The modified-KLICC score: a novel tool to predict outcomes following debridement, antibiotics, and implant retention after early acute periprosthetic hip infection

Abstract

Aims

Two preoperative risk models have been designed to predict debridement, antibiotics, and implant retention (DAIR) failure: KLICC and CRIME-80 scores. However, external validation of both scores is scarce. We aimed to validate these scores in an external cohort and to create a new model with additional risk factors.

Methods

We retrospectively evaluated 96 patients with early acute periprosthetic hip infection treated with DAIR. At a two-year cut-off, failure was defined as the need for second DAIR, implant removal, or 90-day infection-related death. Association between demographic variables and failures was tested. The model discriminatory performance was measured using the time-dependent receiver operating characteristic (ROC) curve and Harrell concordance index (C-index). The ‘calibration in the large’ (CITL) was calculated as the logistic regression model intercept. A modified KLICC score was created by adding the variable time from onset of symptoms to DAIR.

Results

The 24-month cumulative incidence of failure was 23.96% (95% CI 15.9 to 32.8). KLICC’s area under receiver operating characteristic (AUROC) was 0.79 (95% CI 0.67 to 0.90), with a CITL of -0.57 (95% CI -1.16 to -0.01) and a slope of 0.68 (95% CI 0.35 to 1.02). CRIME-80’s AUROC was 0.63 (95% CI 0.51 to 0.76), with a CITL of -1.66 (95% CI -2.13 to -1.19) and a slope of 0.35 (95% CI -0.14 to 0.85). The difference between both AUROCs was statistically significant (p = 0.0138), with the KLICC score performing better. As compared with the original KLICC score, the modified-KLICC improved the AUROC to 0.85 and the beta-slope and α intercept to 1.24 and -0.07, respectively (p = 0.020).

Conclusion

KLICC was superior to CRIME-80 in predicting DAIR failure. The modified-KLICC score improved the model prediction and could be useful to help indicate alternatives to DAIR when the predictive failure is high.

Cite this article: Bone Jt Open 2025;6(12):1532–1541.

Introduction

Surgical intervention and antibiotherapy are the mainstays of therapeutic management in patients with periprosthetic joint infection (PJI).^1-3 Two-stage revision is, for many, the gold standard for treatment of chronic PJI, although one-stage exchange is becoming a more popular treatment option when certain selection criteria are met, with the advantage of cost-containment and improved functional outcomes.⁴ However, for acute PJIs, including Tsukayama types II (acute postoperative) and III (acute haematogenous) infections,⁵ the treatment of choice is debridement, antibiotics, and implant retention (DAIR). This protocol consists of removing all necrotic, infected tissue, irrigating to dilute and decrease bacterial inoculum, and providing appropriate antibiotic therapy while preserving the implants.⁶ The success rate of this procedure remains unclear, ranging from 14% to 95%, and the variability in the reported results is explained by the heterogeneity of the outcome tools and the lack of clear indication criteria.^2,3,7 Diverse factors may influence DAIR outcomes, including immune status of the host, surgical technique and approach, virulence and biofilm-forming capacity of the causative organism/s, time from onset of symptoms to DAIR, arthroplasty of modular components, and type of antibiotherapy, among others.^8-14

For this matter, there is a need to identify which factors may predict success or failure of a DAIR procedure in order to avoid unnecessary expenses and convey clear treatment expectations to patients. It has been shown that results of one- or two-stage revision surgeries for PJI in patients with previous DAIR are controversial, with some reports suggesting higher subsequent failure.^9,15-17

Several predictive models have been developed to help identify risk factors for DAIR failure and, thus, to stratify groups with different risk levels. The two most popular models are the KLICC and CRIME-80 scores, developed for Tsukayama types II and III, respectively.^18,19 The KLICC-score includes the following variables: renal failure (K); liver failure (L); index surgery different from primary total hip arthroplasty (THA) (I); cemented fixation (C); and serum CRP value > 115 mg/l (C).¹⁸ Tornero et al¹⁸ assigned each variable a score corresponding to the relative odds risk (OR) for failure (K = 2; L = 1.5; I = 1.5; C = 2; C = 2.5). On the other hand, the CRIME-80 involves the following variables: chronic obstructive pulmonary disease (COPD) (C), serum CRP value > 150 mg/l (C), rheumatoid arthritis (RA), index surgery due to hip fracture (I), male sex (M), exchange of modular components (EMC), and aged > 80 years (80). As with the KLICC, each risk factor has a value according to the calculated OR (C = 2; C = 1; R = 3; I = 3; M = 1; E=-1; 80 = 2).^19,20 Based on the sum of these values, both scores help stratify risk groups for DAIR failure, suggesting a one- or two-stage revision in case of high risk (KLICC ≥ 4 or CRIME-80 ≥ 3).

