The Comprehensive Hospital Abstract Reporting System (CHARS) is an inpatient discharge database of all non-federal hospitals in Washington State.{Washington State Department of Health, 2011} Hospitals submitting data to CHARS receive quality reports that they then certify as being at least 95% accurate for reporting discharges. We examined CHARS data from 2004 through 2007 for all hospitals and attending surgeons who performed at least one inpatient lumbar fusion for a degenerative spinal diagnosis. A small number of cases were dropped because they appeared to be a duplicate record.

We identified adults (age 20 or older) who underwent an initial inpatient thoracolumbar, lumbar, or lumbosacral fusion operation for degenerative spinal conditions from 2004 through 2007. This starting year (2004) was selected because it corresponds to the first calendar year that codes for three or more disc levels fused (4 or more vertebrae) became available. Patients were identified using relevant diagnosis and procedure codes from the October, 2010 update of the International Classification of Diseases, 9th Revision, Clinical Modification (ICD-9-CM).{Centers for Disease Control & Prevention, 2010} Details of our algorithm for classifying spine-related medical encounters, surgical characteristics, and safety can be obtained from the corresponding author.

CHARS allows up to nine diagnosis codes and up to six procedure codes for each admission. We searched all codes to identify patients undergoing fusion surgery for common degenerative spinal conditions, including disc degeneration (e.g. spondylosis), herniated discs, stenosis, spondylolisthesis, and scoliosis. We did not include patients who had non-degenerative spinal admissions in the previous year, such as spinal fracture, vertebral dislocation, spinal cord injury, or inflammatory spondylopathy. We also excluded patients who, in the previous year, had inpatient admission codes for accidents, neoplasm, HIV or immune deficiency, intraspinal abcess, osteomyelitis, infection, pregnancy, and cervical or thoracic spinal diagnoses and/or procedures.

All lumbar fusions were included, whether or not they were combined with a discectomy or laminectomy. Patients who had other types of spine-related procedures, including artificial disc replacement, corpectomy, osteotomy, kyphectomy, insertion of spacers, and insertion of dynamic stabilizing devices were excluded even if these were performed in conjunction with a fusion operation. We also excluded patients who, in the preceding 10 years, had any type of prior lumbar spine surgery, or had diagnosis or procedure codes that implied a previous lumbar operation (such as “reopening of laminectomy site”, “refusion”, or “removal of an internal fixation device”). Previous surgery has been shown to be an important predictor of higher complication and repeat surgery rates {Deyo 2011}.

We created a composite indicator of 90 day major surgical complications, defined as postoperative device complication, life-threatening complication, wound problem, or death occurring within 90 days. We separately examined the rates of repeat lumbar spine surgery. The presence of each complication (or repeat surgery) was coded as a dichotomous variable based on ICD-9-CM diagnosis and procedure codes. Each complication was attributed to the hospital and attending surgeon performing the initial fusion, and was not counted as an index case or a complication for another hospital or surgeon.

Repeat surgery was identified as the first instance of any second lumbar spine operation (i.e. “reoperation”) and not necessarily a repeat of the same procedure or performed at the same vertebral level. Device complications were based ICD-9-CM diagnosis and procedure codes that indicate a problem with an internal orthopaedic device (e.g. “malfunction of orthopaedic device”). We did not count device complications or repeat spinal surgeries coded during the index admission, and also required that they be co-coded with a lumbar spine-specific diagnosis and/or procedure code.

We counted wound problems, life-threatening complications, and death if they occurred during the 90-day postoperative surveillance or during the index admission. We followed Deyo’s (2010) approach for identifying life-threatening and wound complications in spine surgery.{Deyo et al. 2010} Life-threatening complications included major medical events such as acute respiratory failure, cardiopulmonary resuscitation, endotracheal intubation, pneumonia, stroke, and mechanical ventilation. Life-threatening events have major consequences on health, and their ICD-9-CM coding is more reliable than those of minor complications.{Lawthers at al. 2000} Post-operative wound problems were identified using codes for hemorrhages, excisional debridement of infection, postoperative wound disruption, seroma, and hematoma complicating a procedure. Mortality was identified by linking each index case to the Washington State vital records.

Patient characteristics and operative features may explain variation in rates of complications and repeat surgery across hospitals and surgeons. We adjusted the rates of complications for differences due to patient age, sex, comorbidity, diagnosis, and primary insurance. We used Quan’s “enhanced” version of the Charlson index (categorized “none”, “one”, and “two or more”) to adjust for comorbidity, applying it to admissions occurring at or during the year preceding each index visit.{Quan et al. 2005} However, since the comorbidity index includes myocardial infarctions and strokes, and these are among the life-threatening surgical complications that we sought to identify, we only included these items in the comorbidity score if they occurred during the previous year and did not count them if they occurred during the index admission.

