Article Text
PDF
Abstract
Objectives To (1) estimate the proportion of patients seeking care for low back pain (LBP) who are imaged and (2) explore trends in the proportion of patients who received diagnostic imaging over time. We also examined the effect of study-level factors on estimates of imaging proportion.
Data sources Electronic searches of MEDLINE, Embase and CINAHL databases from January 1995 to December 2017.
Eligibility criteria for selecting studies Observational designs and controlled trials that reported imaging for patients presenting to primary care or emergency care for LBP. We assessed study quality and calculated pooled proportions by care setting and imaging type, with strength of evidence assessed using the GRADE system.
Results 45 studies were included. They represented 19 451 749 consultations for LBP that had resulted in 4 343 919 imaging requests/events over 21 years. Primary care: moderate quality evidence that simple imaging proportion was 16.3% (95% CI 12.6% to 21.1%) and complex imaging was 9.2% (95% CI 6.2% to 13.5%). For any imaging, the pooled proportion was 24.8% (95% CI 19.3%to 31.1%). Emergency care: moderate quality evidence that simple imaging proportion was 26.1% (95% CI 18.2% to 35.8%) and high-quality evidence that complex imaging proportion was 8.2% (95% CI 4.4% to 15.6%). For any imaging, the pooled proportion was 35.6% (95% CI 29.8% to 41.8%). Complex imaging increased from 7.4% (95% CI 5.7% to 9.6%) for imaging requested in 1995 to 11.4% (95% CI 9.6% to 13.5%) in 2015 (relative increase of 53.5%). Between-study variability in imaging proportions was only partially explained by study-level characteristics; there were insufficient data to comment on some prespecified study-level factors.
Summary/conclusion One in four patients who presented to primary care with LBP received imaging as did one in three who presented to the emergency department. The rate of complex imaging appears to have increased over 21 years despite guideline advice and education campaigns.
Trial registration number CRD42016041987.
- lower back
- diagnosis
- MRI
- radiography
- primary care
Statistics from Altmetric.com
Introduction
Low back pain (LBP) is a major contributor to disease burden worldwide,1 with higher prevalence in athletes than in the general population.2 The majority of LBP has no known pathoanatomical cause; presentations due to a specific disease process (eg, infection and malignancy) are uncommon in primary care.3 Diagnostic imaging is only recommended for LBP without radicular pain when there is suspicion of a specific disease process (eg, malignancy, fracture, infection or spondyloarthropathies) that would be managed differently to non-specific LBP.4–6
Overuse of imaging for LBP has been reported for many decades with studies reporting that 20% of patients presenting with LBP received imaging in the UK7 8 and 25% in Australia9 and USA.10 However, the veracity of these estimates is unclear as there has not been a systematic review of studies evaluating the frequency of imaging in patients presenting for care with LBP.
In this systematic review, our aims were to: (1) estimate the proportion of patients seeking care for LBP who are imaged currently and (2) explore trends in the proportion of patients receiving diagnostic imaging over time. We also examined the effect of study-level factors on estimates of the imaging proportion. We hypothesised that the imaging proportion should have decreased over time as a result of clinical practice and therapeutic guidelines to limit imaging and more recently through campaigns such as Choosing Wisely (launched in 2012) warning about overuse of imaging for LBP.11–15
Methods
The study protocol was prespecified, and the review was conducted in accordance with Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) and Meta-analyses Of Observational Studies in Epidemiology (MOOSE) guidelines.16 17 The study protocol was registered with PROSPERO (http://www.crd.york.ac.uk/PROSPERO).
Searches
We searched MEDLINE, Embase and CINAHL for articles published between 1 January 1995 and 9 December 2017 in any language. The rationale for searching from 1995 was that the first evidence-based LBP guideline to provide advice for use of imaging was released in 1994.18 Search terms relating to primary or emergency care, imaging type and LBP were used (online supplementary appendix 1: MEDLINE search string). We supplemented electronic searches with hand searches of reference lists from eligible studies and contacted experts in the field of imaging and management of LBP. Two authors (AD and HJ) independently performed title and abstract screening with full-text articles assessed for study eligibility. Any disagreements were resolved by consensus.