Both predictive models have been described as promising in both the 2015 European Bone and Joint Infection Society (EBJIS) meeting¹⁸ and 2018 International Consensus Meeting (ICM).¹² However, when subjected to external validation, their predictive value decreased, losing clinical applicability when stratifying at-risk groups.^11,12,21 Recently, Shohat et al,³ using machine learning, developed a novel algorithm to improve decision-making in acute PJI cases, with acceptable results when externally validated.¹⁰ Nonetheless, there is no uniformity in the application of artificial intelligence-based models, nor have they been validated by the ICM, EBJIS, or Infectious Diseases Society of America (IDSA).²²

The specific regional epidemiology, as well as the diverse clinical practice of different institutions, considering both social determinants and medical resources, are necessary to adjust for in a correct application and interpretation of predictive models. Therefore, the aim of this study was to perform an external validation of the KLICC and CRIME-80 scores in patients with acute postoperative PJI of the hip treated in a tertiary care centre in Argentina, and to develop a new predictive model adjusted to the characteristics of the studied population.

Methods

Study population and design

This was a retrospective cohort study, approved by the Institutional Review Board of Hospital Italiano Buenos Aires, Argentina (IRB00010193, protocol #6341), of all patients with acute PJI of the hip treated with DAIR at our institution between January 2010 and December 2020. Initial screening was performed through the institutional electronic medical database (digitalized since 2010) using specific keywords in the surgical record. We included all patients aged over 18 years with hip arthroplasty diagnosed with acute PJI, according to the EBJIS classification,¹⁸ treated with DAIR. To classify the infection as acute, the Tsukayama classification was used, where types II (occurring six weeks immediately postoperatively) and III (acute haematogenous, where the onset of symptoms is acute and after six weeks postoperatively) are considered as acute.⁵ From a total of 247 PJIs, we excluded those whose index surgery had been one- or two-stage revision surgery (n = 97), those initially treated at another centre (n = 15), or had less than two years of follow-up still alive at the time of evaluation (n = 19). Patients with bilateral PJI (n = 2) and those with incomplete medical records (n = 18) were also excluded. The final cohort consisted of 96 patients, including 48 females and 48 males (Figure 1).

Fig. 1.

Flowchart of patient selection. DAIR, debridement, antibiotics, and implant retention; PJI, periprosthetic joint infection.

Demographic variables collected were age, sex, laterality, BMI, history of diabetes mellitus (DM), rheumatoid arthritis (RA), COPD, renal failure, liver cirrhosis, immunosuppression, Charlson Comorbidity Index (CCI),²³ American Society of Anesthesia (ASA) classification,²⁴ index diagnosis (femoral neck fracture versus osteoarthritis), type of fixation (in cases of cemented or hybrid prostheses the use and type of antibiotic in the cement was also recorded), the bearing surface, the time from onset of symptoms to DAIR that the patients reported as first having symptoms as documented on the electronic record (measured in days), the time from the index surgery to DAIR (measured in days between both procedures), and death (and cause). Infectiological variables were use of antibiotics prior to DAIR (including type and timing, measured in days), preoperative serum leucocytes value, preoperative serum polymorphonuclears (PMN) value, preoperative serum lymphocytes value, preoperative serum CRP value, preoperative serum ESR value, preoperative serum platelet value, ratio of absolute serum polymorphonuclears:lymphocytes values, and presence of positive blood culture before DAIR. Patients were categorized as immunocompromised if they had history of rheumatoid arthritis, active oncological disease, active corticosteroid therapy longer than eight weeks in duration, diabetes with target organ damage, history of splenectomy, and those with inherited or acquired immune system pathology (e.g. solid-organ transplant-recipients).²⁵

Patient characteristics

The series included 48 females and 48 males, with a mean follow-up of 54.36 months (SD 29.8). The mean age of the study population was 70 years (SD 15), while the mean BMI was 29.2 kg/m² (SD 5.8). A total of 16 patients (16.7%) were categorized as immunocompromised. Table I shows the demographic characteristics of the cohort.

Table I.

Demographic characteristics of the included population.