The primary source of payment was coded into the following groups: “Medicare”, “Medicaid”, “Health maintenance organization” (e.g. Kaiser, Group Health Cooperative), “Commercial insurance” (e.g. Mutual of Omaha, United Health Plan, Safeco), “Workers’ compensation”, “Health service contractor” (e.g. Premera, Premera/Blue Cross), and “Other”. The latter category included charity cases, self-pay, and “other government sponsored patients” (e.g. Tri-care, CHAMPUS, Indian Health, Corrections) that combine to account for 2.5% of the total cases. Variables for race or ethnicity were not provided in CHARS during the study years.

Some operative features that may be associated with outcomes are identifiable from ICD-9 codes, including use of bone morphogenetic protein (BMP), surgical approach (anterior, posterior, or combined/circumferential), an indicator of whether fusion was combined with a decompression, and an indicator of whether 3+ disc levels were fused.{Fritzell et al. 2002, Cizik et al. 2012, Deyo et al. 2010, Deyo et al. 2012, Mirza 2011}. We also included the attending surgeon’s fusion volume in the preceding 365 days to account for any potential association between surgeon experience and outcomes.{Silber et al. 2010}. This measure of recent volume may be a better indicator of surgeon experience than the annual or the cumulative volume, both of which may be prone to misclassification.{French et al. 2011} We grouped the index cases into quartiles based on the ranking of the surgeon’s volume.

Bivariate associations between patient characteristics and complications were assessed using chi-square or t-test comparisons. We examined the risk for complications using logistic regression analysis, including only patients who had a minimum of 90 days of surveillance available in order to assess each outcome. For example, patients who had an initial operation in December of 2007 were not eligible to be assessed for the 90 day outcomes because the data only extend through 2007.

We used a logistic regression model that allowed a random-intercept for each hospital.{Rabe-Hesketh, 2008} These models prevent hospitals with a relatively small volume of cases from being misclassified when high sampling variability may account for apparently poor (or excellent) performance (i.e. through the use of “shrinkage factors”). The models also adjusts the standard errors to account for similarities occurring among patients “clustered” within hospitals.

To compare risk-adjusted outcomes across hospitals, we used the results from the logistic regression models to estimate the risk of complication for an “average” patient. This was accomplished by setting the covariates for age, sex, comorbidity, insurance, and diagnosis to the mean statewide distributions and then reporting the mean rate of complications within each hospital, along with a 95% Empirical Bayes confidence interval.

To examine the influence of surgeon factors on the rates of complications, we then added an additional random-effect parameter to the previous (“hospital only”) model. This parameter represents the variability attributed to surgeons’ “nested” within each of the hospitals he/she operated in.

Model construction was based on the principles by Hosmer & Lemeshow (2000), including checking for interactions. We used likelihood ratio and Aikake’s Information Criteria (AIC) to identify the models with the best fit. An advantage of AIC is that it allows comparisons of models with the same number of parameters; a useful consideration when comparing random-effects models. The best-fitted model, with the lowest AIC, included patient and operative features along with surgeon random-effect. The specification for this final model, along with 95% Bayesian coverage intervals is: Equation 1log(μikKk1-μikKk)=α+Kk+βXikKk~Norm(0,σ2) Equation 295%CIk=(K^k-1.96σ^k2,K^k-1.96σ^k2)

In this specification, μ represents the probability of a complication for patient i, whose initial operation was performed by surgeon k. The terms α combined with represents a surgeon-specific intercept. X represents a vector of covariates with corresponding beta-coefficients for patient characteristics (age, sex, comorbidity, primary diagnosis and insurance) and operative features (surgical approach, multilevel fusion, BMP, surgeon volume, instrumentation).

We found no evidence of a poorly fitted model for complication (p = 0.5615) using Hosmer-Lemeshow goodness-of-fit statistic. We confirmed linearity between the logarithmic odds for reoperations with each of the variables in the model; and no issues were identified that forced us to transform any variables. No changes in parameter coefficients as variables were added to the model suggested evidence of multicolinearity; this was confirmed by examining variance inflation factors and tolerance statistics in the final model.

