Epidemiology of Fracture Nonunion in 18 Human Bones

Key Points

Question  Which patient-specific risk factors other than injury severity increase risk of nonunion of fractures?

Findings  In an inception cohort study of a payer database in which 309 330 fractures in 18 bones were analyzed, only 5 patient-specific risk factors significantly increased the risk of nonunion more than 50% across all bones: multiple concurrent fractures, prescription nonsteroidal anti-inflammatory drug and opioid use, open fracture, anticoagulant use, and osteoarthritis with rheumatoid arthritis.

Meaning  The probability of fracture nonunion can be determined from patient-specific risk factors at presentation.

Importance  Failure of bone fracture healing occurs in 5% to 10% of all patients. Nonunion risk is associated with the severity of injury and with the surgical treatment technique, yet progression to nonunion is not fully explained by these risk factors.

Objective  To test a hypothesis that fracture characteristics and patient-related risk factors assessable by the clinician at patient presentation can indicate the probability of fracture nonunion.

Design, Setting, and Participants  An inception cohort study in a large payer database of patients with fracture in the United States was conducted using patient-level health claims for medical and drug expenses compiled for approximately 90.1 million patients in calendar year 2011.The final database collated demographic descriptors, treatment procedures as per Current Procedural Terminology codes; comorbidities as per International Classification of Diseases, Ninth Revision codes; and drug prescriptions as per National Drug Code Directory codes. Logistic regression was used to calculate odds ratios (ORs) for variables associated with nonunion. Data analysis was performed from January 1, 2011, to December 31, 2012,

Exposures  Continuous enrollment in the database was required for 12 months after fracture to allow sufficient time to capture a nonunion diagnosis.

Results  The final analysis of 309 330 fractures in 18 bones included 178 952 women (57.9%); mean (SD) age was 44.48 (13.68) years. The nonunion rate was 4.9%. Elevated nonunion risk was associated with severe fracture (eg, open fracture, multiple fractures), high body mass index, smoking, and alcoholism. Women experienced more fractures, but men were more prone to nonunion. The nonunion rate also varied with fracture location: scaphoid, tibia plus fibula, and femur were most likely to be nonunion. The ORs for nonunion fractures were significantly increased for risk factors, including number of fractures (OR, 2.65; 95% CI, 2.34-2.99), use of nonsteroidal anti-inflammatory drugs plus opioids (OR, 1.84; 95% CI, 1.73-1.95), operative treatment (OR, 1.78; 95% CI, 1.69-1.86), open fracture (OR, 1.66; 95% CI, 1.55-1.77), anticoagulant use (OR, 1.58; 95% CI, 1.51-1.66), osteoarthritis with rheumatoid arthritis (OR, 1.58; 95% CI, 1.38-1.82), anticonvulsant use with benzodiazepines (OR, 1.49; 95% CI, 1.36-1.62), opioid use (OR, 1.43; 95% CI, 1.34-1.52), diabetes (OR, 1.40; 95% CI, 1.21-1.61), high-energy injury (OR, 1.38; 95% CI, 1.27-1.49), anticonvulsant use (OR, 1.37; 95% CI, 1.31-1.43), osteoporosis (OR, 1.24; 95% CI, 1.14-1.34), male gender (OR, 1.21; 95% CI, 1.16-1.25), insulin use (OR, 1.21; 95% CI, 1.10-1.31), smoking (OR, 1.20; 95% CI, 1.14-1.26), benzodiazepine use (OR, 1.20; 95% CI, 1.10-1.31), obesity (OR, 1.19; 95% CI, 1.12-1.25), antibiotic use (OR, 1.17; 95% CI, 1.13-1.21), osteoporosis medication use (OR, 1.17; 95% CI, 1.08-1.26), vitamin D deficiency (OR, 1.14; 95% CI, 1.05-1.22), diuretic use (OR, 1.13; 95% CI, 1.07-1.18), and renal insufficiency (OR, 1.11; 95% CI, 1.04-1.17) (multivariate P < .001 for all).

