This was a single-center retrospective cohort study conducted at the First Affiliated Hospital of Fujian Medical University. This study was approved by the Institutional Review Board (approval no. MRCTA, ECFAH of FMU [2025] 007), and informed consent was obtained from all participants. The reporting of this study follows STROBE guidelines. Patients who underwent DAIR for acute PJI between 2008 and 2024 were retrospectively reviewed.

Inclusion criteria: (i) Hip or knee arthroplasty patients diagnosed with acute PJI based on the 2011 Musculoskeletal Infection Society (MSIS) criteria (Parvizi et al., 2011), including early postoperative (Tsukayama type 2) and acute hematogenous (Tsukayama type 3) infections (Tsukayama et al., 1996). (ii) Patients treated with DAIR.

Exclusion criteria: (i) Infections occurring >3 months postoperatively or with symptoms lasting >3 months. (ii) Patients treated with one- or two-stage revision. (iii) Incomplete follow-up or lost-to-follow-up cases.

Synovial fluid or pus was cultured in aerobic and anaerobic bottles, and tissue samples were processed for extended cultures (up to 14 days if necessary). Pathogen identification and resistance profiling were performed. Metagenomic next-generation sequencing (mNGS) was also conducted for culture-negative cases (BGI Group, Shenzhen, China). The results were normally available in 48 hours. In this study, resistant pathogens were specifically defined as microorganisms that demonstrated resistance to antibiotics in antimicrobial susceptibility testing.

Our institution functions as a tertiary referral center for bone and joint infections in Fujian Province in the southeast of China. The majority of patients referred to our center had initially undergone arthroplasty procedures at other hospitals before being transferred for management of acute PJI. Upon confirming an acute PJI diagnosis, empirical antibiotic therapy was initiated immediately without awaiting microbiological culture results, followed by prompt DAIR procedures performed exclusively by a senior orthopedic surgeon (W. Zhang) using a standardized protocol.

The surgical procedure commenced with spinal anesthesia and thorough disinfection using iodine tincture followed by ethanol deiodination. Through the original incision, superficial debridement was performed before arthrotomy, during which synovial fluid was aspirated for both conventional cultures and mNGS. Following capsulotomy, a minimum of five periprosthetic tissue specimens were collected from distinct sites for microbiological analysis. The modular components were then exchanged, and all infected/necrotic tissues were systematically debrided. The surgical field underwent povidone-iodine immersion for 30 minutes followed by pulsed lavage with 3 liters of normal saline. After temporary wound closure and complete instrument exchange, the new components were implanted and the wound was closed in layers with drain placement for 48 hours, utilizing tourniquet control when necessary during knee procedures.

Patients were initially treated with an empirical antibiotic regimen consisting of vancomycin (1.0 g intravenous infusion every 12 hours) combined with ceftazidime (2.0 g intravenous infusion every 8 hours) until culture or mNGS results became available. The antibiotic therapy was subsequently adjusted to target the specific pathogens based on antimicrobial susceptibility testing. In selected cases, infectious disease specialists were consulted for multidisciplinary evaluation, with routine monitoring of hepatic and renal function throughout the treatment course. The standard protocol involved 2 weeks of intravenous antibiotic administration followed by 4 weeks of oral antibiotic therapy.

Comprehensive patient data were collected, including baseline characteristics and perioperative parameters (demographics, infection profiles, laboratory markers, microbiological findings, and surgical details). Treatment success was defined as: (1) resolution of clinical symptoms, (2) normalization of inflammatory markers (CRP, ESR, WBC), (3) absence of radiographic loosening, and (4) no requirement for lifelong suppressive antibiotics (minimum 1-year follow-up). Failure was defined as: (1) need for additional surgical interventions (repeat DAIR, debridement, one- or two-stage revision) or (2) indefinite antibiotic suppression. (Zhang et al., 2020) Univariate and multivariate logistic regression analyses were performed to compare variables between groups, followed by the construction of a nomogram prediction model.