Supplemental material
Study selection
Eligible study designs were controlled trials and observational designs (cohort, case–control, cross-sectional and interrupted time series). Studies needed to report on imaging requested or performed for patients presenting to primary or emergency care for LBP. We defined primary care as first contact care with a provider who could refer for imaging, including medical practitioners (eg, general practitioners) and allied health practitioners (eg, physiotherapists chiropractors and osteopaths). We defined emergency care as first contact care in the hospital emergency department setting. Studies were ineligible if not written in English, and translation to English was not feasible, if all participants were imaged, or if greater than 25% of the participant sampling frame was prior to 1995.18
Data extraction and risk of bias assessments
After all eligible studies were retrieved, two of four authors (AD, MH, HJ or CM) independently extracted data and assessed risk of bias with disagreements resolved by consensus, or a third rater, if required. Data from each study were extracted using a prepiloted form. Where available, we extracted data on: year of publication, study design, country, clinical setting, imaging modality, study sampling frame, imaging observation window (period of time between presenting to the clinician and the last time point at which data on imaging request/event was collected), participant characteristics, imaging proportion and study sample size. Authors were contacted to request additional data where required. We extracted imaging proportion based on the entire study sample for observational studies. For controlled trials testing strategies to reduce imaging, we extracted data from the control arm only.
Risk of bias was assessed using the tool developed by Hoy et al 19 for assessing risk of bias of prevalence studies. The tool comprises 10 items scored for risk of bias (low risk and high risk). Modification was made to two of the original 10 items to reflect the aims of this study. Representative population (item 1) was specified as a population seeking primary or emergency care for LBP, and prevalence period (item 9) was specified as the imaging observation window. We generated an overall summary risk of bias score (low, moderate and high risk) based on consideration of the 10 items.19 Any disagreements were resolved by consensus or a third rater.
Data synthesis
We defined simple imaging as plain radiography or ultrasound (U/S); complex imaging as CT, MRI or nuclear bone scan; and any imaging as the aggregate of simple and complex imaging when a study reported both. We defined current imaging for studies with greater than 75% sampling frame from 2010 or later. We extracted data from the most recent study when multiple studies reported on the same source data. The imaging proportion was calculated by extracting imaging counts, either requested or performed as a result of seeking care (numerator), divided by the sample size (denominator). To represent each study sampling frame, a single time-point (year) was calculated using the midpoint of the date range. When the study sampling frame was continuous, a single average imaging proportion was calculated. For discontinuous sampling frames (eg, 2002–2003; 2011–2012), average imaging proportions representing each period were calculated.
Data analysis
Current imaging proportion
To estimate the current imaging proportion, we calculated pooled proportions, grouped by clinical setting (primary or emergency), then by imaging type (simple, complex or any) using random-effects meta-analysis. The relative study weights assigned under a random-effects model are minimally influenced by extremes in study populations.20 21 Outlier studies (identified by visual inspection of the forest plot) were described and excluded from pooled analyses. Some clinical heterogeneity was expected due to variation in study population and clinical features.22 Statistical heterogeneity was assessed; however, meta-analysis was not deemed inappropriate simply due to high I2 values, as long as the individual study estimates fell in a reasonable range.23 24 Sensitivity analyses of pooled imaging proportions were performed based on summary risk of bias (high vs low or moderate risk). Each pooled proportion was graded for quality of evidence.
To assess quality of the evidence for pooled estimates of imaging proportion, we applied GRADE criteria for observational studies.25 26 Two reviewers (AD and HJ) scored four factors for each pooled estimate. Quality of evidence began as high and was downgraded one level for each of limitations in study design (for instance, <50% of studies were observational—potential for selection bias), summary risk of bias (>50% of studies scored moderate or high), inconsistency of results (imaging proportion point estimates had an absolute range >25%) and imprecision (the CI of the pooled estimate spanned >10% above or below the pooled estimate). Thresholds for downgrade were based on consensus.26 Indirectness of evidence was measured as part of the summary risk of bias. Any disagreement was resolved by consensus or a third rater.
