The influence of comorbidity scores on re-operations following primary total hip replacement

The influence of comorbidities on outcome after joint replacement has recently gained a lot of interest. The number of articles in the PubMed index mentioning both arthroplasty and comorbidity increased by more than four-fold between 2001 and 2011 (Fig. 1). We believe that a proper evaluation of the comorbidity measures and their validity is required, especially with the rapid rise in arthroplasty registers and their impact on the evaluation of implants.¹ If the measures are valid, they may improve assessment of the patient case-mix, thereby avoiding premature dismissal of implants, or wrongly highlighting underperforming clinical centres.²

A comorbidity in this case is defined as any disease other than one leading to total hip replacement (THR),³ and can be classified using International Classification of Diseases (ICD)-based scores.⁴ ICD-based scores permit comorbidities to be measured using administrative databases instead of traditional scores,⁵ such as the American Society of Anesthesiologists Anesthesiologists score (ASA).⁶ There are currently two ICD-based scores that are commonly used: the Charlson co­morbidity index⁷ and the Elixhauser comorbidity measure.⁸ The Charlson score has also been simplified and adapted for a surgical setting by the Royal College of Surgeons of England (RCS).⁹

In our opinion a good, valid comorbidity measure should correlate the score with an increase or decrease in risk and have a strong predictive capacity, in that it should have a have a high sensitivity and specificity for individual patients.

The objective of our study was to validate and compare the Charlson, the RCS Charlson and the Elixhauser comorbidity measures with regard to re-operation rates within 12 years after THR.

Patients and Methods

This was a registry study with prospectively collected data on a nationwide cohort. The primary outcomes were re-operation within zero to two years and two to 12 years after primary THR. Re-operation is defined as a surgical procedure localised to the hip joint that can in some way be related to a previous THR, regardless of whether any of the parts or the entire implant is exchanged, removed or left in situ. Closed reductions without surgical intervention after dislocations are not included as procedures in the registry.

Data sources

The Swedish Hip Arthroplasty Register was started in 1979 and provides prospective observational nationwide data. Since 1992 the patient’s personal registration number has been collected, allowing an individual specific follow-up regarding re-operation rate.¹⁰ In order to calculate the comorbidity measures, the patients were matched using their personal registration numbers with data from the Swedish National Patient Register (NPR).¹¹ The NPR was started in 1964 and includes all inpatient care in Sweden since 1987, with discharge codes according to ICD-9 and ICD-10, and admission/discharge dates.

Operations performed between 1992 and 2007 were identified and re-operations were recorded up to 2008. All patients who had received a THR for primary osteoarthritis (OA) were included. Resurfacing THRs and implant systems used < 500 times were excluded. Follow-up time was defined as the period between the initial operation and the day of re-operation, death or the end of the study period, whichever came first.

Comorbidity measures

These were constructed from different predefined disease categories. The Elixhauser score⁸ is the most detailed, with 30 different categories (including two subcategories for hypertension) (Table I). The Charlson score⁷ has 19 categories and the RCS version of Charlson⁹ has 14. There is, as expected, a large overlap between the categories in the three measures. The Elixhauser and the RCS Charlson score are calculated by counting comorbidities without any pre-assigned weights: for example, a score of 2 indicates that the patient has two comorbidities. The original Charlson score is calculated by applying weightings that range between 1 and 6: for instance, myocardial infarction generates one point, whereas metastatic cancer generates six points. All scores were grouped into three categories: none, one to two and three or more points.

Apart from these three ways of calculating the score we also tested the models by applying an unweighted Charlson score, adapted weights of the Charlson score according to Quan et al¹² and a weighted Elixhauser score according to van Walraven et al.¹³

In order to calculate the Charlson and Elixhauser comorbidity scores we used the ICD codes proposed by Quan et al.¹⁴ The RCS Charlson score was calculated by combining the proposed simplified codes for ICD-10⁹ and Quan et al’s¹⁴ ICD-9 codes for the same groups. Transition from ICD-9 to ICD-10 was implemented gradually in Sweden between 1997 and 1998, and therefore the estimated combined results for the full period were compared with the period of 1999 to 2007.