Conclusions and Relevance  The probability of fracture nonunion can be based on patient-specific risk factors at presentation. Risk of nonunion is a function of fracture severity, fracture location, disease comorbidity, and medication use.

The rate of fracture nonunion is estimated to be between 5%¹ and 10%,² and the rate of nonunion may be increasing as the survival rate for patients with severe injuries improves.³ The risk of nonunion is related to the severity of injury resulting in fracture,⁴ and many randomized clinical trials⁵ have shown that variations in nonunion rates are associated with different surgical treatments. However, progression to nonunion is not fully explained by these factors alone.⁶ Determination of the probability for and potential mitigation of nonunion risk is an important clinical objective because patients with nonunion can expect more long-term pain, physical disability, mental health problems, and medical treatment costs as well as a slower return to normal work productivity.⁷ Herein, we describe the epidemiology of fracture nonunion in adults, with a focus on information available to the clinician at patient presentation. We hypothesize that the interplay between a patient’s physiologic risk factors and fracture characteristics increases the risk of fracture nonunion.⁵ We tested this hypothesis in a large payer database of patients with fracture in the United States.⁸

Truven Health Analytics (Durham, North Carolina) compiled patient-level health claims data for medical and drug expenses, together with laboratory test results, hospital discharge information, and death data on 90.1 million patients.⁹ Data were submitted by hospitals, managed care organizations, Medicare and Medicaid programs, and approximately 300 large corporations in exchange for benchmark reports.⁹ This study was approved and exempted from the need for informed consent by the institutional review board of Duke University Medical Center because patient data were deidentified.

The final database contained 1 row per unique fracture, with comma-separated values for patient variables. Variables included patient demographics, treatment procedures as per the Current Procedural Terminology codes; disease comorbidities as per the International Classification of Diseases, Ninth Revision (ICD-9) codes; and drug prescriptions as per National Drug Code Directory (https://www.accessdata.fda.gov/scripts/cder/ndc/) codes.

Study inclusion was limited to patients with a coded bone fracture in calendar year 2011. Patients were excluded if they had less than 12 months of continuous enrollment following fracture so as to capture all coded nonunions.

Fractures were identified based on 5-digit ICD-9 codes. Rule-out codes were not counted; such codes are used to order radiography in some patients who may not have a fracture. In addition, codes with an unspecified character string in the definition were not used because such codes are replaced with a specific code defining the location of the fracture. Nonunion was determined by the presence of either a nonunion code or a code for prescription use of an electrical bone stimulation device since such devices are used to treat nonunion. Patients who used low-intensity pulsed ultrasound devices for a fresh fracture were excluded because this prescription device may increase the healing rate for bone.

Disease comorbidities were identified using ICD-9 primary disease codes. Secondary conditions arising from a chronic disease condition (eg, diabetic retinopathy) were not used as proxies for the primary disease because of the risk of double counting. Thus, our analysis would not identify patients with diabetes diagnosed before 2011, although medications used to treat diabetes would be captured. Medications were identified using National Drug Code Directory codes, which are for oral medications purchased in a retail pharmacy. Such codes contain a range of medications; the opioid class contains analgesics but can also contain opioid agonists used to treat addiction. Medications were assumed to be part of long-term therapy, with the exception of antibiotics, thrombolytics, analgesics, and corticosteroids.

Analysis focused on the cohort of patients aged 18 to 63 years at the time of the fracture. This age range was chosen because skeletal maturity is achieved by approximately 18 years.¹⁰ Patients younger than 18 years were abundant in the database, but their healing rate was high, so it was less compelling to identify risk factors for those who failed to heal. Patients older than 63 years were excluded because the requirement for 12 months of continuous enrollment created an artifact as patients transitioned to Medicare and no longer appeared in the database. Older individuals were also excluded because only some purchase Medicare supplemental coverage and thus are not representative of other Medicare patients.