In all models we assumed that the random effects were additive and normally distributed. Empirical Bayes predictions of the random-effects were assessed for normality using a Q-Q plot. There were no serious violations in the random-effect distribution through the bulk of the data. However, at the high-end of the random-effect spectrum, several surgeons’ complication rates deviated higher than predicted; indicating that that they had unusually high complication rates. This finding translates into a conservative estimation (underestimate of their actual rates) of the complication rates for these particular surgeons, because they are “over-fitted” toward the overall state mean. Our final model had a C-statistic of 0.662 for the complication model and 0.660 for reoperation model.

We then estimated the proportion of total variation that was explained by the inclusion of hospital (or surgeon) effects. The total variability for a random-effect model of a dichotomous outcome is calculated as the sum of hospital (surgeon) -specific variance(s) (ψ) plus the mathematical constant π²/3 {Rabe-Hesketh, 2008}. The proportion of total variation explained by covariates that are added to the previously null model (a model with random-effects parameter but no explanatory variables) is calculated using: Equation 31-var(FullModel)var(NullModel)

An Intraclass Correlation Coefficient (ICC) was calculated for hospital (surgeon) effects to estimate the proportion of variation explained within the hospitals (surgeons) by patient characteristics and operative features. The ICC represents the percentage of total variability due to between- hospital (surgeon) differences not accounted for by covariates in the model. For a dichotomous outcome the ICC is calculated using the following equation: Equation 4ψψ+π2/3

To measure the proportion of hospital (surgeon) variation that was explained by the addition of covariates, a model with only random-effects (null) was compared to subsequent models with covariates. A reduction in the ICC as covariates are added to the model represents the proportion of hospital (surgeon) variation that is explained at that level by the added covariates.

We also examined our findings after excluding 6 (24%) hospitals that included only one surgeon who performed fusions because, in these instances, the random-effect models cannot differentiate the variation in complications at the hospital level from that of the surgeon level. All analyses were performed using Stata-MP, version 11 (College Station, TX), with hypothesis testing performed using a two-sided alpha level set at 0.05.

A total of 7,680 patients were identified as having an initial inpatient fusion for lumbar degenerative conditions between 2004 and 2007. A total of 1,589 (20.7%) were excluded for reasons shown in table 1. The predominant reason for exclusion was lumbar surgery in the previous 10 years. Excluded patients were statistically older (57.7 years versus 56.4 years, p=0.001), but this was not clinically different. They were also less likely to be female (55.1% versus 60.7%, p<0.001), more likely to be insured by workers’ compensation (16.6% versus 11.3%, p<0.001), to have a comorbidity score greater than one (41.8% versus 27.2%, p<0.001), less likely to have spondylolisthesis (33.1% versus 54.1%) and were more likely to have degenerative disc disease (24.9% versus 18.0%) and herniated disc (16.2% versus 9.4%).

The study population consisted of 6,091 patients who had an initial inpatient spinal fusion during the study period (table 2). The mean age of the cohort was 56.4 years (sd 13.9); 31.3% were insured by Medicare, 6.1% by Medicaid, 11.3% by workers’ compensation; and 27.2% had a comorbidity index greater than zero. A diagnosis of spondylolisthesis accounted for the largest fraction of the fusion operations (54.1%), followed by degenerative disc disease (18.0%), herniated disc without myelopathy (9.4%), spinal stenosis (9.0%), scoliosis (7.1%), and herniated disc with myelopathy (2.3%). The index operations were performed by 298 surgeons, who on average performed 20 fusions each (range 1 – 333) during the study period. These were performed within 43 hospitals that performed an average of 142 (range 3 – 533) fusions during the study period.

Age between 61 and 80, greater comorbidity, and Medicare and Medicaid insurance were associated with higher risk of having a complication within 90 days (table 2). Workers’ compensation (a public payer in Washington State) and Health Maintenance Organizations had the lowest 90 day rates of complications among all types of insurance. Having 3+ disc levels fused, use of bone morphogenetic proteins, and anterior surgical approaches were associated with higher complication rates.

Unadjusted complication rates for herniated disc without myelopathy (2.6%) were lower than for disc degeneration (4.2%), spondylolisthesis (4.5%), spinal stenosis (6.4%), herniated disc with myelopathy (8.6%), or scoliosis (9.7%). Complications were not mutually exclusive. For example, the same patient may have had a device complication and a reoperation. Wound problems were the most common type of complications (3.1%), followed by repeat surgery (2.2%), life- threatening complications (2.0%), device problems (0.8%), and death (0.2%). Those with a diagnosis of scoliosis had a higher rate of complications than all other diagnoses, including a mortality rate of 1.2%.

