The research, led by Giulia Zumbo and María-Dolores del-Toro-López of the Hospital Universitario Virgen Macarena in Seville, together with collaborators across eight Spanish hospitals, is published in the Journal of Bone and Joint Infection. The team assembled a retrospective cohort of 224 adults with prosthetic joint infection treated with DAIR between 2006 and 2023, each followed for a minimum of twelve months after surgery. The stakes of getting this prediction problem right are considerable. Prosthetic joint infection affects between 0.5 and 2.3 percent of patients within two years of a primary joint replacement, and those figures climb to between roughly 5.5 and 8.4 percent after revision hip or knee surgery. With millions of arthroplasties performed worldwide each year, even small improvements in patient selection could translate into thousands of avoided reoperations.

The study’s first target was the KLIC score, a prediction tool developed in 2015 that assigns points for kidney dysfunction, liver disease, index surgery, a cemented prosthesis, and elevated C-reactive protein. KLIC was derived from a cohort of 222 early postoperative infections and designed to predict failure within sixty days of DAIR, when the observed failure rate was just 23.4 percent. In the new Spanish cohort — older in follow-up horizon and different in case mix — KLIC collapsed. Its area under the receiver operating characteristic curve, the standard measure of a diagnostic or prognostic model’s ability to separate those who fail from those who succeed, was 0.53, with a confidence interval spanning 0.44 to 0.64. An AUC of 0.5 represents discrimination no better than random chance. The finding aligns with a string of previous external validation studies that reported AUCs ranging from 0.53 to 0.76, suggesting that KLIC’s performance depends heavily on how closely the population and outcome horizon resemble its original derivation setting.

The contrast with the newly developed model could hardly be sharper. Using a generalized linear mixed-effects model with hospital as a random intercept — a statistical technique that accounts for clustering of outcomes within centres — the researchers identified five independent predictors of DAIR failure, all measurable before surgery. These were a revision prosthesis rather than a primary implant, a Charlson Comorbidity Index of three or higher, a C-reactive protein level above 150 milligrams per liter, a longer interval between the index arthroplasty and the DAIR procedure, and infection with Gram-negative bacilli or multiple microbial species. The full model achieved an AUC-ROC of 0.85, with a 95 percent confidence interval of 0.79 to 0.90, and bootstrap internal validation with one thousand resamples confirmed limited optimism, yielding an optimism-corrected AUC of 0.83 and a calibration slope of 0.91 — indicators that the model is neither overfitted nor systematically miscalibrated.

Translating the statistical model into something a clinician could compute at the bedside, the team derived a simple point score: three points for a revision prosthesis, five points for a Charlson index of three or more, four points for C-reactive protein above 150 milligrams per liter, four points for Gram-negative or polymicrobial infection, and one point for every twenty-five days elapsed between the original joint replacement and the DAIR procedure. When patients were grouped by score, the observed failure rates separated dramatically: 25.5 percent in the low-risk group, 45.6 percent in the intermediate group, and a striking 87.9 percent in the high-risk group, a difference that was highly statistically significant. The score itself retained good discrimination, with an AUC of 0.81, only marginally below the full model.

The microbiological component proved particularly informative. Infections caused by Gram-negative bacilli carried an observed failure rate of 66.7 percent, and polymicrobial infections failed 68.0 percent of the time, compared with the overall cohort failure rate of 54 percent. When the researchers removed microbiological aetiology from the model entirely — simulating the common clinical scenario in which culture results are not yet available when surgery is planned — discrimination fell from 0.85 to 0.81 for the multivariable model and from 0.81 to 0.77 for the corresponding score. This suggests that while waiting for microbiology costs some predictive power, the clinical variables alone still carry substantial prognostic information, an important practical consideration for surgeons who must decide on a surgical strategy before cultures finalize.

