Prediction of Early Periprosthetic Joint Infection After Total Hip Arthroplasty

Abstract

Purpose: To develop a parsimonious risk prediction model for periprosthetic joint infection (PJI) within 90 days after total hip arthroplasty (THA).

Patients and methods: We used logistic LASSO regression with bootstrap ranking to develop a risk prediction model for PJI within 90 days based on a Swedish cohort of 88,830 patients with elective THA 2008-2015. The model was externally validated on a Danish cohort with 18,854 patients.

Results: Incidence of PJI was 2.45% in Sweden and 2.17% in Denmark. A model with the underlying diagnosis for THA, body mass index (BMI), American Society for Anesthesiologists (ASA) class, sex, age, and the presence of five defined comorbidities had an area under the curve (AUC) of 0.68 (95% CI: 0.66 to 0.69) in Sweden and 0.66 (95% CI: 0.64 to 0.69) in Denmark. This was superior to traditional models based on ASA class, Charlson, Elixhauser, or the Rx Risk V comorbidity indices. Internal calibration was good for predicted probabilities up to 10%.

Conclusion: A new PJI prediction model based on easily accessible data available before THA was developed and externally validated. The model had superior discriminatory ability compared to ASA class alone or more complex comorbidity indices and had good calibration. We provide a web-based calculator (https://erikbulow.shinyapps.io/thamortpred/) to facilitate shared decision making by patients and surgeons.

Keywords: clinical decision-making tool; external validation; orthopaedics; prediction model; total hip arthroplasty; web calculator.

Figure 1

Flowchart with inclusion criteria and number of patients. Data from the Swedish Hip Arthroplasty Register were used for model derivation and internal validation.

Figure 2

Flowchart with inclusion criteria and number of patients. Data from the Danish Hip Arthroplasty Register were used for external model validation.

Figure 3

Receiver operating characteristics (ROC) curves combines sensitivity and specificity to illustrate discriminative abilities of the different models. The main and reduced models performed almost identical for prediction of periprosthetic joint infection (PJI) within 90 days after surgery. They both performed better than all other models. Area under the curve (AUC) are stated for each curve within parenthesis.

Figure 4

Area under the receiver operating characteristics curve (AUC) as a measure of predictive discriminative ability with 95% confidence intervals. The reduced model performed no different than the main model on the Swedish data, and both of these models performed better than all other models (left panel). Similar models were fitted to the Danish cohort (right panel). The reduced model (SE) with coefficients based on the Swedish data, performed almost as good as the reduced model (DK) with coefficient values refitted to the Danish cohort.

Figure 5

Calibration for the reduced model. Proportions of periprosthetic joint infection (PJI) above 10% were rarely observed and are therefore omitted. Internal calibration (95% confidence band between the green lines) was good; predicted probabilities were similar to observed proportions, as indicated by close proximity to the diagonal line. External calibration with the same model (red) indicated some over-estimation and less accuracy. This was expected due to smaller sample size in the Danish cohort, and due to national differences. Calibration improved after re-calibration of the model intercept to better resemble the Danish base incidence (blue). DK, Denmark; SE, Sweden.

Figure 6

Patients with observed periprosthetic joint infection (PJI) had, on average, higher predicted probabilities for this adverse event, compared to patients with no PJI. The x-axis is log-transformed for visual clarity and each curve was normalized on the y-axis which therefore has no direct interpretation.