We conducted a retrospective matched cohort study of new episodes of uncomplicated non-specific low back pain. We used data from the U.S. Department of Veterans Affairs (VA) Corporate Data Warehouse, a national repository of extracts of electronic health records. A summary of the databases used can be found in Appendix 1. The study protocol was approved by the Administrative Panels for the Protection of Human Subjects of Stanford University.
Individuals were eligible for inclusion if they had a visit for uncomplicated low back pain in a VA primary care clinic between June 1, 2015, and June 30, 2016. Uncomplicated low back pain visits were characterized with a diagnosis of non-specific pain of the lower back: lumbar sprain, strain, spondylosis, or disk degeneration. Visits characterized exclusively by diagnosis codes for generalized back pain were excluded: unspecified backache (ICD-9 code 724.5), other dorsalgia, or unspecified dorsalgia (ICD-10 codes M5489 and M549). Visits were excluded if they were also characterized with a specific diagnosis of myelopathy, radiculopathy, sciatica, injury to the coccyx, spinal stenosis, or disk herniation.
Visits were excluded if there was a “red flag” condition, as defined by the Medicare performance measure for appropriateness of lumbar spine MRI (LS-MRI),16 present in VA hospital and outpatient data. These conditions included trauma in the prior 45 days; lumbar spine surgery in the past 90 days; cancer, neurologic impairment, infections, or injection drug use in the past year; or autoimmune, inflammatory, conditions, neoplastic abnormalities, radiation therapy, or congenital malformation in the last 5 years (i.e., a lookback period beginning June 1, 2010).
We considered a visit the start of a new episode if there was no VA-provided or VA-sponsored care in the prior 6 months with a diagnosis involving pain to the thoracic, lumbar, sacral, or sacrococcygeal spine. When patients had more than one episode in the study year, the first episode was selected for study. Less than 5% of patients had multiple episodes in the study year.
Early MRI Scan
Scans were identified using Current Procedure Terminology codes for LS-MRI (72148, 72158). Scans provided within 6 weeks of the index visit were deemed “early,” which is consistent with guidelines for treatment of low back pain.6
Patients with early MRI were matched to patients without an early MRI using coarsened exact matching (CEM). CEM assigns individuals from each group to strata within which individuals share the same coarsened values.17,18,19 Individuals exposed to an early MRI were assigned a value of 1. Unexposed individuals were assigned a weight that accounts for the number of exposed and unexposed observations within that strata and the total number of exposed and unexposed observations that were matched. The matching was based on patient baseline characteristics: demographics (age and sex); index visit pain score, having an assigned primary care provider; residing ≥ 40 miles from the nearest VA facility; and a history of a pain condition, a mental health condition, or a prescription for opioids in the previous 12 months.
Outcomes were measured from 43 to 365 days post-index visit, and costs were also measured for the first 42 days. We excluded the 42-day post-index visit, as this was the time period used to assess whether an individual received an early MRI scan.
Probability of Lumbar Surgery
Lumbar-related spinal surgeries included surgical procedures conducted on the lumbar, lumbosacral, and thoracolumbar regions in an inpatient or outpatient setting. Surgeries provided in the hospital were identified with ICD-9 or ICD-10 procedure codes. Surgeries provided in outpatient settings were identified using CPT codes when the encounter included a diagnosis code for low back pain.
Opioid prescriptions were defined as long-acting opioids, short-acting opioids, and tramadol, with doses converted to morphine equivalents using Centers for Disease Control conversion factors.20
During most encounters, VA patients are asked to report their pain on a scale from 0 to 10. We used the first value recorded on any given date. Values were found for 95.0% of index visits. Follow-up scores recorded 43–365 days after the index visit were found for 77.4% of episodes. We used the last score reported in this period. The mean time to this follow-up was 272 days after the index visit.
Cost was tallied for the 6-week period of exposure to early MRI as well as the follow-up period from weeks 7 to 52. Sub-totals were calculated for back-related care and care unrelated to the back. Sub-totals were also calculated for outpatient and inpatient care. Costs included both VA-provided care and care purchased by VA. The costs of long-term care and homeless domiciliary services were excluded. Hospital stays were considered back-related if the primary diagnosis was for low back pain; outpatient visits were considered back-related if any of the assigned diagnoses was back-related (no diagnosis is considered primary in outpatient records).
Baseline characteristics of patients with early MRI were compared with those of patients without early MRI by calculating the standardized difference of the means using pooled variance. Both a raw standardized difference and a CEM-weighted standardized difference were compared.
The association between early MRI and each outcome was determined by multivariable regressions adjusted with CEM weights and controlling for baseline covariates: demographics (age and sex); primary care provider characteristics (whether the patient was assigned to a provider, whether the assigned provider was seen on the index visit, provider’s panel size, and type of clinic visited at the index visit); case-mix (pain score at the index visit, history of opioid use in the prior year, time between the index visit and the last low back pain visit, total VA costs in the prior year, and total morphine equivalents in the prior year); and indicators for chronic conditions, mental health diagnoses, and pain conditions.
Logistic regression was estimated for the lumbar surgery and prescription opioid outcomes. The adjusted absolute proportion of each outcome was found as the mean of the predicted values of the logistic regression. Log-binomial and Poisson regressions were used to estimate relative risks.21 Generalized linear model (GLM) regressions were specified for continuous variables. Box-Cox regression was used to identify the link function and a modified Park test was used to identify the appropriate distribution.22 GLM with a log link and gamma distribution accommodates the skewed distribution and heteroscedastic errors of costs.23, 24 These specifications were used for all cost outcomes except acute cost in the first 42 days of the episode, for the appropriate distributional assumption was inverse normal. Absolute difference in means between those with an early scan and the comparison group was estimated as the difference in the mean antilog of the fitted values, with standard errors estimated using the delta method.25 The ratio of costs of the early MRI group relative to the comparison group was estimated by adding one to the regression parameter for the indicator of early exposure to MRI. The parameter for an indicator variable in a log model represents the proportional change, so adding one to this value is the ratio of the fitted value of observations with the indicator (those with early MRI) to the fitted value of observations without the indicator.
Two-part models were used to accommodate the large number of zero values for the last recorded pain score and quantity of opioid use outcomes.26 Logistic regression modeled the probability of a non-zero value, while GLM regressions were used for the second part. An identity link and inverse normal distribution were specified for the last recorded pain score. A log link and Poisson distribution were specified for the quantity of opioids. We estimated the distribution of the standard errors using 1000 bootstrap replicates.26 Analyses were conducted using Stata version 15.27