Trends in frequency of diagnostic imaging
We explored trends in imaging over time for simple and complex imaging, using mixed-effects meta-regression with important prespecified study-level factors considered as covariates.27 28 We considered factors for regression if reported by greater than 85% of studies.27 We performed a univariate analysis for each factor. We built two multivariate models (simple imaging and complex imaging) with study year selected and other factors added. Final selection of study-level factors was based on clinical rationale (not data driven), testing for collinearity (variance inflation factor (VIF) <4)29 and ensuring the model was not overfit based on the number of available studies. We identified, then considered excluding from regression, extreme outlier studies based on a plot of standardised shrunken residuals as recommended by Harbord and Higgins.30 Statistical analyses used STATA-IC V.15—metareg,30 and Comprehensive Meta Analysis V.3.3 (Biostat, USA).
Results
The electronic database search and citation tracking identified 5011 potential studies of interest. After screening of titles and abstracts, we retrieved full-text copies of 191 articles. Forty-five studies were included in the review (42 unique data sources). Key reasons for exclusion included: imaging proportion inestimable, inappropriate study design and >25% of the participant sampling frame prior to 1995 (figure 1).
PRISMA flow. PRISMA, Preferred Reporting Items for Systematic Reviews and Meta-Analyses.
Characteristics of included studies
The 45 studies we identified represented 19 451 749 unique consultations to primary or emergency care, which resulted in 4 343 919 imaging requests/events over a 21-year period (September 1994–July 2015).7–10 14 31–70 Study sample size ranged from 55 to 10 255 661 participants,49 with government-supported studies the the most common. The majority of studies were from North America (Canada [2 studies], USA [29]), followed by Oceania (Australia [5] and New Zealand [1]), Europe (Germany [1], Italy [1], Poland [1] and Spain [3]) and UK (England [2]). Two studies studied exclusively elderly participants.10 60 The majority of studies were retrospective reviews of clinical records or insurer data. Most commonly reported modalities were radiography, CT and MRI. A small proportion of patients presenting with LBP received diagnostic U/S: Britt et al 34 (0.4%: 2002–2005; 0.6%: 2009–2012) and Allen et al 31 (0.6%: 2001–2009). However, neither study provided details about what structures were scanned. U/S is not appropriate for identifying the usual or common serious causes of LBP including fracture and cancer but may be considered in patients with suspicion of abdominal aortic aneurysm71 or renal colic.72 73 Radionuclide bone scan was also reported by Britt et al 34 (0.3%: 2002–2005, 0.2%: 2009–2012). Thirty-six of the 45 studies (80%) had an imaging observation window within 3 months. The imaging observation window ranged from same day (eg, imaging data from the emergency setting)62 out to 1 year (eg, review of private health insurer data).43 Study characteristics are provided in table 1.
Characteristics of included studies
The majority of studies scored moderate or high for summary risk of bias (34 studies, 76%). The most frequent reasons for high risk of bias were non-representative sample (eg, by excluding the elderly), broad case definition or imaging observation window greater than 4 weeks (table 2).
