Data

Hospital Episodes Statistics (HES) is the national hospital administrative database for England, covering all NHS (public) hospitals. Each row of the database covers a “finished consultant episode”, the period during which a patient is under the care of a senior physician (consultant). Using HES’s own patient identifier “HESID”, which involves a hierarchy of their NHS Number (unique to each person) where given and their date of birth, sex and postcode if not given, we linked records into admissions and linked admissions ending in transfer to another hospital into “superspells”; we analysed superspells and will refer to them as simply “admissions”. Outcome data were assumed to be complete for every patient (or at least assumed to be missing completely at random). With a national database, the main losses to follow-up occurs when patients move or are treated abroad. HES has 20 diagnosis fields, which use ICD10 coding, and 24 procedure fields, which uses the UK’s own “OPCS” coding system. The diagnosis fields capture comorbidities and complications. Each admission and procedure is dated. Urgency is indicated by the method of admission field (elective or emergency). We extracted records with discharge dates in the financial years 2006/7 to 2013/14 inclusive. Further information is given on NHS Digital’s website (https://digital.nhs.uk/).

Definition of index and revision procedures

Index elective (planned) hip and knee replacements (HR, KR) for any indication were identified using the method of admission field and any of the OPCS codes given in the Appendix in any procedure field. Index procedures were those with any of these codes between April 2009 and March 2014 in order to allow three full years of follow-up for all patients. Revisions were defined as any of the list of OPCS codes in the NJR’s definition in their report (S1 Table). The few missing procedure dates were taken to be the date of admission. Joint laterality was recorded in 99% of index procedures; where this was unknown, we assumed that the revision related to the most recent joint replacement. Revisions were included at any time between one day and three years after the index procedure. Revisions were matched on laterality to the index.

Statistical analysis

Our previous work comparing two popular comorbidity indices, Charlson [20] and Elixhauser [21], recommended using the Elixhauser set plus dementia [22,23]. We therefore first tabulated each of this combined set of comorbidities and excluded those with a prevalence of <0.1%. We then built three-level multiple logistic regression models using the remaining variables plus age group, sex, Carstairs socio-economic deprivation fifth, number of admissions in the year before the index operation, with patient, surgeon and hospital as the three levels [24]. To aid model fitting, consultant codes with five or fewer procedures between 2009/10 and 2013/14 were combined into one new pseudoconsultant. Likewise, we combined hospital trusts fewer than 50 procedures during that time into one new pseudohospital. Model fit was checked using global goodness-of-fit tests and by inspecting residuals and standard errors. The effect of mortality was checked by running cause-specific survival analysis with death as the competing event for revision and found to be limited.

The population attributable fraction was calculated for each covariate with a significantly (p<0.05) elevated odds ratio using the formula by Levine [25]; adjusted odds ratios stood in for relative risks, for which they will be a good approximation given that the outcome here is rare.

Hip replacement

Patients aged under 60 were the most likely to need an early revision, though PAFs were low. Males were 19% more likely than females to need earlier revision. The PAF for males was larger than for age at 7.2%.

Resurfacing was nearly twice as likely to require early revision than cemented (OR 1.82). Cementless HR had a 34% greater odds. Partial HR had higher odds, though a partial cemented replacement was not statistically different from a total cemented. Resurfacing had a PAF of 1.6% and uncemented HR had a PAF of 13.8%.

Other predictors of early revision were primary diagnosis of the index admission, previous all-cause emergency admissions and twelve comorbidities–all with PAFs of at most 2.5%. For the primary diagnosis of the index admission, significantly higher odds were seen for polyarthrosis, rheumatoid arthritis, trauma (fractured neck of femur) and other, the latter largely explained by osteonecrosis and non-union of fracture (pseudoarthrosis). Socio-economic status was not significant at p<0.05.

Knee replacement

Patients aged under 60 were most likely to need an early revision. Those aged 40 to 44 years had the highest odds ratio at 1.63, but the PAF was highest for those aged 55 to 59 (1.2%). Males were 24% more likely than women to need early revision (PAF 7.1%).