Variable	Data	p-value*
Mean age, yrs (SD)	70.03 (15.07)	0.330
Aged > 80 yrs, n (%)	28 (29)	0.210
Female sex, n (%)	48 (50)	0.588
Right side, n (%)	47 (49)	0.487
Mean BMI, kg/m2 (SD)	29.21 (5.83)	0.200
BMI > 30 kg/m2, n (%)	42 (43.75)	0.441
Condition, n (%)
Diabetes mellitus	12 (12.5)	0.350
Immunosuppression	16 (16.6)	0.062
COPD	7 (7.3)	0.041
Rheumatoid arthritis	6 (6.25)	0.030
Liver cirrhosis	6 (6.25)	0.040
Chronic renal failure	21 (21.87)	0.020
Charlson Comorbidity Index, n (%)
0	7 (7.3)	0.511
1 to 3	46 (47.9)
>3	43 (44.8)
ASA grade, n (%)
I	6 (6.25)	0.372
II	48 (50)
III	40 (41.66)
IV	2 (2.09)
Index diagnosis, n (%)
Primary osteoarthritis	65 (68)	0.490
Femoral neck fracture	31 (32)
Type of fixation, n (%)
Cemented	8 (8.5)	0.371
Hybrid	38 (39.5)
Uncemented	50 (52)
Use of antibiotics in cemented/hybrid prosthesis, n (%)	43 (93.5)	0.362
Type of antibiotics used in cemented/hybrid prosthesis, n (%)
Vancomycin	8 (18.6)	0.583
Gentamicin/tobramycin	31(72)
Vancomycin + gentamicin	4 (9.4)
Bearing surface, n (%)
Ceramic-on-ceramic	11 (11.45)	0.272
Metal-on-polyethylene	61 (63.55)
Ceramic-on-polyethylene	24 (25)
Approach, n (%)
Posterior	96 (100)
Skin closure, n (%)
Staples	58 (60.4)	0.361
Absorbable suture	13 (13.5)
Non-absorbable suture	25 (26.1)
Mean time from index surgery to DAIR, days (SD)	28.45 (16.92)	0.490
Mean time from onset of symptoms to DAIR, days (SD)	8.06 (8.13)	0.003
Time from onset of symptoms to DAIR, days, n (%)
< 7	58 (60.4)	≤0.001
7 to 14	26 (27.1)
> 14	12 (12.5)
Antibiotherapy prior to DAIR, n (%)	47 (49)	0.352
Mean time of antibiotherapy prior to DAIR, (SD)	9.17 (7.2)	0.332
Type of antibiotic previous to DAIR, n (%)
Trimethoprim/sulfamethoxazole	18 (38.3)	0.451
Cephalosporin	8 (17)
Combined antibiotherapy	9 (19.2)
Others (quinolones, amoxicillin, unknown, etc.)	12 (25.5)
Mean leucocytes, mm3 (SD)	9,277.29 (2,961.57)	0.270
Mean ESR, mm/h (SD)	59.05 (29.017)	0.620
Mean serum CRP, mg/l (SD)	89.11 (74.44)	0.010
Mean platelet count, mm3 (SD)	332,154.17 (108,406.05)	0.450
Mean polymorphonuclears, mm3 (SD)	6,528.97 (2648.17)	0.500
Mean lymphocytes, mm3 (SD)	1,433.24 (604.74)	0.383
Mean ratio polymorphonuclears/lymphocytes (SD)	5.57 (3.78)	0.454
Blood culture, n (%)
Positive	8 (8.3)	0.090
Negative	88 (91.7)
Death, n (%)	15 (15.6)	0.541
Primary arthritis, n (%)	4 (26.7)
Femoral neck fracture, n (%)	11 (73.3)

^*Association with DAIR failure.

ASA, American Society of Anesthesiologists; COPD, chronic obstructive pulmonary disease; DAIR, debridement, antibiotics, and implant retention.

DAIR protocol

The approach to treat PJI was multidisciplinary, involving one of four fellowship-trained hip surgeons and an infectious diseases specialist. An extended incision with removal of the previous scar was performed in all cases. After joint dislocation, the stability of the non-modular components was manually assessed. Three to five representative tissue samples were taken for culture and pathology analyses. Irrigation was then performed using at least six litres of fluid using either saline, dilute povidone-iodine, or other solution, depending on the surgeon in charge. Immediately after surgery, broad-spectrum intravenous antibiotherapy was started until identification of the causative organism, from which time oral administration was indicated whenever possible.