Our overall hypothesis was that the probability of fracture nonunion can be determined with the use of risk factors derived from patient demographics, using Current Procedural Terminology, ICD-9, and National Drug Code Directory codes. Possible risk factors for nonunion were identified in a literature search,⁵ with a focus on risk factors likely to be of concern to orthopedic surgeons. We requested information on 257 potential nonunion risk factors, including fracture type, fracture cause, patient demographics, and medication use. We focused on 18 bones most frequently fractured. An operative treatment variable was defined for patients who received any fracture surgery and we compared them with patients who did not undergo surgery. Statistical analyses were performed using SAS, version 9.4 (SAS Institute Inc). The critical value for significance was set at P < .05.

Because so many variables were available for each patient, it was important to group variables into manageable categories. Ultimately, data were pooled to obtain 45 variables of interest (eAppendix in the Supplement). For example, patients had as many as 15 separate fractures, but we binned them into a smaller number of categories for analysis (eg, 1-2 fractures, 3-5 fractures, and ≥6 fractures). Multivariate logistic regression was used to control for correlations among the various risk factors.

We did not adjust for multiple comparisons because showing 95% CIs for each odds ratio (OR) achieves the same end. Furthermore, an OR significant at P < .001 is comparable to an OR significant at P = .05 that has been Bonferroni corrected for 50 comparisons. Correcting for additional comparisons would be likely to lead to type II (false-negative) errors. In an exploratory context such as this, P values should be interpreted as a measure of statistical evidence rather than a test of hypothesis.

In parallel to the logistic regression analysis, we also conducted random forest decision tree modeling using the same covariates.¹¹ The random forest method is invariant to interactions and terms of higher dimension, such as quadratic terms. The random forest method generally performs better than regression methods but, unlike regression analysis, is harder to interpret. We compared the methods using the C statistic (area under the receiver operating characteristic curve) and found that the models were substantially equivalent; the C statistic of the random forest model was only slightly larger than the C statistic of the logistic regression model, differing in the third or fourth decimal place. We elected to report only the main effects logistics regression model herein. Nonunion rates for individual bones were also compared with those in the literature. Data analysis was performed from January 1, 2011, to December 31, 2012.

A flowchart (eFigure in the Supplement) shows how the patient sample was assembled. There were 309 330 fractures in patients ranging from 18 to 63 years (mean [SD], 44.48 [13.68]), or approximately 6725 patients in each of the 46 age classes. The overall nonunion rate was 4.9% (Table 1), with substantial variation from bone to bone. The metatarsal was the most frequently fractured bone, with a nonunion rate of 5.7%. The lowest nonunion rates were for the metacarpal (1.5%) and radius (2.1%) bones. The highest nonunion rates were for the scaphoid (15.5%), followed by the tibia and fibula (14%) and femur (13.9%). These were the only bones for which the nonunion rate was greater than 10%. If 4.9% of all patients had nonunion fractures (Table 1), then there were approximately 330 nonunion patients per age class. This relatively small number of nonunion fractures per age class could potentially result in uncertainty in estimating nonunion rates as a function of age.

There were clear demographic differences between patients who healed and those who failed to heal (Table 1). Women had more fractures, but men had a higher proportion of nonunions. Open fractures represented 3.9% of all fractures, but 10.9% of open fractures were nonunion and 4.7% of closed fractures were nonunion. Multiple fractures were more likely result in nonunion; nonunion frequency was 4.4% in patients with 1 fracture but 24% among patients with 7 or more fractures.

