Development and temporal validation of a clinical prediction model to identify bacterial infection in hospitalised patients with systemic lupus erythematosus Authors: Rong-qia Wu, Ze-xin Zheng, Qiong-bing Zheng, Hong-jin Liang, Yao Gong, Dan-min Wang, Zhi-duo Hou, Ling Lin Published: 2026 ABSTRACT AbstractObjectiveTo develop and temporally validate a clinical prediction model for identifying bacterial infection (BI) among hospitalised patients with SLE.MethodsWe retrospectively included hospitalised patients with SLE admitted for BI or active disease between January 2023 and January 2026. Admissions from January 2023 to January 2025 formed the training cohort, and admissions from February 2025 to January 2026 formed the temporal validation cohort. BI included infection with or without concomitant active SLE; non-infected active (NA) SLE was the comparator. Candidate predictors were prespecified according to clinical relevance, prior literature, routine availability and missingness. LASSO (Least Absolute Shrinkage and Selection Operator) regression was applied across imputed datasets, and final coefficients were estimated using multivariable logistic regression and pooled using Rubin’s rules. Performance was assessed by discrimination, calibration, full-workflow bootstrap internal validation, same-centre temporal validation and decision curve analysis.ResultsThe training cohort included 395 admissions (170 BI, 225 NA), and the validation cohort included 200 admissions (75 BI, 125 NA). The final model included age, disease duration, previous hospitalisation for BI, fever, white cell count, C reactive protein (CRP), procalcitonin, C4 hypocomplementemia and ordinal SLE-DAS (Systemic Lupus Erythematosus Disease Activity Score) category. Training area under the curve (AUC) was 0.869 (95% CI 0.833 to 0.905), exceeding CRP (0.771), procalcitonin (0.682) and the erythrocyte sedimentation rate/C reactive protein (ESR/CRP) ratio (0.736). Full-workflow bootstrap validation yielded an optimism-corrected AUC of 0.837 and corrected Brier score of 0.162. Same-centre temporal validation showed stable discrimination (AUC 0.852, 95% CI 0.794 to 0.910), acceptable calibration and preserved net benefit.ConclusionsA routinely available, interpretable model identified BI among hospitalised patients with SLE using NA SLE as the comparator. C4 hypocomplementemia and higher SLE-DAS categories should be interpreted as markers of disease activity predominance rather than protective factors against infection. Multicentre external validation and clinical impact studies are required. WHAT IS ALREADY KNOWN ON THIS TOPIC Bacterial infection (BI) and active SLE frequently overlap clinically, and single biomarkers provide incomplete discrimination. Bacterial infection (BI) and active SLE frequently overlap clinically, and single biomarkers provide incomplete discrimination. WHAT THIS STUDY ADDS Using non-infected active SLE as the comparator while retaining BI with concomitant active SLE in the BI group, this study developed and temporally validated an interpretable model based on routinely available clinical and laboratory information. Using non-infected active SLE as the comparator while retaining BI with concomitant active SLE in the BI group, this study developed and temporally validated an interpretable model based on routinely available clinical and laboratory information. The model incorporated ordinal SLE-DAS (Systemic Lupus Erythematosus Disease Activity Score) and achieved area under the curves of 0.869 in the training cohort and 0.852 in same-centre temporal validation. The model incorporated ordinal SLE-DAS (Systemic Lupus Erythematosus Disease Activity Score) and achieved area under the curves of 0.869 in the training cohort and 0.852 in same-centre temporal validation. The negative associations of C4 hypocomplementemia and higher SLE-DAS categories reflected the active-SLE comparator context and should not be interpreted as protective effects against BI. The negative associations of C4 hypocomplementemia and higher SLE-DAS categories reflected the active-SLE comparator context and should not be interpreted as protective effects against BI. HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY The model may support structured bedside assessment and provides a basis for multicentre external validation and prospective impact studies. The model may support structured bedside assessment and provides a basis for multicentre external validation and prospective impact studies. Introduction SLE is a multisystem autoimmune disease marked by immune