KLICC and CRIME-80 scoring data

While the KLICC score was originally created to predict outcomes of DAIR for early acute PJIs,¹⁸ CRIME-80 was described to predict failure in late acute (i.e. haematogenous) rather than in early acute PJIs;^19,20 however, both scores include similar variables. The variables of each score, the KLICC (renal failure, liver failure, index surgery other than THA for primary osteoarthritis, type of cemented fixation, CRP > 115 mg/l) and the CRIME-80 (COPD, CRP > 150 mg/l, index surgery performed for intracapsular hip fracture, male sex, exchange of modular components, age > 80 years), were calculated in isolation to then calculate the final scores for each patient. Renal failure was defined when the patient presented chronically, prior to DAIR with a history of creatininaemia > 1.3 mg/dl and with a creatinine clearance < 72 ml/min/body surface area (BSA) in patients aged 30 to 39 years, < 67 ml/min/BSA in patients aged 40 to 49 years, < 62 ml/min/BSA in patients aged 50 to 59 years, or > 56 ml/min/BSA in patients aged 60 to 72 years.26 Liver failure was considered when the patient presented the following laboratory findings: thrombocytopenia (< 160,000/mm), hypoalbuminaemia, elevated liver transaminases, decreased prothrombin time, and hyperbilirubinaemia.²⁷ The mean KLICC score of the series was 3 (SD 2.29), while the mean CRIME-80 score was 2.48 (SD 2.16).

Outcome measures

The primary outcome measure of interest was to perform external validation of both the KLICC¹⁸ and CRIME-80 scores.^19,20 This was done calculating the calibration and discrimination of both scores. DAIR failure was considered as any new unplanned surgery including a new DAIR, partial or complete implant removal (Girdlestone procedure), revision surgery (either in one or two stages), or any 90-day infection-related death.

A secondary outcome of interest was to modify the best-performing model adjusting it for the most significant demographic variables and to develop a new predictive model adjusted to the characteristics of the included population. In order to certify the validity of this new model, both the calibration and discrimination were evaluated in comparison with the original KLICC/CRIME-80 models.

Statistical analysis

The normality of the distribution of the continuous variables was tested with the Schapiro-Wilk test. If continuous variables had a normal distribution, they were expressed as mean (SD) and compared independently with the independent-samplest-test; whereas variables with non-normal distribution were reported as median (IQR) and compared with the Mann-Whitney U test. Categorical variables were expressed as absolute frequency and percentages, and were compared using the chi-squared test or Fisher’s test when appropriate. The association between demographic variables and DAIR failure was assessed.

A time-to-event analysis was performed describing time to DAIR failure, analyzing the cumulative incidence of failure using death as a competing event, reporting cumulative incidence curves adjusting for competing events (mortality) with the respective 95% CI.²⁸ Time-to-failure of DAIR was necessary to define a cross-sectional timepoint in order to perform subsequent external validation analyses of both predictive models, although most external validations had been performed at two-year follow-up.¹¹

External validation was performed calculating the calibration and discrimination. We used the Cox approach to assess calibration, calculated as the logistic regression model intercept. With this approach, the perfect prediction is described with an α intercept (‘calibration in the large’ (CITL)) of 0 and a β slope of 1.¹⁰ The discrimination was quantified with a Harrel concordance statistic test (‘c’) that is identical to the area under the receiver operating characteristic (AUROC) curve for a binary outcome.²⁹ The ‘c’ statistic with a 95% CI was used as a measure of discrimination of each model and the Hanley-McNeil test was used to compare both models.³⁰

As per the secondary outcome of interest, in order to certify the validity of a new model, both the calibration and discrimination of the new model were compared with the original KLICC/CRIME-80 models. In case the best-performing model was the KLICC score, since it is a predictive algorithm where each variable has a score ranging from 0.5 to 2.5 points based on the regression coefficients of the model,¹⁸ we performed a transformation of the categories of the included variable(s) in the new model using the same transformation proposed in the original KLICC (0.5 score points for each point of β coefficient).

The inclusion of any quantitative variable into the original KLICC or CRIME-80 scores required transformation of the variable into a categorical (i.e. qualitative) one, for purposes of the scoring system. To do so, the likelihood-ratio (Wilks) test was used to compare two different maximum likelihood estimates of a parameter in order to decide whether to reject or not a restriction on the parameter.³¹ Statistical significance was set at p < 0.05.