Table 3 presents the multivariate logistic regression models with surgeon random-effects used to estimate the risk-adjusted rates for complications within 90 days. The overall 90 day complication rates for surgeons, adjusted for patient age, sex, insurance type, comorbidity and diagnosis was 4.1% (95% CI 2.7 – 6.4), and the adjusted mean 90 day reoperation rate was 1.7% (95% CI 0.9, 3.7).

Anterior operative approaches were associated with a significantly higher risk for 90 day repeat surgery (OR 3.33; 95% CI 2.03 – 5.44), even after adjusting for patient characteristics, comorbidity, diagnosis, insurance status, and other operative features. Circumferential fusions were associated with a higher 90 day risk for complications (OR 1.24; 95% CI 0.82–1.89) and reoperations (OR 1.13; 95% CI 0.59–2.17); but did not reach statistical significance in this small subset of patients. Patients who had 3+ disc levels fused had a higher risk for complications (OR 1.64 95% CI 1.12 – 2.40).

Higher surgeon case volume was not associated with a significantly lower 90 day repeat surgery or complication rate. The use of bone morphogenetic proteins (BMP) was associated with a non-significantly higher risk for complication (OR 1.20; 95% CI 0.88 – 1.63) and repeat surgery (OR 1.78; 95% CI 1.17 – 2.69). Having 3+ disc levels fused was also associated with an increased risk for complication (OR 1.64 95% CI 1.12 – 2.40) and repeat surgery (OR 1.78 95% CI 1.17 – 2.69).

Scoliosis was associated with a higher risk for complications within a given surgeon, while the risk for stenosis, spondylolisthesis, and herniated disk did not significantly differ from that of disc degeneration. Stenosis was associated with higher reoperation rates within a given surgeon, compared to degenerative disc disease. Adjusting for operative features lowered the odds ratio for complications among those with scoliosis from 1.94 (95% CI 1.23 – 3.09, not shown) to 1.58 (1.25 – 3.87, final model.)

The rates for complications are shown in figure 1, with each spike representing a single hospital or surgeon. We found that the nearly eightfold variation in risk-adjusted repeat surgery across surgeons was substantially attributed to a few surgeons (11/298, 3.7%) with rates significantly above the state mean (represented by a solid horizontal line). Limiting our analysis to the hospitals where more than one surgeon performed a fusion did not substantially change the variance estimates or lead us to alter our conclusions.

Table 4 provides the random-effect variance and fit parameters for a model without any covariates (null model, containing only random effects), as well as for models adding patient characteristics and operative features. In the null model for reoperations, the proportion of total variation due to hospital effects decreased from 8.8% to 4.0% with the addition of a parameter for the surgeon effects (a reduction of 54.5%). Similar reductions in hospital-level variation were observed in the models containing patient characteristics and operative features. Hospital effects accounted for a smaller proportion of the total variation in complication rates (3.6%), but were substantially reduced by the inclusion of a surgeon-parameter.

Surgeon effects accounted for 14.4% of the total variation in reoperations and 5.0% of the total variation for complications. Surgeon-level variation for both repeat surgery and complications was greater than the variation observed among hospitals. For example, the variance for the surgeon effects for repeat surgery was 0.341 (9.0% of total), compared to 0.153 (4.0%) for hospitals.

The addition of patient characteristics did not reduce the variability in reoperation rates, but accounted for 32.6% of the between-surgeon variation in complication rates. The addition of operative features accounted for 4.4% of the total variation in reoperations (30% of the variability between surgeons), and 2.5% of the total variation in complications (50% of the between-surgeon variation).

The unadjusted rate of repeat lumbar spine surgery in our cohort was 2.2% at 90 days and 5.0% at 1 year. We found that 115/137 (84%) of the reoperations were performed by the same surgeon who performed the initial surgery. The most common diagnoses at the time of these reoperations was spinal stenosis (21.2%), followed by disc degeneration (18.3%), spondylolisthesis (17.5%), disc herniation with myelopathy (13.1%), disc herniation (6.6) and scoliosis (5.8%). The remaining 16.1% were for non-degenerative diagnoses such as removal of internal orthopaedic device or surgical aftercare. Device complication codes were included in 36/137 (26%) of the all repeat surgeries within 90 days, and the code for “arthrodesis status” (implying a problem at the same vertebral level as the index surgery) in 53/137 (38.7%). The most common procedures coded at the time of reoperation were fusion or re-fusion (51.8%), decompression only (22.6%), and other spinal procedures such as removal of orthopaedic devices (25.5%).