dysregulation, complement activation, autoantibody production and fluctuating disease activity. Infection is a major complication of SLE and contributes substantially to hospitalisation and mortality. Susceptibility to infection reflects overlapping disease-related and treatment-related factors, including lymphopenia, complement consumption, organ involvement, glucocorticoid exposure and immunosuppressive therapy.1–4 Differentiating bacterial infection (BI) from active SLE is a persistent challenge in hospitalised patients. Fever, fatigue, abnormal white blood cell count, elevated inflammatory markers, serositis and pulmonary manifestations may occur in either condition. Diagnostic misclassification can have immediate therapeutic consequences. Escalating immunosuppression in an infected patient may worsen infection, whereas treating active SLE as infection may delay anti-inflammatory or immunomodulatory treatment.1 5–7 Several biomarkers, including C reactive protein (CRP), procalcitonin (PCT), white blood cell (WBC) count, neutrophil-to-lymphocyte ratio, erythrocyte sedimentation rate (ESR)/CRP ratio, neutrophil CD64 and presepsin, have been studied for distinguishing infection from flare in SLE, and recent studies continue to support the diagnostic relevance of PCT, ESR/CRP ratio and neutrophil-to-lymphocyte ratio in febrile patients with SLE.8–16 No single marker, however, captures the clinical, inflammatory and immunological context of hospitalised SLE. A multivariable clinical prediction model may better integrate these complementary signals, and recent work has further explored machine learning approaches for infection-related prediction in lupus nephritis using immune and laboratory variables.6 7 13 17–19 A recent multicentre study in febrile patients with SLE further highlighted the ongoing difficulty of discriminating flare from infection and the need for clinically interpretable tools based on routinely available clinical and laboratory information.13 Existing calculators and biomarker studies have advanced the distinction between flare and infection in febrile patients with SLE, but several clinically relevant gaps remain. First, infection-risk prediction and concurrent differential diagnosis represent distinct clinical questions: factors associated with future susceptibility to infection do not necessarily distinguish BI from active SLE at the time of hospitalisation. Second, many previous differential-diagnosis studies have treated infection and disease flare as competing states, although BI and active SLE may coexist. Third, disease activity has predominantly been represented by Systemic Lupus Erythematosus Disease Activity Score (SLE-DAI)-based measures or conventional activity markers in previous differential-diagnosis studies,6 7 13 17 18 whereas SLE-DAS20 21 has been less frequently evaluated in this setting. We therefore developed an interpretable model to estimate the probability of concurrent BI among hospitalised patients with SLE, using non-infected active SLE as the clinically relevant comparator while retaining patients with concomitant active SLE in the BI group. We further incorporated ordinal SLE-DAS into model development and examined continuous SLE-DAS to clarify its contribution to infection-versus-disease-activity discrimination. Model robustness was evaluated using full-workflow bootstrap internal validation and same-centre temporal validation. Methods Study design and participants We retrospectively reviewed patients with SLE admitted to the Department of Rheumatology and Immunology, First Affiliated Hospital of Shantou University Medical College, for BI or active disease between January 2023 and January 2026. Admissions from January 2023 to January 2025 formed the training cohort for model development and bootstrap internal validation. Admissions from February 2025 to January 2026 formed the temporal validation cohort. The same eligibility criteria, predictor definitions and outcome definitions were applied to both cohorts. The unit of analysis was hospitalisation because the intended prediction task concerned the presence of BI during each admission, and infection status, disease activity and other admission-level clinical features could differ between hospitalisations. Therefore, repeated eligible admissions from the same patient during the study period could be included as separate observations. Eligible patients fulfilled the 2019 European League Against Rheumatism (EULAR)/American College of Rheumatology (ACR) classification criteria for SLE.22 Patients were excluded if they were 14 