Power analysis

For purposes of the power analysis, the KLICC or CRIME-80 scores were considered as a single variable and one extra parameter was planned to be added to either one of the models in order to improve their calibration. The selection of this additional variable was done after the univariate analysis, looking for the variable that, together with the CRIME-80 or KLICC scores, best explained and improved the model. To build a predictive model with approximately two estimated parameters (either the CRIME-80 or KLICC scores plus one additional parameter/variable), ten to 20 events per estimated parameter would be needed, i.e. between 20 and 40 DAIR failure events (‘rule of thumb’).³² Considering that DAIR failure was expected to be 25% to 35%, it was estimated that a sample size of between 100 and 140 patients was needed.

Results

In total, 12 of the 96 patients (12.5%) had diabetes, 44 patients (45%) had a CCI score > 3, and 42 patients (43.75%) had an ASA score > 2 (Table I). Overall, 65 patients (68%) had an initial diagnosis of primary osteoarthritis of the hip, while 31 patients (32%) had an intracapsular hip fracture. All patients were classified Tsukayama type II; i.e. as early acute postoperative PJI. A total of 47 patients (48.9%) received oral antibiotic therapy prior to DAIR. The mean preoperative antibiotic therapy time was 9.17 (SD 7.2) days. Modular component exchange was done in only one case, since this is not a regular institutional practice. Use of infra- and/or supra-fascial vancomycin powder was also done at surgeon’s discretion, with 14 cases (14.6%) receiving intraoperative powdered antibiotics before closure. Mean postoperative antibiotic treatment was 9.8 (SD 7) weeks. Six patients received suppressive (i.e. > six months) antibiotic therapy. Table II shows the characteristics of the DAIR protocols included in this series.

Table II.

Characteristics of the debridement, antibiotics, and implant retention protocols included in the series.

Variable	Data	p-value
Surgeon’s experience, n (%)
Consultant	53 (55.2)	0.622
Fellow	43 (44.8)
Exchange of modular components	1 (1.0)	N/A
Irrigating solution, n (%)
Saline	24 (25)	0.462
Betadine	56 (58.4)
Chlorhexidine	8 (8.3)
Others (rifocin)	8 (8.3)
Use of local antibiotic power, n (%)	14 (14.6)	0.321
Causative organism, n (%)
MSSA	20 (20.8)	0.061
MRSA	9 (9.4)
S. epidermidis	12 (12.5)
Escherichia coli	6 (6.3)
Pseudomonas aeruginosa	7 (7.3)
Others (Streptococcus agalactiae, Streptococcus viridans, Cutibacterium acnes, Proteus mirabilis, Corynebacterium spp, Enterobacter, etc)	13 (13.5)
Polymicrobial	10 (10.4)
Negative culture	19 (19.8)
Mean duration of postoperative antibiotherapy, mnths (SD)	39.33 (28.4)	0.581
Chronic suppressive antibiotherapy, n (%)	6 (6.25)	0.342

^*Association with DAIR failure.

DAIR, debridement, antibiotics, and implant retention; MRSA, methicillin-resistant Staphylococcus aureus; MSSA, methicillin-sensitive Staphylococcus aureus; N/A, not applicable; spp, species.

Time to DAIR failure

There were a total of 24 (25%) DAIR failures. Of these, 13 (13%) were new unplanned DAIRs and 11 (11%) revision surgeries. There were no infection-related deaths. Most failures occurred within the first two years. The cumulative incidence of DAIR failure was 23.96% (95% CI 15.9% to 32.8%) at 24 months and 25.09% (95% CI 16.9% to 35%) at 35 months (Figure 2). The incidence density of DAIR failure was 0.00761 events per person/month of follow-up; that is, 0.91 reoperations per ten person-years of follow-up. Considering these outcomes, two-year follow-up was used as an appropriate timepoint to perform subsequent external validation analyses.

Fig. 2.

Graph showing cumulative incidence function of debridement, antibiotics, and implant retention failure using death as a competing event. CIF, cumulative incidence function.

External validation of the KLICC score

Using the KLICC predictive model and the 24-month time interval to perform validation, the calculation of the calibration through a Cox approach obtained a CITL of -0.57 (95% CI -1.16 to 0.01) and a slope of 0.68 (95% CI 0.35 to 1.02), while the AUROC for discrimination was 0.79 (95% CI 0.67 to 0.90) (Figure 3 and Figure 4).

Fig. 3.

Image showing the calibration of the KLICC (left) and CRIME-80 (right) scores, with the corresponding calibration in the large (CITL) and slope values. DAIR, debridement, antibiotics, and implant retention.