Risk of bias assessment
Current imaging proportion (2010 or later)
Sixteen studies provided information on current imaging, of which 12 collected data from primary care,14 34–36 43 44 50 53 55 61 67 68 and 4 from emergency care.59 61 62 65 Two studies measured imaging in both settings, either reported separately61 or combined.63
Current imaging in primary care
The pooled estimate of current proportion for simple imaging in primary care (7 studies; n=1 574 236) was 16.3% (95% CI 12.6% to 21.1%), rated as GRADE: moderate quality evidence. We considered the imaging estimate from Carey et al 35 an outlier so did not pool (estimate: 56.9%, 95% CI 49.9% to 63.7%; potentially influenced by participant self-report) (online supplementary appendix 2). For complex imaging (8 studies; n=2 323 559), the pooled proportion was 9.2% (95% CI 6.2% to 13.5%), GRADE: moderate quality evidence. We excluded Carey et al 35 from pooling as above (estimate: 80.1%, 95% CI 74.2% to 84.9%). For any imaging (8 studies; n=1 675 720), the pooled proportion was 24.8% (95% CI 19.3% to 31.1%), GRADE: moderate quality evidence (figure 2). Summary risk of bias (high vs low or moderate risk) did not significantly influence pooled proportions in primary care (p=0.21, 0.59, between group mixed-effects analyses for simple and complex imaging, respectively) (online supplementary appendix 3).
Current imaging proportion in primary and emergency care ordered by summary risk of bias (high to low). Pooled imaging proportions were calculated using random-effects meta-analysis. The ‘Any imaging’ subgroup contains only studies that reported both simple and complex imaging types. The imaging estimate from Carey et al 35 was considered an outlier so was not pooled. Pooled estimates were assessed for quality using GRADE criteria, downgraded for: *greater than 50% of studies were scored moderate or high risk for summary risk of bias (ROB); †point estimate range was greater than 25%.
Current imaging in emergency care
The pooled estimate of current proportion for simple imaging in emergency care (4 studies; n=16 552) was 26.1% (95% CI 18.2% to 35.8%), GRADE: moderate quality evidence. For complex imaging (4 studies; n=16 552), the pooled proportion was 8.2% (95% CI 4.4% to 15.6%), GRADE: high-quality evidence. For any imaging (4 studies; n=16 552) the pooled proportion was 35.6% (95% CI 29.8% to 41.8%), GRADE: high-quality evidence (figure 2).
Trends in frequency of diagnostic imaging over time
After removing duplicate datasets, 42 studies were available for meta-regression.46 56 63 See online supplementary appendix 2 for imaging proportion of all available studies. Of 12 prespecified study-level factors, four were ineligible (reported by less than 85% of studies) (table 3). To explore trends over time, we adjusted for clinical setting and imaging observation window. Univariate analysis for each of the eight remaining study factors are reported in online supplementary appendix 4.
Prespecified study-level factors
Simple imaging
We included 36 studies in the adjusted simple imaging model (figure 3A). Carey et al 35 and Tacci et al 66 were extreme outliers, so were excluded from the model (shrunken residual=3.4 and 3.1, respectively). We found no significant change in the proportion of simple imaging over 20 years from 21.2% (95% CI 16.2% to 27.2%) for imaging requested in 1995 to 21.3% (95% CI 16.4% to 27.2%) for imaging requested in 2015. Similarly, clinical setting and imaging observation window were not associated with frequency of simple imaging in the adjusted model.
Influence of study-level factors on imaging proportion. Panels A and B represent trends in frequency of simple and complex diagnostic imaging since the introduction of clinical imaging guidelines. All models were calculated in the logit space, adjusted for clinical setting and imaging observation window. Circles represent the imaging proportion for each study sized according to inverse of the within-study variance but do not reflect the study weighting in the statistical model. Each fitted line was calculated using mixed effects meta-regression back transformed from the logit scale. *Adjusted mean proportion and their 95% CI estimates were statistically adjusted for the presence of factors in the mixed effects model including year, clinical setting and imaging observation window held constant at their means. †Significant influence (p<0.05) for each factor in the adjusted model.
