
Following a discussion among clinical colleagues, we assumed that a score with a predictive value of or near 5% would be a good cut-off for the lowest risk group, and that a 40% predictive value would be an appropriate cut-off for recommending further assessment and therapeutic measures. Applying these criteria to the 6-month prediction, score cutoffs were less than -4 for low-risk and above +7 for high-risk groups. The low risk group included 47% of all patients, the mid-range risk group 38%, and the high-risk group 15%. The resulting proportions of chronic pain patients in the three risk groups were 3.8%, 14.1% and 39.7%, respectively.
Applying the 2-year decision rule in the 2-year follow-up dataset, we obtained scores between −18 and 28.5 and found that relatively low scores had a higher than 5% risk of developing chronic pain. We, therefore, chose a 10% cutoff for the low-risk group, maintained the 40% cutoff for the high-risk classification, and thus classified 49% as low risk at a score of ≤1, 36% as mid-level risk at scores of >1 and <9, and 15% as high risk with scores of ≥9.
Likelihood ratios for correctly classifying patient into low, medium, and high-risk categories were 0.26 (95% CI 0.14–0.48), 1.08 (0.79–1.5), and 4.35 (3.0–6.3) at six months and 0.50 (0.34–0.72), 1.12 (0.82–1.52), and 3.14 (2.06–4.78), respectively, at 2 years (Table 4).
To the best of our knowledge, this is the first attempt to develop a clinical decision rule (CDR) for the prediction of chronic LBP among patients with strictly defined acute LBP of less than four weeks duration in the US. A variety of methods exists for developing such rules.38,39 This CDR was developed using multivariable logistic regression to help primary care clinicians decide whether a patient who presents with a new episode of non-specific LBP with or without sciatica is at risk of developing chronic pain and may warrant closer follow-up and potentially a more intensive therapeutic intervention. The CDR is limited to patients who had no LBP in the previous year and never had spine surgery. Prior CDRs were developed and validated in patients with LBP of any duration; a majority of these had pain for more than 3 months and already had a higher pre-test probability for persistent pain. When a CDR constructed of items that were identical or highly similar to the 9-item STarT-Back from the UK was applied in our sample, its performance was found unsatisfactory in patients with truly acute LBP.40
The variables included in the new CDRs reflect risk factors that have been found in prior studies: College education was protective and the only significant demographic predictor.41 Pain spreading to the upper back was a consistent clinical risk factor. Sciatica42 and difficulty sleeping10 predicted poor outcomes at 6 months, the inability to walk for 1 hour10 poor outcomes at 2 years. At both follow-ups, a coping style of watching TV or listening to music43 was protective, whereas catastrophizing44 and coping with pain by ignoring13 were psychological risk factors. Five of eight predictor variables were identical for outcomes at our 2 follow-up time points, three were different. The expectancy of pain to persist was maladaptive at 6 months,45,46 while a low willingness to tolerate pain14 and perceived stress41,42 increased risk at 2 years. We do not have an explanation for the difference.
The observed likelihood ratios for the 6-month CDR of 0.26 (95% CI 0.14–0.48) for the low-risk and 4.4 (95% CI 3.0–6.3) for the high-risk classifications are moderately accurate.47 The rule is likely to be clinically useful, as almost half of all patients fell into a low-risk group that was unlikely to develop chronic pain at 6 months. Approximately 15% of patients were classified as high-risk and may warrant more intensive interventions. The remaining 38% of mid-level risk patients, assessed at an average of 2 weeks (range 2–30 days) after pain onset, had a mean risk of 14% for developing chronic pain. They may warrant closer oversight by their primary care clinician than the low-risk group, but it may be justified to suggest waiting a bit longer before prescribing more intense and costly interventions.
We previously reported for this cohort that due to the recurring course of cLBP, individuals with persistent pain at 6 months were not identical to those at 2 years, and that the proportion of persistent pain patients had increased between the two follow-up time points.16 Creating a 2-year decision rule with the 2-year follow-up made it impossible to use the 5% criterion as cutoff for low risk classification. We therefore used a 10% criterion but maintained the 40% risk for the high-risk classification cutoff. The results show that the prediction of the longer-term outcome is challenging in patients with strictly defined acute LBP. Longer-term predictions over years appear to be less precise than the prediction for 6 months, which is not surprising as the outcome is much further into the future than for the 6-month CDR.
A scoring method that assigns weights to individual predictor variables may best be used by a programmed risk classification calculator rather than by hand but it increases the precision of the prediction.38 Whereas the 6-month CDR maybe most useful for the primary care clinical practice, the 2-year CDR maybe useful for long-term clinical research. Five of the predictive items for the 6-month rule are identical with those for the 2-year rule. For validation of the rules in a separate sample and in particular for clinical application we would recommend assessing all eleven items, and then applying the relevant 8 items for 6-month predictions and the slightly different set of 8 items for a 2-year prediction.
The main limitation of our study is that we have not assessed the decision rules’ performance in a separate validation sample. The observed variance in these predictor item scores among study participants with acute LBP early into their episode is rather large and reduces their predictive power. Moreover, as shown for this cohort, a high recurrence rate leads to different individuals having persistent pain at different time points.16 This variance creates a challenge for creating a rule that performs strongly with patients where it is most needed, early in the course of a new episode of LBP.
A second limitation is that we included only questionnaire items that were known to be potentially predictive at the time of the study’s implementation. Somatization,48 reduced levels of body awareness,49 and potentially many others may be further parameters of predictive value and were not included in our questionnaire. However, we included a wide range of demographic, clinical and psychological predictor variables carefully chosen according to the best knowledge of the time.
A third limitation is that we relied on diagnostic codes from electronic medical records created by primary care providers and patient self-report. It is possible that clinical findings that indicate a more severe baseline condition, such as positive signs for spinal nerve compression or spinal claudication from spinal stenosis, can be identified as important risk factors for chronic pain at the very first onset of LBP by clinical exam and imaging studies. However, current clinical guidelines do not recommend imaging in the first weeks after new-onset LBP in patients who most likely would not need an immediate referral to a spine surgeon. Future studies would benefit from clinical exams at study entrance.
A fourth potential limitation is that the study population did not include uninsured patients; only 3% reported annual household incomes below $25,000. However, income level was not predictive of the outcome.
Despite these limitations, we conclude that our study provides a clinical decision rule that is urgently needed for one of the most frequent and most costly conditions in primary care.50 It contains 8 items for the 6-month and 8 items for the 2-year risk classification (5 are common to both) into 3 levels of risk for developing chronic pain in patients presenting in primary care with a new-onset episode of strictly defined acute low back pain. The next step is to prospectively validate this tool in an independent population.
This study was supported by NIH/NCCAM grants K23 AT002298 (Mehling), R21 AT004467 (Mehling), NIH/NCCAM K24 AT007827 (Hecht), the Research Evaluation and Allocation Committee (REAC) of the University of California San Francisco, and the Mount Zion Health Fund, San Francisco. The funding agencies played no role in design and conduct of the study; collection, management, analysis, and interpretation of the data; and preparation, review, or approval of the manuscript.
We would like to acknowledge the cooperation of the many patients who took time to answer our surveys in spite of being in pain and thank them. We would like to thank Timothy S. Carey, MD, MPH, for his contributions to study design, item selection and manuscript review, Alice Pressman for creating the computer program for the electronic medical records, Viranjini Gopisetty and Elizabeth Bartmess for project management, Pete Bogdanos for research assistance and the many volunteers who helped with the interviews.
Publisher's Disclaimer: This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final citable form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.