Fig. 4.

Image showing the discrimination results of both the KLICC (blue line) and CRIME-80 (red line), with the corresponding receiver operating characteristic (ROC) curve.

External validation of the CRIME-80 score

As per the CRIME-80 score, the calibration obtained a slope of 0.35 (95% CI -0.14 to 0.85) and a CITL of -1.66 (95% CI -2.13 to -1.19), whereas the discrimination had an AUROC of 0.63 (95% CI 0.51 to 0.76) (Figure 3 and Figure 4).

Comparison between the KLICC and CRIME-80 predictive models

The difference in the AUROC curve between the KLICC (ROC = 0.78) and CRIME-80 (ROC = 0.63) scores was statistically significant (p = 0.014), with the KLICC score being superior (Figure 3 and Figure 4).

Creation of a new predictive model: the modified-KLICC score

From the univariate analysis, the variable ‘time from onset of symptoms to DAIR’ was found to have a significant association with subsequent DAIR failure (p = 0.003; Table I and Table II). Therefore, considering that the KLICC score was superior to the CRIME-80 in both discrimination and calibration, the variable ‘time from onset of symptoms to DAIR’ was merged into the original KLICC model creating a new predictive model termed modified-KLICC score. Afterwards, both scores were compared with each other.

The quantitative variable ‘time from symptom onset to DAIR’ was categorized dichotomously into < seven days from onset of symptom to DAIR, ≥ seven days and < 14 days from symptoms to DAIR, and ≥ 14 days. The choice of the number of days was supported by the likelihood-ratio test which showed that the binary split of seven and 14 days significantly reduced the residual variability of the model compared with other cut-off points, after performing an analysis of the local polynomial regression through locally weighted scatterplot smoothing (LOWESS). This was done by giving 0 points to a patient who had less than seven days of infection symptomatology, 1 point when the patient had between seven and 14 days from the onset of symptoms until the DAIR, and 3 points when the DAIR was performed at, or after, day 14 of symptomatology. The choice of these scoring points was based on the same scoring transformation proposed by the authors of the KLICC score.

The new model with the variable ‘time from onset of symptoms to DAIR’ presented a CITL of -0.07 (95% CI -0.68 to 0.53) with a slope of 1.24 (95% CI 0.65 to 1.82), and an AUROC of 0.85 (95% CI 0.74 to 0.95), as shown in Figure 5.

Fig. 5.

Image showing the calibration curve (left) of the modified-KLICC score, with the corresponding calibration in the large (CITL) and slope values; as well as the discrimination area under the receiver operator curve (AUROC) (right) of both the original KLICC (blue line) and modified-KLICC (red line) scores.

Comparison between the KLICC and modified-KLICC predictive models

The difference in the AUROC curve between the KLICC and modified-KLICC scores was statistically significant (0.79, 95% CI 0.67 to 0.90 vs 0.85, 95% CI 0.74 to 0.95), p = 0.032), with the latter model performing better.

Discussion

Nowadays, the treatment of choice for acute PJI of the hip is DAIR, considering that during the acute period the causative microorganism has not yet formed a mature biofilm.^9,33 However, the success of this procedure is variable, ranging from 10% to 95%,⁸ depending on numerous factors, many of which are interrelated. The KLICC and CRIME-80 scores, designed for this matter, stratify multiple risk factors which ultimately define the clinical status of the host and, together with the type of causative organism, and the way the DAIR protocol was applied, help predict outcomes and patient selection.³⁴ Therefore, external validation of such predictive models is required for proper patient selection worldwide, since the last 2018 International Consensus Meeting, although supporting the indication of DAIR in acute PJI,⁹ suggested that there was moderate evidence on the accuracy and clinical applicability of prognostic scoring systems for prediction of DAIR failure.¹²