Patellofemoral joint replacement conferred a significantly greater risk of early revision (OR 5.16, and the largest PAF of 7.6%), as did a partial knee replacement (OR 2.81, PAF 9.4%).

Other predictors of early revision were primary diagnosis of the index admission, previous all-cause emergency admissions, socioeconomic status and six comorbidities, with a lower odds for renal disease (OR 0.82). For the primary diagnosis of the index admission, significantly higher odds were seen for the “other” group only (compared with gonarthrosis). This was largely explained by the ICD10 code M2556 (pain in knee, with revision rate 2.5%) and several T84 codes (complications of internal orthopaedic prostheses, implants and grafts, mostly “other” and “mechanical” complications, with revision rates of around 20%). PAFs for these predictors were small.

Comparison with other studies

Younger age in our study was associated with a greater risk for early revision in both hip and knee replacement surgery, consistent with other studies [26–30]. This has partly been thought to be due to polyethylene wear causing loosening in younger, more active patients, although this usually results in late wear [8,31]. On the other hand, confounding variables such as co-existing diagnoses may explain the inclusion of this factor. Age should be a factor included in the conversation when patients are counselled about the risk of early revision in primary arthroplasty. At a population level, the PAFs indicate that this is of modest importance.

Like ours, prior studies have identified males as carrying a higher risk of early revision [14,26,27]. This has been attributed to a higher rate of dislocations in males than females for HR [27]. There is no such apparent explanation with regards to early revision in males undergoing KR. One study has contextualised it with KR being increasingly performed in younger males [14]. It is recommended that this risk be explained in the consent process. From a population perspective, sex has an important impact, second only to procedure type.

Regarding the procedure subtype, Australian and New Zealand databases have similarly shown significantly higher rates of revision in partial KR [32,33]. On the other hand, a study using NJR data found smaller differences between the revision rates of partial KR versus Total Knee Replacement [34]. The authors believed this to be due primarily to progression of the degenerative process. Other reasons could include poor patient selection for partial KR and the threshold for revision for partial KR being lower than that for total KR as it is easier to revise.

In HR, we found that resurfacing and uncemented HR had higher revision rates than cemented HR, similar to a Swedish study, which found poorer performance of uncemented cups and unrecognised intraoperative femoral fractures [28]. Kandala et al was a UK-based study using NJR data that also confirms that cementless hip replacements needed earlier revision than cemented prostheses [35]. There is a large variation in practice, particularly in Nordic registries, with likely reasons including a lack of randomised controlled trial evidence, lenient regulatory requirements and intensive marketing [35]. Furthermore, Conroy et al, who investigated the Australian Joint Registry for those factors associated with revision for dislocation following primary HR, showed the same results as our study. They demonstrated a relative risk of 2 for fractured neck of femur and rheumatoid arthritis and 1.5 for avascular necrosis [36].

From the population perspective based on our data, the PAF was 13.8% for the uncemented prosthesis and 1.6% for resurfacing. This means that preventative measures at the population level should be targeted at uncemented arthroplasty. On the other hand, individual surgeons may have satisfactory outcomes with the type of HR to which they have become accustomed. Therefore, a change in practice might lead to worse outcomes.

Our study found higher rates of early revision in KR in areas with the lowest socioeconomic status at p<0.05. Dy et al also found this to be the case for KR but not HR [13,14]. Their finding suggested the need for higher levels of social support in impoverished areas to improve post-operative outcomes. However, others have demonstrated increased rates of revision in higher education levels [37]. This was thought to be due to more post-operative visits amongst this group for discussion of their progress and hence potentially earlier identification of patients who would benefit from revision [38,39].

Patient co-morbidities are perhaps the most relevant for the surgeon to consider when offering primary arthroplasty. Of our comorbidities that were highly significantly (p<0.01) associated with early revision, only Parkinson’s disease was common to both HR and KR. The higher number of significant comorbidities in HR could be due to the higher risk of dislocation in cases of neurological disorders, alcohol and drug abuse, hence the need for early revision. The outcome rates for the two joints were similar and the number of procedures for KR was higher than for HR, which argues against statistical artefact as an explanation for this difference. Depression has also been linked to early revision in other studies, thought to be due to lower levels of satisfaction and pain relief postoperatively[40]: we found as association for HR but not for KR regarding depression.