Multivariate analysis determined the ORs of nonunion with comorbid disease when adjusted for other risk factors (Table 2). Multivariate evaluation of fractures associated with comorbidities (Table 2) indicated that 3 risk factors (osteoarthritis [OR, 1.45; 95% CI, 1.39-1.52], osteoarthritis with rheumatoid arthritis [OR, 1.58; 95% CI, 1.38-1.82], and type 1 diabetes [OR, 1.40; 95% CI, 1.21-1.61]) increased the odds of nonunion by at least 40%. Odds ratios for individual nonunion risk factors were generally small, with 13 of 16 multivariate ORs less than 1.40 (Table 2). Two risk factors were inversely associated with nonunion (OR, <1): cardiovascular disease (OR, 0.94; 95% CI, 0.90-0.98) and allergy (OR, 0.90; 95% CI, 0.86-0.93) (Table 2). The number of risk factors and the complex interactions possible between risk factors may explain the absence of multivariate ORs higher than 1.58 (Table 2).

Use of certain medications increased the nonunion risk (Table 3). After controlling for confounding variables, the most powerful risk factor was use of nonsteroidal anti-inflammatory drugs (NSAIDs) and opioids (multivariate OR, 1.84; 95% CI, 1.73-1.95). Other pain medications, such as opioids alone, and anticonvulsants, with or without benzodiazepines, were moderately strong positive risk factors, whereas antidiabetics other than insulin (OR, 0.92; 95% CI, 0.86-0.99) and oral contraceptives (OR, 0.88; 95% CI, 0.81-0.95) were inversely associated with nonunion.

Nonunion ORs were significantly increased for many risk factors (Table 4 and Table 5), including number of fractures (OR, 2.65; 95% CI, 2.34-2.99), use of prescription analgesics (NSAIDs and opioids; OR, 1.84; 95% CI, 1.73-1.95), operative treatment (OR, 1.78; 95% CI, 1.69-1.86), open fracture (OR, 1.66; 95% CI, 1.55-1.77), anticoagulant use (OR, 1.58; 95% CI, 1.51-1.66), osteoarthritis with rheumatoid arthritis (OR, 1.58; 95% CI, 1.38-1.82), anticonvulsant use with benzodiazepines (OR, 1.49; 95% CI, 1.36-1.62), opioid use (OR, 1.43; 95% CI, 1.34-1.52), type 1 diabetes (OR, 1.40; 95% CI, 1.21-1.61), high-energy injury (OR, 1.38; 95% CI, 1.27-1.49), osteoporosis (1.24; 95% CI, 1.14-1.34), male gender (OR, 1.21; 95% CI, 1.16-1.25), insulin use (OR, 1.21; 95% CI, 1.10-1.31), diagnosed smoking (OR, 1.20; 95% CI, 1.14-1.26), diagnosed obesity (OR, 1.19; 95% CI, 1.12-1.25), antibiotic use (OR, 1.17; 95% CI, 1.13-1.21), osteoporosis medication use (OR, 1.17; 95% CI, 1.08-1.26), diagnosed vitamin D deficiency (OR, 1.14; 95% CI, 1.05-1.22), diuretic use (OR, 1.13; 95% CI, 1.07-1.18), and renal insufficiency (OR, 1.11; 95% CI, 1.04-1.17) (all, multivariate P < .001).

Relatively few risk factors affected multiple bones (Tables 4 and 5). The need for operative treatment was associated with nonunion in 15 bones, anticoagulant use was associated with nonunion in 14 bones, use of analgesics (NSAIDs and opioids) affected 12 bones, and osteoarthritis and use of anticonvulsants each affected 11 bones. Overall, the largest risk factor for nonunion was the number of fractures (OR, 2.65; 95% CI, 2.34-2.99).

A limited number of putative risk factors were inversely correlated with nonunion (Tables 4 and 5), including oral contraceptive use (OR, 0.88; 95% CI, 0.81-0.95), allergy (OR, 0.90; 95% CI, 0.86-0.93), and age (OR, 0.97; 95% CI, 0.95-0.98); each was apparently protective in at least 2 bones. Some bones have several protective factors; both radius and ankle had 3 protective factors. A total of 12 bones had at least 1 protective factor.