One of the main findings of our study was that the KLICC score was a significantly better predictor of failure than the CRIME-80. This can be due to the fact that the CRIME-80 score was originally described to predict failure in late acute (i.e. haematogenous) rather than in early acute PJIs,^19,20 despite including several risk factors that could also explain suboptimal outcomes in early acute postoperative PJIs (e.g. male sex, older age, non-exchange of modular components, and rheumatoid arthritis).^35,36 However, both the KLICC and CRIME-80 models performed inferiorly in our cohort than in their original reports. Our cohort did not include cases of DAIR in revision surgeries and focused exclusively on primary cases, and this could alter the predictive value of both models given the higher failure rate of DAIR in such cases when compared with primary arthroplasties.³⁷ The KLICC and CRIME-80 predictive models function as a scoring algorithm stratifying ordinal groups that, in increasing order, increase the risk of DAIR failure. Both scores share some variables, including the type of index surgery and the baseline serum CRP at the time of diagnosis, although the cut-off points for DAIR failure differs subtly between the two models. The original study by Tornero et al¹⁸ reported a DAIR failure rate of 23.4% at a 24-month follow-up; and the KLICC score obtained an AUROC curve of 0.839 to discriminate this outcome, while in our cohort the AUROC was 0.79. The original KLICC includes, among its variables, cemented fixation as a risk factor; however, in our population, most intracapsular fractures in elderly patients were treated with this type of fixation, being thus a confounding factor. In the setting of late acute PJI, the authors of the CRIME-80 model reported that a score ≥ 3 had a probability of DAIR failure of 65%, recommending revision surgery in such cases.^19,20 However, they validated the score without defining an AUROC, thus our CRIME-80 AUROC of 0.63 cannot be compared with the original.

Both scoring systems have had very few external validations reported in the literature. Unlike our study, none of them were done calculating both the calibration and discrimination. Chalmers et al performed an external validation analysis of both scores in two separate populations, an acute postoperative cohort (n = 122) to validate the KLICC score, and an acute haematogenous cohort (n = 134) to validate the CRIME-80 score.²¹ In patients with acute postoperative PJI, the KLICC had an AUROC curve of 0.64 and 0.63 at 90 days and two years, respectively, proving to be of limited value. On the other hand, the CRIME-80 had an AUROC curve of 0.77 and 0.65 at 90 days and two years, respectively, to predict DAIR failure in the acute haematogenous PJI cohort,²¹ showing a moderate performance. In comparison, the KLICC performed better in our cohort, whereas the CRIME-80 obtained a similar (poor) validation. Similarly, Lowik et al¹¹ analyzed the KLICC score in a cohort of 386 acute PJI cases and obtained an AUROC curve lower (0.64) than both our’s and the original’s; however, the authors highlighted the usefulness of the model when stratifying the total score into low-risk (< 3.5 points) and high-risk (> 6 points) patients. Similar poor validation results of the KLICC score had been reported by Duffy et al³⁸ in a retrospective cohort of 59 knee PJIs and by Bernaus et al³⁹ in a retrospective series of 455 acute PJIs however, none of these reports calculated neither the calibration nor AUROC value. There is no doubt that an ideal prognostic system for DAIR surgery should have similar validation results worldwide; nonetheless, local epidemiology does differ between nations and regions, including dissimilar international organism profile of PJIs.⁴⁰ Therefore, equivalent outcomes after surgery may not be expected.

In this scenario, and taking into account the limited clinical applicability of these scores, including a risk factor that remains unchanged among the different populations may be of value. Considering that ‘time from onset of symptoms to DAIR’ was the variable with the highest statistical association to DAIR failure in the univariate analysis, we created a new predictive model for DAIR failure called modified-KLICC score, which combines the aforementioned variable with the original KLICC model. A recent systematic review of cohort studies showed that in studies where the median time from the onset of symptoms to debridement was > seven days, the pooled proportion of success was 51.8% (170/329 patients; 95% CI 46.1% to 57.2%), whereas in those where the median time from the onset of symptoms to debridement was < seven days, the pooled proportion of success was 72.0% (198/275 patients; 95% CI 66.3% to 77.2%).⁷ Although the time window for performing DAIR with the greatest likelihood of success is not precisely established, most authors propose a seven-day period from infection diagnosis as the ideal for performing DAIR.^1,34,41-43 The underlying pathophysiological reason for this is the time of mature biofilm formation, which ranges from four to seven days from contact with the implant surface,^44-46 at which point mechanical debridement and irrigation with different types of solutions would not be therapeutically effective, while other alternatives such as one- or two-stage revision could improve the success rate.⁴⁷ In the original KLICC score, all cases had a duration of symptoms shorter than 21 days, and the variable ‘time between PJI diagnosis and surgical debridement’ was categorized into ≤ four days and > four days, without being statistically significant between the remission (n = 170) and failure cases (n = 52).¹⁸ Likewise, the authors of the CRIME-80 score used a ‘duration of symptoms > ten days’ as a cut-off value to analyze outcomes, despite having a statistically significant association with DAIR failure (adjusted OR for failure 1.21 (95% CI 0.54 to 2.74); p = 0.64).²⁰ By adding the variable ‘time from onset of symptoms to DAIR’ to the original KLICC score, a marked improvement in the performance of the KLICC model was noted, increasing the AUROC curve from 0.79 to 0.85 (p = 0.0322), and the CITL from -0.57 with a slope of 0.68 CITL to -0.07 with a slope of 1.24. In this sense, like Tsang et al,⁷ we believe that timing of debridement after the onset of symptoms of infection is a determinant of outcome that should be included in any prognostic model analyzing DAIR failure.