Complex imaging
We included 27 studies in the adjusted complex imaging model (figure 3B). Carey et al 35 was an extreme outlier, so was excluded from the model (shrunken residual=7.4). We found an absolute predicted increase in imaging proportion (p=0.03) from 7.4% (95% CI 5.7% to 9.6%) for imaging requested in 1995 to 11.4% (95% CI 9.6% to 13.5%) for imaging requested in 2015, equivalent to a relative increase in complex imaging of 53.5%. Clinical setting was associated with frequency of complex imaging (p=0.001) with an imaging proportion of 17.8% (95% CI 13.5% to 23.0%) for imaging requested in primary care and 10.9% (95% CI 9.9% to 12.1%) for imaging requested in emergency care. Length of observation window was also associated with frequency of complex imaging (p=0.001) with an imaging proportion of 8.4% (95% CI 7.3% to 9.6%) for imaging requested within 4 weeks of the initial visit and 11.7% (95% CI 10.2% to 13.3%) when imaging was measured across the whole study observation window. These three factors accounted for most of the variance in frequency of complex imaging (adjusted R2=75.3%).
Discussion
Statement of principal findings
There is moderate quality evidence from eight studies that during the ‘current’ phase, approximately one-quarter of patients who presented to primary care were referred for imaging, and high-quality evidence from four studies that approximately one-third of patients who presented to emergency care were imaged. Based on 27 studies (n=8 742 444), we found a 53% relative increase in complex imaging from 1995 to 2015. When all studies were considered, more complex imaging was requested in primary care compared with emergency care. We found no change in frequency of simple imaging over the same period.
Strengths and limitations of review
The strengths of this systematic review include use of a prespecified protocol, inclusion of studies published in languages other than English and consideration of all studies published after the introduction of the first clinical imaging guideline.18 We located studies from primary and emergency care as representative of settings where patients may seek care for LBP59 and provide summary data in a graphical format that enables clinicians to interpret unbiased imaging estimates, assessed for quality using the GRADE system.
One limitation was the magnitude of between study variance when estimating imaging proportion. To address this, we first grouped by setting and imaging type before applying random-effects meta-analysis. We applied mixed effects meta-regression to further explain sources of heterogeneity with the number of factors in the adjusted model constrained to avoid overfitting. In addition, we adjusted for imaging observation window for trends in frequency of diagnostic imaging over time. Due to the greater percentage of North American studies (69%), we advise caution when interpreting data based on geographic region. Eleven studies (24%) counted imaging requests instead of imaging events alone. The calculated imaging proportion from these studies may be overestimated given that not all imaging requests will be realised due to a range of issues (eg, patient choice and radiologist clinical consultation). Compared with imaging events, requests for simple imaging were higher (unadjusted model), with no significant difference found between requests and events for complex imaging (online supplementary appendix 4). We were unable to extract sufficient data on some prespecified study-level factors (eg, older age, duration of episode and presence of radicular syndrome) that may have influenced imaging rates. It remains unclear how these factors are associated with imaging proportions.
In relation to other studies
We believe this is the first systematic review of how commonly imaging was performed for patients who seek care for LBP. As such, we are unable to compare our results with previous reviews. Our study is a clear advance over non-systematic/narrative estimates from individual studies. For example, one study74 used the proportion of elderly patients who underwent imaging for acute LBP60 to estimate the potential cost saving across the adult US population in a campaign that targeted unnecessary imaging.
Imaging for LBP in the absence of indications of underlying pathology does not improve clinical outcomes,75 but we found that radiography ordering did not diminish over 20 years. Furthermore, we found complex imaging (which includes CT imaging) had increased over the same period. These findings align with a recent study by Morrisroe et al,76 who reported a relative increase of 74% in Medicare-funded CT scans in Australia for LBP (195 000 in 2004 vs 340 000 in 2015), while billing for radiography remained static over the same period. Similarly, Deyo et al 77 described a relative increase of 307% in Medicare Part B claims for lumbar spine MRI in the 12 years from 1994.