Smoking was not identified as a major risk factor in this study (Tables 4 and 5). However, our data included only diagnosed past or current smoking, which most reliably may mean that someone was offered smoking cessation therapy. Thus, our count of smokers is likely an underestimate. Only 10.2% of patients were coded as past or current smokers (Table 1), whereas 18% of the general population is expected to smoke¹² and other nonunion cohort studies have reported a prevalence of smoking higher than the US average.¹³ Similarly, we may underestimate obesity prevalence; only diagnosed obesity was analyzed, so many people with nonunion fractures may have been obese but we have no record of it.

Lack of convergence of the model was a rare problem (Tables 4 and 5) except for coagulants, which appear to be rarely used among patients with fractures (Table 3). The number of patients with multiple scaphoid fractures may have been too small to find a solution for nonunion associated with multiple fractures. Only 10 bone–risk factor ORs other than coagulants could not be estimated.

There were clear demographic differences between patients whose fractures healed and those whose fractures failed to heal (Table 1). Most risk factors conferred a relatively small increase in multivariate nonunion risk (Table 2), perhaps because there are complex interactions between and among risk factors. Use of certain medications was an important determinant of nonunion (Table 3). Adjusting medication use by patients may enable physicians to improve the odds that a patient will heal. Nonunion rates varied among bones, and the contributions of various risk factors showed a complex interplay (Tables 4 and 5). In general, nonunion rate appears to be a function of fracture severity, fracture location, disease comorbidity, and medication use.

The distinction between univariate and multivariate ORs is important. For example, type 2 diabetes was associated with a univariate OR of 1.60 and a multivariate OR of 1.15 (Table 2). If that diagnosis is the only information that a clinician has about a patient, then it is reasonable to conclude that this patient has a 1.60-fold higher risk of nonunion than does a person without type 2 diabetes. As other variables become known and can be incorporated into a risk assessment, the risk associated specifically with type 2 diabetes decreases. In a multivariate analysis, which controls for many other factors, the nonunion risk associated with type 2 diabetes was 1.15-fold times the risk of nonunion in a person without that disease (Table 2). It is almost certainly the case that type 2 diabetes has not been diagnosed and treatment has not been instituted in some patients; thus, the disease would not have been analyzed in this study. Diabetes medications other than insulin appear to provide protection from nonunion (Tables 4 and 5), although the mechanism of such a protective effect is not known. When working with large patient databases (eg, big data), unanticipated associations are likely to be found. Causality cannot be tested without using an experimental approach; therefore, big data projects should be regarded as an opportunity for hypothesis generation rather than hypothesis testing.

The overall nonunion rate that we report was 4.9% (Table 1). This rate is somewhat lower than others reported in the literature,^1,2 although the healing rate for individual bones aligns well with previously published information. For example, the tibial nonunion rate we report was 7.4% for 12 808 fractures. The literature suggests that the expected nonunion rate for tibial fractures is 7.6%, a value derived by collating the reported healing rate in 46 publications spanning the period from 1976 to 2014 and including 5920 fractures treated with a variety of conservative and operative methods.^14-58 The agreement between our database and the literature is striking since both samples include many patients who received a wide range of treatments.

Similarly, the nonunion rate we report for clavicle fractures is 8.2% for 7414 fractures. Collation of literature on clavicle fractures suggests that the expected nonunion rate is 8.6%. This percentage was derived from the reported healing rate in 12 separate publications^59-70 spanning the period from 2004 to 2013 and including 3168 patients treated with a range of conservative and operative methods. Both samples are large, and the nonunion rate in our database again differs by less than a percentage point from the literature values.

The overall nonunion rate we report (4.9%) is slightly lower than the 5% to 10% that reviews suggest,^1,2 which may reflect reporting bias. Clinicians generally report more often on fractures that heal poorly.⁷¹ Metacarpal fractures heal with a nonunion rate of just 1.5% and have been reported only 226 times in the literature (per PubMed search, December 21, 2015: search terms, bone fracture healing and metacarpal). Conversely, tibia fractures have a nonunion rate of 7.4% and have been reported 2578 times in the literature (per PubMed search, December 21, 2015: search terms, bone fracture healing and tibia). Similarly, femur fractures have a nonunion rate of 13.9% and have been reported 2791 times in the literature (per PubMed search, December 21, 2015, search terms, bone fracture healing and femur).