Other novel prognostic scores had been described to estimate failure following DAIR. Recently, Shohat et al,³ using machine learning, developed a predictive algorithm to improve decision-making in cases with acute PJI, with good external validation results.¹⁰ The authors reported, by order of importance, several non-modifiable demographic variables that may preclude failure: 1) higher serum CRP levels; 2) positive blood cultures; 3) indication for index arthroplasty other than primary osteoarthritis; and 4) no exchange of modular components. The authors acknowledged, however, that hip compared with knee locations as well as early-acute compared with late-haematogenous PJIs exhibited dissimilarities as per the ten most important variables detected by the random forest on each location and type of infection separately.³ Although not changing the modular components was overall the fourth most important factor for septic failure, this factor did not have any relative importance in THA re-infections. While days of symptoms prior to DAIR was the most relevant risk for failure in late haematogenous PJIs, it had no relative importance in the outcomes after early acute PJIs. Also, while time from index surgery to DAIR was the fourth most important factor precluding DAIR failure in infected knee arthroplasties, it had no relevance in the outcomes of infected THAs. Therefore, we feel there is no uniformity in the clinical application of such artificial intelligence-based models, nor have they been validated in the international consensus meetings addressing these infections.

Our study was not without limitations. First, we included a small number of acute primary THAs (n = 96), while the power analysis estimated a sample size of between 100 and 140 patients, so we believe that the study might be underpowered. However, we have excluded acute PJIs after revision surgeries in order to strengthen the predictability of the risk factors precluding failure since we believe they could be considered as a confounding factor.⁴⁸ Second, we acknowledge that the practice of DAIR was performed in a non-standardized fashion, with diverse surgical techniques and irrigation solutions, and with some surgeons even indicating oral antibiotics previous to the irrigation and debridement procedure. Additionally, exchange of modular components was not performed in the majority of cases; therefore, the survival analysis can be considered as a best-case estimate. It has been shown that changing the modular components may increase the success rate following DAIR.¹ However, even without this practice, the survival rate following DAIR was acceptable (76.04% at two years), being similar or even higher than other cohorts in which modular components had been exchanged.^7,8 Also, diverse causative organisms were identified and although they showed borderline statistical significance (p = 0.06), there are several bacteria with potential of producing more resistant biofilm, affecting DAIR outcomes.⁴⁹ Finally, the criteria used to define DAIR failure were not completely the same as the original cohorts, as we did not consider cases receiving suppressive antibiotic treatment as failures, which might make the survival analysis a best-case estimate by increasing the infection-free survivorship.⁵⁰ It is not completely clear whether patients with suppressive therapy should be considered as failures or part of the treatment strategy which is undertaken in order to prevent further surgery.⁵¹ However, we acknowledge certain strengths in that this is the first external validation study calculating both calibration and discrimination, and a thorough analysis of risk factors was undertaken to perform the optimization of the KLICC score.

In conclusion, there is no universal prognostic score with perfect results applicable to all populations. In this population, the KLICC score was superior to the CRIME-80 one in predicting DAIR failure, although both validations performed poorer than the original ones. The modified-KLICC score, created in the present study including the variable ‘days of symptoms to DAIR’, improved the prediction of DAIR failure. This score may help select more accurately patients who are likely to benefit from DAIR, and to indicate alternatives to DAIR in those with a high probability of failure. Large cohort studies are yet necessary to validate the prognostic algorithms precluding failure in acute PJI cases.

Take home message

- Proper external validation of predictive models precluding debridement, antibiotics, and implant retention (DAIR) failure is required for accurate patient selection.

- The KLICC score was a significantly better predictor of DAIR failure than the CRIME-80 score.

- As compared with the original KLICC score, the newly introduced modified KLICC improved the discrimination and calibration of the score.

Data Availability

The datasets generated and analyzed in the current study are not publicly available due to data protection regulations. Access to data is limited to the researchers who have obtained permission for data processing. Further inquiries can be made to the corresponding author.