years or younger, had another systemic autoimmune disease or had neither BI nor evidence of active SLE. Outcome definition Patients were classified into a bacterial infection (BI) group or a non-infected active (NA) SLE group. The BI group included patients with BI, including those with concomitant active SLE. The NA group included non-infected hospitalised patients with active SLE. Accordingly, the model was designed to estimate the probability of BI rather than to define infection and disease activity as mutually exclusive clinical states. BI was defined by at least one of the following: positive microbiological culture, imaging evidence consistent with infection or clinical response to antibacterial treatment. Clinical response to antibacterial treatment was assessed retrospectively from the documented overall clinical course and was not defined by a prespecified numerical change in any single clinical or laboratory parameter. Rather, adjudication considered documented improvement in infection-compatible clinical manifestations following antibacterial therapy, with laboratory and/or imaging changes used as supportive evidence when available. Outcomes were initially extracted by one investigator and checked by a second investigator with clinical experience in rheumatology. Uncertain cases were discussed with a senior rheumatologist until consensus was reached. Formal blinding to predictor information was not feasible because of the retrospective design. Candidate predictors We collected demographic characteristics, previous hospitalisation for BI, fever, treatment history, SLE-DAS score and activity category, complete blood count, inflammatory markers, immunoglobulins, complement levels, anti-double-stranded DNA antibodies and peripheral lymphocyte subsets. Previous hospitalisation for BI was defined as any documented inpatient admission for BI before the index admission that was identifiable in the electronic medical records of the study hospital. The current electronic medical record system was introduced in 2019; therefore, prior infection-related hospitalisations were retrospectively ascertained from 2019 onwards. No additional fixed look-back window was imposed. Fever was defined as a documented body temperature of ≥37.3°C before or at admission. C4 hypocomplementemia was defined according to the lower limit of the study hospital laboratory reference range. SLE-DAS was categorised using guideline-based cut-offs: ≤2.08, >2.08 to ≤7.64 and >7.64.21 23 For model interpretation, these groups were labelled as low, intermediate and high SLE-DAS categories, respectively, rather than as formal remission or flare definitions. Treatment-related variables were considered as candidate predictors but were not retained in the final model after variable selection. Missing data and sample size Missing predictor values were handled using multiple imputation by chained equations with the mice package in R.24 Variables with more than 30% missingness were excluded from model development. The imputation model included the outcome and candidate predictors, while identifiers and free-text diagnostic fields were excluded. Five imputed datasets were generated for the primary analysis, and variable-level missingness before imputation is summarised in online supplemental table S7. The outcome was included in the imputation model for retrospective model development and validation performance estimation, consistent with the analytic objective of reducing bias in model evaluation. This does not represent implementation-time handling of missing predictors, for which complete predictor availability or a prespecified missing-input strategy would be required. In temporal validation, model performance was assessed across the imputed validation datasets and summarised using the final pooled model coefficients from the development cohort. Sensitivity analyses included complete-case temporal validation, predictor-only imputation in the temporal validation cohort and an additional analysis using 20 imputed datasets to assess the stability of predictor selection and model performance. Sample size adequacy for model development was assessed using the Riley approach,25 based on an anticipated C-statistic of 0.85, an outcome prevalence of 0.430, a Cox-Snell R-squared of 0.3415 and 17 candidate predictor parameters considered before LASSO regression after preliminary screening. The minimum required sample size was 377 observations with 163 outcome events. This calculation assumed a global shrinkage factor of approximately 0.90, a maximum acceptable difference