Meaning of the study: possible explanations and implications for clinicians and policy makers
We found that imaging for LBP remains high and has not decreased despite guideline advice, education campaigns and imaging referral decision systems. This pattern is consistent with a recent systematic review that found most interventions do not reduce imaging.78 There is a need for more research in this area to develop new strategies to reduce unnecessary imaging. This investment in research can be justified by the ‘costs’ of unnecessary imaging. Unnecessary imaging wastes scarce health resources and in the case of radiographs, CT and nuclear medicine, increases the risk of iatrogenic disease (cancers) because of cumulative ionising radiation.79 80 Another cost is that the risk of overdiagnosis increases with imaging (especially with complex imaging).12 This can promote poorer health outcomes through misguided patient or clinician concern,81 82 medicalisation of pain83 or through unfounded confidence that incidental findings on imaging are the cause of LBP.82 84 85 The implication is that high levels of non-indicated imaging may contribute to the disease burden of LBP,1 iatrogenic disease and perpetuate low value care.86–88
Unanswered questions and future research
The drivers of excessive imaging are multifactorial and incorporate many aspects of the health system including sluggish imaging guideline reform,89 reliance on individual red flags that offer little or no diagnostic value when a patient is triaged towards further investigations including imaging,90 91 regional variation (eg, different interpretations of legislation),44 cultural practices (eg, patient/practitioner beliefs)92 or financial interest (eg, clinicians with financial interest in MRI scanners).93 The majority of studies in our review either did not explore drivers of excessive imaging, or focused on a single issue such as health insurer variation36 or effect of clinical decision support.33 The complete picture of what drives excessive imaging when patients present with LBP remains unanswered.
There is a paucity of research that has investigated effectiveness of interventions to reduce imaging for LBP. A systematic review by Jenkins et al 78 found only in-hospital imaging decision support and targeted reminders reduced imaging referral, but recommendations were limited due to the low number of included studies, study heterogeneity and risk of bias. A recent study that investigated the effectiveness of decision support during imaging requests in the ED found a reduction in the volume of imaging after implementation.94 Involvement across multiple levels of healthcare (eg, clinicians, policy makers, payers and technology developers) has been recommended to help facilitate the adoption of clinical imaging decision support systems.95 Artificial intelligence algorithms may also assist clinicians with appropriate decisions about imaging96 but have yet to be tested in the initial management of LBP. Similarly, natural language processing algorithms when applied to large volumes of imaging request/report data may assist researchers to build improved clinical decision models for management of LBP.97 98 Further research to evaluate strategies aimed at reducing imaging as a contributor to overdiagnosis must be prioritised.12
Conclusion
We report moderate quality evidence from ‘current data’ that about one-quarter of patients who presented to primary care for LBP were referred for imaging, and high-quality evidence that about one-third of patients who presented to emergency care were imaged. Importantly, complex imaging has increased by 50% over 21 years despite guideline advice and education campaigns. These results draw attention to high levels of imaging in both primary and emergency care settings.
What is already known
The vast majority of low back pain has no pathoanatomical cause; patients should not undergo routine diagnostic imaging.
Overuse of imaging for low back pain has been reported for decades.
What are the new findings
We have moderate quality evidence that about one-quarter of patients who presented to primary care for low back pain were imaged. We have high-quality evidence that about one-third of similar patients who presented to emergency care were imaged.
The rate of complex imaging per patient increased by 50% from 1995 to 2015.
References
Footnotes
Contributors Conception and design: AD, MH, CGM and HJ. Analysis and interpretation of the data: AD, MH, HJ, CGM, MU and RB. Drafting of the article: AD, MH, HJ and CGM. All authors critically revised the article for important intellectual content and approved the final article. Statistical expertise: AD, MH and CGM. Administrative, technical or logistic support: CGM and MH. Extraction and assembly of data: AD, HJ, MH and CGM. The corresponding author (AD) attests that all listed authors meet authorship criteria and that no others meeting the criteria have been omitted. AD is guarantor.
Funding The authors have not declared a specific grant for this research from any funding agency in the public, commercial or not-for-profit sectors.
Competing interests For support outside the submitted work MU declares: funding by UK National Institute for Health Research and Arthritis Research UK, SERCO Ltd, personal fees from UK National Institute for Health and Care Excellence, personal fees from UK National Institute for Health Research and other from Clinvivo Ltd.
Provenance and peer review Not commissioned; externally peer reviewed.
Data sharing statement No additional available.
Patient consent for publication Not required.