Abstract
Background
National joint replacement registries inform prosthesis selection and health policy using a consistently recorded set of patient, implant, and surgical variables. However, preoperative radiographic alignment and detailed diagnostic complexity are rarely, if ever, captured. Surgeons pay close attention to limb alignment and deformity because both reflect the complexity of the knee being replaced and are associated with the risk of failure. However, the incremental value of recording these measurements, beyond the variables that registries already collect, has not been quantified. As preoperative alignment can be read from the same long-leg radiographs that many surgeons already obtain, even a small improvement in identifying which knees will be revised could justify the modest cost of recording it.
Questions/purposes
In a 15-year, single-center, TKA group with comprehensive preoperative alignment assessments and national joint replacement registry linkage, we asked: (1) How well do the variables that a national joint replacement registry routinely records identify which knees will later be revised (as quantified by a metric called discrimination, where 0.50 is no better than chance and 1.00 is perfect prediction)? (2) Does adding diagnostic-complexity variables improve that discrimination? (3) Does adding preoperative alignment variables improve it further? (4) When alignment is left out, by how much do the risk estimates that registries report for common surgical decisions (their HRs) change?
Methods
A retrospective analysis of 2341 primary TKAs in 1917 patients performed between January 2010 and May 2024 at a single tertiary center was conducted, all with national-registry linkage and a documented preoperative long-leg standing radiograph. The primary outcome was aseptic revision, modeled with nested Fine-Gray regression and death treated as a competing risk, meaning that a patient who dies can no longer be revised and is not simply removed from the denominator. To check that the findings were not an artifact of incomplete imaging, we repeated the entire analysis in a more completely imaged group (a prespecified sensitivity analysis), which was a subset of 1567 procedures with both preoperative and postoperative long-leg radiographs (the paired-imaging subset). We built the models (M) stepwise to determine how much each kind of information added. The M1 used the variables that a registry routinely records, M2 added three diagnostic-complexity variables, and M3 added three preoperative alignment variables. To measure how much each group of variables contributed, we compared neighboring models statistically, using an intermediate model (M2.5) to isolate the share attributable to limb alignment. We measured how well each model separated the knees that were later revised from those that were not using a score called the concordance index (C-index): 0.50 means that the model is no better than a coin toss, 1.00 means it is always right, and 0.66 means it ranks a revised knee above a nonrevised one about two-thirds of the time. Because any model flatters itself on the same knees that were used to build it, we also tested each model on knees it had not seen by rebuilding it on part of the data and checking it on the rest, repeatedly, keeping both knees of a patient together. That cross-validated C-index is the honest figure and the one we rely on. The in-sample figure is reported alongside it because earlier studies report it, and the gap between the two shows how much of the apparent gain is optimism rather than real predictive information.
Results
The registry-only variable set (M1) identified the risk of aseptic revision only modestly (apparent C-index 0.66 in the primary group and 0.67 in the sensitivity group; cross-validated 0.59 and 0.61). Adding diagnostic-complexity variables (M2) did not improve this. Adding preoperative alignment (M3) raised apparent discrimination to 0.68 in the primary group and 0.69 in the sensitivity group, but it did not improve cross-validated discrimination in either. Taking everything the models added from M1 to M3, that addition was detectable as a whole only in the sensitivity group, not in the primary group. Within it, limb alignment carried most of the information in the sensitivity group (68%; p 180° is valgus [95% confidence interval (CI) 1.00 to 1.12]; p = 0.03); the same association was present but not statistically distinguishable from the effects of chance in the primary group (subdistribution HR 1.04 [95% CI 0.99 to 1.09]; p = 0.11). Leaving alignment out changed some registry-style HRs. For example, the HR for tibial stem use shifted by 17% in the primary group. The findings were consistent, and stronger, in the more completely imaged subset (n = 1567; 68 revisions). When alignment was added, about 30% of knees in the primary group and 34% in the subset moved into a different third of predicted risk.
Conclusion
The variables that registries routinely record identify the risk of aseptic TKA revision only modestly. Preoperative limb alignment adds a small but consistent and clinically interpretable amount of predictive information, and it is stronger in the subset of knees for which both preoperative and postoperative imaging existed, although we did not test postoperative alignment as a predictor in its own right. Leaving alignment out also shifts the registry-style HRs that surgeons and policymakers rely on for some surgical decisions. Registries should therefore consider recording at least one preoperative alignment field. In practice, this could begin as a simple category (varus, neutral, or valgus) read from films that surgeons already obtain, beginning as a pilot before any wider rollout, because even partial recording would add value.
Level of Evidence
Level III, therapeutic study.