of 0.05 between apparent and adjusted R-squared and a 0.05 margin of error for intercept estimation. The training cohort included 395 admissions and 170 BI events, meeting the required sample size. Model development A multistage variable selection strategy was used. Candidate predictors were prespecified according to clinical relevance, prior literature, routine availability and missingness. Univariable analyses were used descriptively to inform candidate reduction, rather than as the sole criterion for predictor eligibility. LASSO regression with 10-fold cross-validation and the 1-SE criterion was applied within each imputed dataset.26 The predictors selected by lambda.1se in each imputed dataset were recorded, and predictors selected in at least three of five imputed datasets were retained as the final predictor set. A multivariable logistic regression model containing this fixed predictor set was then fitted across the imputed datasets, and regression coefficients and standard errors were pooled using Rubin’s rules. Thus, multivariable logistic regression was used to estimate the final model coefficients rather than to perform additional predictor removal. Model coefficients were presented as ORs with 95% CIs, and the full logistic regression coefficients were provided to support model reproducibility. Because the final model was selected after LASSO penalisation followed by multivariable logistic regression, CIs for individual predictors may not fully account for variable selection uncertainty and should be interpreted with caution. PCT was modelled per 0.1 ng/mL for clinical interpretability. The model was displayed as a nomogram, and a web-based calculator was developed for supplementary clinical translation (online supplemental figure S2). Model evaluation and temporal validation Discrimination was assessed using receiver operating characteristic (ROC) curves and the area under the curve (AUC). Calibration was assessed using calibration plots and the Hosmer-Lemeshow test. Bootstrap internal validation with 200 resamples was repeated across the full model-building pipeline, including multiple imputation, candidate predictor processing, LASSO selection, final logistic regression fitting and performance estimation.27 Optimism was estimated as the difference between bootstrap apparent performance and performance when the bootstrap-derived model was applied to the original training cohort. Clinical utility was evaluated with decision curve analysis and net benefit.28 Decision curve analysis was interpreted as the potential net benefit of using the model to prompt further infection work-up, including microbiological testing, imaging reassessment, infectious-disease consultation or structured clinical reassessment, rather than as a direct rule for antibiotic initiation. In temporal validation, the final pooled model derived from the training cohort was applied unchanged across the imputed validation datasets, and performance was assessed using ROC analysis, calibration analysis, Brier score and decision curve analysis. Model performance was compared with CRP, PCT and the ESR/CRP ratio. Exploratory threshold-dependent metrics, including sensitivity, specificity, positive predictive value and negative predictive value, were calculated using the Youden index and reported in online supplemental tables. Threshold-dependent metrics were reported descriptively and should not be interpreted as validated decision thresholds. Logistic regression was also compared with random forest, Extreme Gradient Boosting (XGBoost), support vector machine (SVM) and artificial neural network (ANN) models using fivefold cross-validation; these machine learning comparisons were reported as supplementary analyses. Further analysis of SLE-DAS To examine how disease activity contributed to model interpretation, we plotted predicted BI probability across ordinal SLE-DAS categories by observed outcome. Restricted cubic spline analysis evaluated the association between continuous SLE-DAS score and BI probability. The overall p value tested the joint association of the spline terms with BI, whereas the p value for nonlinearity tested whether the association departed from linearity. Reporting This manuscript was prepared in accordance with Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis + Artificial Intelligence (TRIPOD+AI) guidance for reporting clinical prediction models.29 Patient and public involvement Patients and the public were not involved in the design, conduct, interpretation, reporting or dissemination plans of this retrospective