This research has several limitations. First, Truven Health Analytics is a payer database that excludes unemployed or indigent patients who might have had a higher rate of nonunion. Second, some diagnoses may have been coded incorrectly as fracture if coding was performed before radiographic imaging. However, this error is unlikely because rule-out codes were not counted; such codes are used to order imaging if a fracture is suspected. Third, some fractures that received treatment in 2011 could have occurred earlier. Fourth, uncoded patient data cannot contribute to conclusions. Data were more likely to be missing from the database if information was not crucial to reimbursement; smokers not receiving medication for smoking cessation would likely not be coded as smokers. Nevertheless, missing data are a problem in every study, including randomized clinical trials. Therefore, we used statistical methods, including random forest analysis, that are robust and resilient despite missing data.¹¹ Additional limitations are characteristic of claims databases in general; there is imprecision in ICD-9 coding schemes,⁷² and coding errors are common but are assumed to distribute randomly and to be minimized by the need for accurate reporting for claims reimbursement and by legal penalties for fraudulent reporting.⁷³ Limitations do not appear to be a substantial issue because the nonunion rates we report are similar to others in the literature.

An important strength of this research is that it has contributed novel insights into the cause of fracture nonunion. For example, use of certain medications is a key risk factor for fracture nonunion (Table 3). Because medication use is a modifiable risk factor, our findings suggest that clinicians could counsel patients about use of medications. Other strengths of this research are that all data were collected prospectively, the 12-month follow-up time represents a longer follow-up interval than used in most other published studies,⁵ and the outcomes reported reflect real-world outcomes.

Fracture nonunion can result from the interplay of many risk factors. Key risk factors include features of the fracture, such as severity and location, as well as issues such as comorbidities and use of certain medications. Fracture severity, fracture location, comorbidity, and medication use could be incorporated into an algorithm that would help clinicians determine which fractures are at greatest risk of nonunion. Medications have a significant effect on fracture healing and can potentially be altered after fracture to improve healing.

Corresponding Author: R. Grant Steen, PhD, Medical Affairs, Bioventus LLC, 4721 Emperor Blvd, Ste 100, Durham, NC 27703 (grant.steen@bioventusglobal.com).

Accepted for Publication: May 23, 2016.

Published Online: September 7, 2016. doi:10.1001/jamasurg.2016.2775.

Author Contributions: Dr Steen had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.

Concept and design: Zura, Xiong, Watson, Ostrum, Della Rocca, Mehta, McKinley, Steen.

Acquisition, analysis, or interpretation of data: Zura, Xiong, Einhorn, Watson, Prayson, Della Rocca, McKinley, Wang, Steen.

Drafting of the manuscript: Zura, Della Rocca, Steen.

Critical revision of the manuscript for important intellectual content: All authors

Statistical analysis: Xiong, Wang, Steen.

Administrative, technical, or material support: Watson, Della Rocca, Mehta, Wang, Steen.

Study supervision: Zura, Einhorn, Ostrum, Prayson, Della Rocca, Mehta, Steen.

Conflict of Interest Disclosures: Drs Zura, Xiong, Einhorn, Watson, Ostrum, Prayson, Della Rocca, Mehta, McKinley, and Wang are paid consultants to Bioventus LLC; Dr Steen is an employee of Bioventus LLC.

Funding/Support: All financial and material support for this research was provided by Bioventus LLC.

Role of the Funder/Sponsor: Bioventus LLC had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, and decision to submit the manuscript for publication; however, they approved the final manuscript.

Additional Contributions: Peter Heeckt, MD, PhD, and Neill M. Pounder, PhD (Bioventus LLC), read the manuscript critically. John Jones, MS (Bioventus LLC), supervised formulation of the statistical analysis plan that guided this work. There was no additional financial compensation for these contributions.