The construction of the predictive model followed a rigorous multi-step analytical process. Initial screening of potential predictors was performed through univariate logistic regression analysis, retaining variables with P-values ≤0.2 for further evaluation. These candidate variables were subsequently analyzed using least absolute shrinkage and selection operator (LASSO) logistic regression with 10-fold cross-validation to identify features with non-zero coefficients. The optimal lambda parameter was determined based on the one-standard-error rule (1se) from the minimum deviance. Variables selected through LASSO regression were then entered into a bidirectional stepwise multivariate logistic regression analysis, with statistical significance set at P<0.05 for both entry and retention in the final model. The resulting statistically significant predictors were incorporated into the nomogram construction. The developed nomogram underwent extensive validation procedures, including internal validation with 1000 bootstrap resamples to assess model stability. Calibration curves were generated to evaluate the agreement between predicted and observed outcomes. The model’s discriminative ability was examined through receiver operating characteristic (ROC) curve analysis, with calculation of the area under the curve (AUC), sensitivity, and specificity. Clinical utility was further assessed using decision curve analysis (DCA) to quantify net benefit across various probability thresholds. Additional validation was performed through 10-fold cross-validation.

All statistical analyses were conducted using specialized R packages: the “glmnet” package for LASSO regression, “rms” for nomogram construction and calibration plotting, “pROC” and “ggplot2” for ROC curve generation, and “ggDCA” for decision curve analysis. A two-tailed P-value <0.05 was considered statistically significant for all analyses.

All data processing and statistical analyses were performed using R software (version 4.4.1). Continuous variables were presented as mean ± standard deviation when normally distributed, with between-group comparisons conducted using independent samples t-tests. Non-normally distributed continuous variables were expressed as median [P25, P75] and analyzed using the Wilcoxon rank-sum test. Categorical variables were summarized as counts (percentages), with between-group comparisons performed using chi-square tests or Fisher’s exact test when appropriate. Missing data were handled through multiple imputation using the “missRanger” package in R.

The study cohort comprised 90 patients who met the inclusion criteria for the final analysis. The population consisted of 45 male and 45 female participants. The mean patient age was 67.1 ± 11.5 years, with an average body mass index (BMI) of 25.5 ± 3.2 kg/m². Regarding infection sites, 40 cases (44.4%) developed following hip arthroplasty procedures, while 50 cases (55.6%) occurred after knee arthroplasty. Among the hip procedures, total hip arthroplasty (THA) accounted for 31 cases (77.5% of hip infections) and hemiarthroplasty for 9 cases (22.5%). For knee procedures, total knee arthroplasty (TKA) represented 41 cases (82.0% of knee infections) and unicompartmental knee arthroplasty (UKA) comprised 9 cases (18.0%). Based on the Tsukayama classification system, the cohort included 66 cases (73.3%) of early postoperative infection and 24 cases (26.7%) of acute hematogenous infection.

The study demonstrated an overall treatment success rate of 85.6% (77/90), with 77 successful cases and 13 failures. Baseline characteristics between groups showed no statistically significant differences in age, gender, BMI, Charlson comorbidity index (CCI), or presence of oozing (all p>0.05). (Table 1) Significant between-group differences were observed in both infection location and type (p=0.023 and p=0.002, respectively). The success group comprised 38 hip cases (95% success rate among total hip PJIs) and 39 knee cases (78% success rate among total knee PJIs), while the failure group included only 2 hip cases but 11 knee cases. This disparity suggests greater DAIR failure susceptibility in knee procedures. Among knee infections, success rates were comparable between TKA (78.0%; 32/41) and UKA (77.8%; 7/9). Regarding infection classification, the success group consisted of 61 early postoperative infections and 16 acute hematogenous infections, whereas the failure group contained 5 early infections versus 8 hematogenous cases. This distribution indicates substantially higher failure rates for acute hematogenous infections compared to early postoperative cases (33.3% vs 7.6% failure rate).

Table 2 presents the comparison of perioperative parameters between the two patient groups.