The script calculating the comorbidity measures is available online.¹⁵ This identifies any registered ICD code one year prior to surgery.^9,16 The ICD codes indicating acute diseases (e.g. myocardial infarction) were only considered valid if applied in association with a hospital admission before admission for THR. The calculation also acknowledged the hierarchy of diseases: for example, cancer does not give additional points if the patient has a metastasis code registered.

The calculation script was validated by testing for the ICD codes from Quan et al’s article.¹⁴ We also validated the script manually on a small data subset. Owing to minor changes in the Swedish version of the ICD-10 and the previous adaptation of the ICD-9, we checked that the international comorbidity codes agreed with the Swedish ones. We found no relevant changes to the any of the scores’ calculations.

Confounding variables

We used age, gender and whether it was the first or the second THR as possible patient-related variables. Implant-related variables were restricted to four general fixation types: cemented, uncemented, hybrid or reverse hybrid. Surgery-related confounding variables were also controlled for by the type of hospital, hospital THR volume and year of surgery. The clinic volume was grouped by number of THRs per year into five groups, with an increment of 50 patients where the largest group contained centres performing > 200 THRs per year.

Statistical analysis

The calculations were carried out using R version 2.15.2 (R Foundation for Statistical Computing, University of Vienna, Austria), with Cox’s regression from the rms package.¹⁷ The package cmprsk¹⁸ was used to calculate competing risk regression to validate the results from the Cox’s regression. Competing risks is similar to regular Cox regression, except that it takes into account patients who have died and therefore cannot be re-operated upon. The graphs were created using the package Gmisc.¹⁹

As age is an important variable we chose to model it as a spline instead of traditional cut points, or a straight line. A spline is a line that can bend according to certain rules. This allows for a smooth relationship throughout the variable’s span, with minimal residual confounding.²⁰ We used restricted cubic splines; these are splines that use cubic terms in the centre of the data and restrict the ends to a straight line, thus helping to avoid the centre distorting the ends. The flexibility of a spline is chosen by the number of ‘knots’: more knots allow a more detailed description of the relationship. Somewhat simplified, each knot connects two sections of the line, for instance a linear spline with one knot will have a V-shape, while two knots allow for a N-shaped relation. As the sharp twists of a linear spline are undesirable, the cubic function is generally applied to smoothen the curve over the different knots. The position of the knots is in most cases chosen by different quantiles depending on the number of knots. In order to avoid over-fitting the regression model by choosing too many knots, the number of knots was chosen using the Akaike information criterion (AIC).^21,22

The models were stratified for first or second THR, operation time in intervals of three years between study start and end, type of clinic and clinic volume. Effect modification for implant fixation was investigated for gender, age >  75 years and age < 50 years. An effect modification between gender and implant fixation would indicate that men and women might respond differently to different fixation types. We also tested for effect modification between fixation and the extremes of age, as we believed that bone quality or patient activity might affect the impact of fixation. The ages between 50 and 75 years served as a reference in the model in this case.

The proportional hazards assumption for the Cox’s regression model was investigated using the Grambsch and Therneau algorithm,²³ and by visually inspecting the Schoenfeld residuals. The visual inspection is performed by plotting the residuals against time: this line should be straight, indicating no interaction between time and the studied variable, thus respecting the proportionality assumption.²³

The predictive ability of each score was evaluated by C-statistic (area under the Receiver Operating Characteristic (ROC) curve). This is a measure of the sensitivity and specificity of either a score or a logistic regression model. In this study sensitivity translated to how well the score/model correctly identified patients who were re-operated on, whereas specificity translated to how many were incorrectly identified as re-operated on. The C-statistic value can range between 0.5 and 1, where a high value indicates that the model has a good discriminative ability.