The authors are careful to frame these variables as prognostic markers rather than causal determinants of failure. The association between longer time to DAIR and worse outcomes, for instance, is consistent with previous studies, but it may capture differences in clinical presentation, diagnostic complexity, or referral pathways rather than a direct effect of treatment delay itself. Similarly, the Charlson threshold of three was selected during model development after exploring alternative parameterizations, and the subgroup of patients with high comorbidity burden was small — just 23 patients in the overall cohort — producing a wide confidence interval around the effect estimate. The C-reactive protein cutoff of 150 milligrams per liter is supported by earlier international work on DAIR outcomes, but because alternative cut-points were explored within this cohort, its performance also requires independent confirmation.

The study’s methodology reflects modern standards for clinical prediction research. The team followed the TRIPOD+AI 2024 reporting guideline, used multiple imputation by chained equations to test the robustness of their complete-case analysis against missing C-reactive protein values in 38 patients, and ran sensitivity analyses showing that failure rates were similar across the 2006–2014 and 2015–2023 recruitment eras, with adjustment for calendar period leaving the predictor estimates essentially unchanged. The composite outcome definition was deliberately broad, counting as failure any persistence or recurrence of infection symptoms, any need for further revision surgery, any requirement for long-term suppressive antibiotics, or infection-related death. That breadth partly explains why the overall failure rate of 54 percent exceeds many figures in the literature, and the authors caution that differences in outcome definitions and follow-up duration must be considered when comparing success rates across studies.

Important limitations remain before this tool reaches the clinic. The retrospective design may have introduced selection and information bias, the risk thresholds were derived within the same cohort they were applied to, and — critically — the model has not yet been validated in an independent population. The study also cannot answer the question that matters most for the highest-risk patients: whether they would fare better with a one-stage or two-stage prosthesis exchange instead of DAIR. The model estimates failure risk conditional on DAIR being performed; it does not establish that switching strategies would improve outcomes for those flagged as high risk. Nor does it account for intraoperative findings or the quality of surgical and antimicrobial management, both of which influence results. The authors also note that potential differences in model performance across sociodemographic groups — a question of algorithmic fairness — should be examined during external validation.

Even with those caveats, the implications are significant. If external validation in independent cohorts confirms these findings, the score could complement clinical judgment by quantifying baseline failure risk at the moment a treatment strategy is chosen, helping surgeons and patients weigh the trade-off between the lower morbidity of implant retention and the higher certainty of exchange. For now, the message from the Spanish cohort is clear and somewhat uncomfortable for the field: a decade-old prediction tool widely cited in guidelines and reviews may be offering false reassurance, while a handful of variables already sitting in every patient’s chart — comorbidity burden, inflammatory markers, implant history, timing, and microbiology — hold the information needed to see the risk clearly. The researchers emphasize that the observed risk strata should not yet be used as treatment-selection thresholds, but the path toward a validated, practical decision aid for one of orthopedic surgery’s hardest calls has just become considerably shorter.

Subject of Research: Prediction of failure of debridement, antibiotics, and implant retention in prosthetic joint infection

Article Title: External validation of the KLIC score and development of a novel prediction model for debridement, antibiotics, and implant retention failure in prosthetic joint infection: a multicentre cohort study

Article References: Zumbo, G., Martínez-Alemany, I., Bravo-Ferrer, J. M., Nieto Diaz De Los Bernardos, I., Praena, J., Del Arco, A., Nuño, E., Corzo Delgado, J. E., Brun, F., Natera, C., Romero Palacios, A., Sobrino, B., Rodríguez-Baño, J., & del-Toro-López, M.-D. (2026). External validation of the KLIC score and development of a novel prediction model for debridement, antibiotics, and implant retention failure in prosthetic joint infection: a multicentre cohort study. Journal of Bone and Joint Infection, 11(5), 591-599. https://doi.org/10.5194/jbji-11-591-2026

Image Credits: AI Generated

DOI: 10.5194/jbji-11-591-2026

Keywords: prosthetic joint infection, DAIR, KLIC score, prediction model, arthroplasty, risk stratification, C-reactive protein, orthopedic surgery, external validation, clinical prediction, implant retention, multicentre cohort study