The analysis of perioperative indicators showed no statistically significant differences in ASA classification between the success and failure groups. However, the failure group demonstrated significantly elevated levels across multiple inflammatory markers compared to the success group. Preoperative peripheral blood tests revealed higher white blood cell counts (9.5×10^9/L vs 8.2×10^9/L, p=0.043), CRP levels (79.6 mg/L vs 42.4 mg/L, p<0.001), and ESR values (80.6 mm/h vs 60.5 mm/h, p=0.002) in the failure group. Synovial fluid analysis similarly showed more pronounced inflammation in cases of DAIR failure, with significantly higher leukocyte counts (35,300×10^6/L vs 21,843×10^6/L, p=0.043) and polymorphonuclear neutrophil percentages (PMN) (91.7% vs 83.8%, p<0.001). These findings collectively indicate that both systemic and local inflammatory responses were more severe in patients who ultimately failed DAIR treatment, with CRP and PMN% showing particularly strong discriminatory value for predicting treatment outcomes.

Microbial culture results were positive in 74 cases, yielding a culture positivity rate of 82.2% (74/90). Among the 16 culture-negative cases, 4 were subsequently identified as positive through mNGS testing. The pathogens identified in these four mNGS-positive cases were burkholderia, Microcystis aeruginosa, staphylococcus and Staphylococcus haemolyticus. Figure 1 presents the comparative microbiological profiles between the two outcome groups. In the success group, microbial identification was achieved in 65 cases, while all patients in the failure group had confirmed microbiological results. (Figure 1)

The extended microbiological analysis revealed that several factors did not significantly influence surgical outcomes, including pathogen-negative status, presence of polymicrobial infection, and Gram-positive staining characteristics. However, a notable trend was observed regarding antimicrobial resistance patterns, with the failure group demonstrating a higher incidence of resistant pathogen infections compared to the success group, though this difference did not reach conventional statistical significance (p=0.106). (Table 3)

The univariate regression analysis (Table 4) identified several potential risk factors for DAIR failure: infection site (hip/knee, p=0.036), infection type (early postoperative/hematogenous, p=0.004), preoperative peripheral WBC count (p=0.011), CRP level (p=0.003), ESR value (p=0.041), synovial fluid PMN% (p=0.006), duration of infection symptoms (p=0.062), and presence of resistant pathogens (p=0.114). To address potential multicollinearity among these variables, all factors with P ≤ 0.1 in the univariate analysis were included in the LASSO regression analysis. The LASSO regression with 10-fold cross-validation identified the optimal lambda value (logλ≈-3.5, Supplementary Figure 1), selecting four significant variables for the final multivariate model: infection type, preoperative CRP, synovial fluid PMN%, and microbial resistance. The multivariate logistic regression analysis (Table 5) confirmed these as independent predictive factors, with hematogenous infection (OR = 11.704, 95% CI: 1.957-119.357) and preoperative CRP (OR = 1.022, 95% CI: 1.009-1.040) demonstrating the strongest predictive value for DAIR failure. (Table 5)

The predictive performance of inflammatory markers was further evaluated through ROC curve analysis (Figure 2). CRP demonstrated significant discriminative ability for DAIR treatment failure with an AUC of 0.797 (95% CI: 0.692-0.902). Synovial fluid PMN% showed superior predictive accuracy compared to CRP, achieving an AUC of 0.825 (95% CI: 0.726-0.924). Using the Youden index, optimal cutoff values were determined to maximize predictive performance. For CRP, the threshold of 63.3 mg/L provided perfect sensitivity (100%) while maintaining 61% specificity. Similarly, a PMN% cutoff of 87.4% achieved 100% sensitivity with 64.9% specificity. These results suggest that while both markers show excellent sensitivity for identifying potential treatment failures, PMN% may offer marginally better overall performance in clinical decision-making.

The predictive nomogram (Figure 3) was developed incorporating the four independent risk factors identified in the multivariate analysis, with a total possible score ranging from 0 to 180 points. The nomogram displays individual variable scores on the top scale, where clinical values can be converted to points by drawing vertical lines. The sum of these points corresponds to a total score on the bottom scale, which can then be projected downward to estimate the probability of treatment failure. The model’s predictive performance was rigorously evaluated through multiple validation methods. ROC curve analysis demonstrated strong discriminative ability, while calibration plots showed excellent agreement between predicted and observed outcomes. Decision curve analysis further confirmed the model’s clinical utility across various probability thresholds (Supplementary Figures 2–4).