It is interesting that diabetes was not found to be relevant in HR, but significant in KR, while the converse was true for obesity. It is possible that the differences do exist but we were unable to detect them due to how these factors were recorded. Obesity in our database was dichotomous, i.e. either recorded or not, and dependent on ICD10 coding. The BMI was not recorded, which would have offered greater statistical power to detect associations. Diabetes was dichotomous for the same reason, with HbA1c lacking in the database, but the type and level of glycaemic control can determine the risk of infection and need for early revision [13,14]. Another research group also found obesity to be only linked to an increased risk in the HR population and explained this difference in a similar fashion [15,16]. The causal mechanism is thought to be the greater biomechanical load exerted in obese patients leading to early failure and an increased risk of infection in poorly controlled diabetic patients [14,41].

From a population perspective, we found that comorbidities are less important than the type of operation. The most relevant were obesity in HR (PAF 2.5%) and chronic lung disease in KR (0.8%). Although this is of modest importance at the population level, an awareness of these comorbidities is crucial, particularly in the pre-assessment clinic and the consent process when explaining to the patient their risk of early revision. It enables better patient selection as both the patient and the surgeon are informed of the relevant factors and the degree to which they are important in the need for early revision. It also allows targeted pre-operative optimisation of patients with these particular disorders.

Finally, in both HR and KR, early revision was likelier with increasing numbers of pre-operative emergency admissions. These could relate to the hip or knee pain, an unrelated new issue or a known comorbidity. This factor has to our knowledge not been identified in previous studies. It could reflect a combination of the previously mentioned factors, as patients with comorbidities are more likely to need to present to the emergency department. It could relate to various other patient characteristics, e.g. lower satisfaction levels [37,40]. Omitting this variable from the models had little impact on the size of the odds ratios, indicating an effect independent of our set of comorbidities.

Limitations

Our study carries limitations inherent to the use of national administrative data. The accuracy of the primary diagnosis and procedure fields in HES is very high at 96% [42], and there are financial incentives to capture every admission. Secondary diagnosis coding is known to be less complete and vary by comorbidity, though studies of the relation between comorbidity and outcomes have shown that this underrecording does not lead to bias in the odds ratios [43]. We have already mentioned the lack of BMI and HbA1c. Joint laterality was 99% complete, but with administrative data, it is difficult to reflect case complexity. We had to exclude a small proportion of records due to invalid consultant codes but have no reason to think that the relations between predictors and early revision would differ greatly between those included and those excluded.

Unlike some administrative data sets in North America, for instance, HES data lack present-on-admission (POA) flags, i.e. they cannot distinguish between comorbidities and complications. While we tried to minimise this issue by using dementia, Parkinson’s disease plus the Elixhauser set (devised specifically for databases with no POA information), a minority of our flags such as renal disease and fluid and electrolyte disturbances could have occurred after the index procedure for some patients. This would limit their usefulness in pre-operative management.

We could not assess factors such as operative times and hospital volume, although it could be argued that in the case of the latter there is a more uniform distribution due to a nationalised health service in the UK [14,44]. Early revision is not always a true indicator of failure of primary procedure as patients may not seek advice or be offered revision. They may decline from other comorbidities and be unable to undergo the procedure, which is not reflected in our study. Analysis of the failure rate at specific post-operative intervals would be helpful, as the cause of failure within 30 days may be very different from that at 3 years. Furthermore, our data do not describe the extent to which each clinician excluded extra-articular causes of postoperative pain such as vascular pathologies, bursitis, iliotibial band syndrome and psychological illness, which have previously been described in patients following primary arthroplasty [45].

Finally, as there is evidence that revision due to dislocation or infection may have different risk factors [46] our analysis could be extended to examine the risk factors for different primary diagnoses for the revision admission separately.