Results

We identified 134 423 THRs in 114 072 patients with primary OA (Table II). Of these, 1826 hips (1.4%) had been re-operated on within two years. The two main reasons for these early re-operations were dislocation and infection. From the original cohort, 118 065 THRs had been observed for more than two years, and the mean follow-up was 7.1 years (sd 3.2; interquartile range (IQR) 4.2 to 10.1). In this group 4244 hips (3.6%) had undergone re-operation. The main reason for surgery in the later period was aseptic loosening, and the second most common reason was surgery due to dislocation (data regarding dislocated hips without surgery were not available).

The presence of any comorbidity was found to increase over the study period. The prevalence was highest for the Elixhauser score, followed by the Charlson score and finally the RCS Charlson score (Fig. 2), in a reflection of the number of categories included within each (Table I).

Early re-operations

The risk for re-operation increased continuously in the RCS Charlson and Elixhauser scores (Fig. 3). The Elixhauser score was the only measure where both comorbidity groups were significantly different from the healthy group, and this was least affected by the confounding variables (Table III). None of the score estimates showed significant changes when only ICD-10 codes were used. The C-statistic for the best-performing measure, the Elixhauser score, was 0.52 (95% confidence interval (CI) 0.51 to 0.53) (the full model with all covariates 0.59). We also compared the results with an unweighted Charlson score,¹² and the difference in the estimates was negligible. The weighted van Walraven version¹³ of the Elixhauser score exhibited similar estimates to the unweighted score, although it lacked the continuous increase.

Late re-operations

Over the period between two and 12 years none of the scores had any significant impact on the re-operation rates. When we applied a non-weighted Charlson score, the category ≥ 3 had a close to significant estimate (adjusted hazard ratio (HR) 1.9 (95% CI 1.0 to 3.7); p = 0.053). This estimate also increased when only patients with ICD-10 codes were considered, with a HR of 2.6 (95% CI 1.1 to 6.2). The van Walraven version of the Elixhauser score did not perform better than the unweighted score.

Implant fixation and gender

Women had an overall lower re-operation rate in the first period (adjusted HR 0.68 (95% CI 0.62 to 0.75)) and also in the later period (adjusted HR 0.67 (95% CI 0.63 to 0.72)). During both periods we observed a difference in the effect of implant fixation between the genders. Women had a higher hazard ratio for all non-cemented implants compared with fully cemented devices, the only exception being reverse hybrids in the later period. In men the choice of fixation had no measurable influence during any of the periods (Table IV).

Age

The age estimate changed considerably between the two time periods. In the early period, zero to two years, the effect was negligible, whereas in the later period, two to 12 years, the risk for re-operation decreased with increasing age (Fig. 4).

Other analyses

We also looked at the different causes for re-operation over the 12-year period (Fig. 5). The early causes, such as dislocation, infection and fracture, exhibited similar patterns to the previous results. Aseptic loosening failed to correlate with the Elixhauser score (p = 0.299), the Charlson score (p = 0.645) or the RCS Charlson score (p = 0.5425, all ANOVA) (Fig. 5).

None of the scores violated the proportional hazards assumption during the period from zero to two years. In the later period, two to 12 years, the RCS Charlson score did violate the assumption. When inspecting the Schoenfeld residuals this violation affected only category 1 and 2, where there was a minor decline in the estimate over time.

During the first two years 4321 THRs (3.2%, 3.7% of patients) were censored due to death, and in the next ten years 23 087 THRs (19.6%, 20.9% of patients) were censored due to death. When re-analysing a similar model using the competing risk regression, thereby adjusting for these deaths, the estimates for the scores were marginally lower, but without any impact when comparing scores.

Discussion

The quality of these data is robust: the Swedish Hip Arthroplasty Register has coverage at a national level of 98.5%,²⁴ and the patient registry has a high sensitivity and specificity for primary diagnoses.²⁵ Of the three measures, the unweighted Elixhauser comorbidity score was closest to fulfilling the qualities that we sought in the first two years. In the later period none of the measures proved satisfactory, although patients with an unweighted Charlson score ≥ 3 did seem to have a worse outcome.

The main strengths of our study are that we analysed a homogeneous population and a large sample size. The choice of implant might be influenced by the patient’s comorbidities,²⁶ and therefore we believe it is vital to exclude less commonly used, unproven implants, and to adjust for the fixation type. By investigating only patients with primary OA and excluding uncommonly used implants, we believe that we were able to create an ideal population to evaluate the impact of the various comorbidities.

Age is closely related to comorbidity and often regarded as the most important comorbidity measure. In this study we used a spline for age instead of age groups, as commonly used^27-29; this minimises residual confounding for age and reduces the risk for measurement error or cut-point bias.²²

Our study has limitations. We had no information concerning technical outcome, patient compliance or surgeon’s experience. Although we stratified for type of hospital and surgical volume, we lacked detailed knowledge about the surgeons’ experience, a factor that is undoubtedly important.^30-32 Also, we probably failed to identify all comorbidities owing to the unrefined nature of registry data.

It is noteworthy that the estimates were unchanged when we looked at the ICD-10 subgroup: these patients were studied during a later period with a higher prevalence of comorbidity (Fig. 2). We interpreted this increase in the later study period as being due to improved registration of diagnoses. This is supported by the fact that the Elixhauser score, with less severe comorbidity groups, increased more than the other scores, whereas the age distribution remained unchanged during the study period.

Patients with comorbidities might also be subjected to a surveillance bias. These patients might be more familiar with hospital settings and seek medical attention sooner, which affects the estimates. Although this effect is hard to rule out, we believe that the clinical problems leading to further surgery are likely to bring the majority of patients to seek medical attention.

In this study we looked at re-operation rates, and it is important to recognise that this is just one of many possible outcome measures after THR. Other outcomes, such as quality of life, non-surgically treated infections or dislocations, or early medical readmissions are also important and require separate validation.

It is also important to note that re-operation is a wider definition of failure than revision. Previous studies using revision as an outcome^16,27,28 have found similar estimates for the Charlson score. This strengthens the external validity of our findings.

We constructed the Elixhauser comorbidity score simply by counting disease categories. Elixhauser et al⁸ questioned this approach, as the impact of different disease categories may have very different effects on overall outcome. We believe that using a score instead of enumerating all the 30 comorbidities reduces the risk of overfitting the model. We investigated an alternative weighted Elixhauser score, as suggested by van Walraven et al,¹³ but this did not outperform the unweighted score. Combined with the similarity of the different Charlson scores, there may be a lack of difference between weighted and unweighted scores. It is possible that this could be due to a surgeon coding bias: for example, a more clinically demanding patient will probably have a more rigorous pre-operative preparation and thus also a more complete coding of the comorbidities.

We believe that predicting arthroplasty outcome from comorbidity scores may prove to be extremely difficult despite the expediential increase in publications on the subject between 1998 and 2012 (Fig. 1). Our C-statistic of 0.52 for the best score indicates a poor ability to differentiate and although our complete model had a C-statistic of almost 0.6, this is very different from published mortality models that range between 0.7 and 0.9.^12,13,33 These results are cause for caution when interpreting comorbidity-adjusted results, especially when the scores have not been validated for the studied outcome.

In conclusion, we failed to validate any of the scores in terms of predicting re-operation after THR. The Elixhauser score may be useful for estimating the comorbidities relevant to the risk of re-operation within two years, but none showed any convincing results in the period between two and 12 years.

The authors would like to thank Professor A. Odén for his invaluable aid in choosing the statistical analyses, and Dr M. Kay for her help in making the article more readable. They also wish to thank all of the Swedish orthopaedic departments for kindly supplying these high-quality data.

No benefits in any form have been received or will be received from a commercial party related directly or indirectly to the subject of this article.

This article was primary edited by D. Rowley and first proof edited by G. Scott.