study. Results Baseline characteristics The training cohort included 395 admissions from hospitalised patients with SLE, of which 170 were classified as BI and 225 as NA. The temporal validation cohort included 200 admissions, of which 75 were classified as BI and 125 as NA. Selected baseline characteristics are shown in table 1, and detailed clinical, treatment and laboratory variables are provided in online supplemental table S1. In the training cohort, patients in the BI group were older, had longer disease and treatment duration, and more often had fever or previous hospitalisation for BI. CRP, PCT, ESR and WBC were higher in the BI group. C4 hypocomplementemia and the high SLE-DAS category were more frequent in the NA group. In the temporal validation cohort, previous hospitalisation for BI, fever, CRP, PCT, C4 hypocomplementemia and ordinal SLE-DAS category showed similar between-group patterns, whereas age, disease duration, treatment duration, WBC and ESR were not significantly different. Among BI admissions, three admissions in the training cohort and two in the temporal validation cohort were classified as BI solely on the basis of clinical response to antibacterial treatment. These admissions were categorised as other infection sites because the specific infection site was uncertain. Infection sites and the proportion of multiple-site infections among BI admissions are summarised in online supplemental table S3. Peripheral lymphocyte subset variables are presented in online supplemental table S1 and were treated as exploratory immune-status markers. Table 1Selected baseline characteristics of the training and temporal validation cohortsCharacteristicsTraining BI(n=170)Training NA (n=225)P valueValidation BI(n=75)Validation NA (n=125)P valueMale, n (%)21 (12.4)25 (11.1)0.8246 (8.0)8 (6.4)0.886Age, years39.0 (31.0 to 52.8)33.0 (26.0 to 45.0)0.00238.0 (27.0 to 52.5)35.0 (27.0 to 47.0)0.362Disease duration, years8.0 (3.0 to 15.0)5.0 (2.0 to 10.0)<0.0017.0 (1.0 to 12.5)5.0 (0.8 to 10.0)0.137Treatment duration, years7.0 (2.0 to 14.0)5.0 (1.0 to 10.0)0.0146.0 (0.2 to 10.0)3.0 (0.0 to 9.0)0.097Previous hospitalisation for bacterial infection, n (%)78 (45.9)29 (12.9)<0.00136 (48.0)14 (11.2)<0.001Fever, n (%)70 (41.2)27 (12.0)<0.00129 (38.7)9 (7.2)<0.001WBC, 109/L6.15 (4.25 to 9.15)5.31 (3.64 to 7.26)<0.0015.24 (3.75 to 7.71)5.28 (3.38 to 7.46)0.527ESR, mm/hour39.0 (18.2 to 68.8)27.0 (13.0 to 55.0)0.00540.0 (19.5 to 72.5)28.0 (15.0 to 55.0)0.053CRP, mg/L14.80 (5.03 to 36.00)2.93 (1.60 to 7.30)<0.00113.80 (3.85 to 50.25)3.33 (1.85 to 7.65)<0.001Procalcitonin, ng/mL0.11 (0.05 to 0.27)0.05 (0.05 to 0.08)<0.0010.08 (0.04 to 0.15)0.05 (0.03 to 0.06)<0.001C4 hypocomplementemia, n (%)43 (25.3)104 (46.2)<0.00133 (44.0)76 (60.8)0.031SLE-DAS category, n (%)<0.001<0.001Low SLE-DAS category61 (35.9)37 (16.4)36 (48.0)18 (14.4)Intermediate SLE-DAS category29 (17.1)46 (20.4)8 (10.7)26 (20.8)High SLE-DAS category80 (47.1)142 (63.1)31 (41.3)81 (64.8)Data are median (IQR) unless otherwise stated. BI includes patients with bacterial infection, including those with concomitant active SLE. NA indicates non-infected active SLE. SLE-DAS categories were defined as low (SLE-DAS ≤2.08), intermediate (SLE-DAS >2.08 to ≤7.64) and high (SLE-DAS >7.64). P values compare BI and NA within each cohort; for SLE-DAS category, P values refer to the overall comparison across the three categories.BI, bacterial infection; CRP, C reactive protein; ESR, erythrocyte sedimentation rate; NA, non-infected active; SLE-DAS, Systemic Lupus Erythematosus Disease Activity Score; WBC, white blood cell. Selected baseline characteristics of the training and temporal validation cohorts Data are median (IQR) unless otherwise stated. BI includes patients with bacterial infection, including those with concomitant active SLE. NA indicates non-infected active SLE. SLE-DAS categories were defined as low (SLE-DAS ≤2.08), intermediate (SLE-DAS >2.08 to ≤7.64) and high (SLE-DAS >7.64). P values compare BI and NA within each cohort; for SLE-DAS category, P values refer to the overall comparison across the three categories. BI, bacterial infection; CRP, C reactive protein; ESR, erythrocyte sedimentation rate; NA, non-infected active; SLE-DAS, Systemic Lupus Erythematosus Disease Activity Score; WBC, white blood cell. Variable selection and final model After univariable screening, clinical filtering and LASSO selection across imputed datasets, the final predictor set included age, disease duration, previous hospitalisation for BI, fever, WBC, CRP, PCT, C4 hypocomplementemia and ordinal SLE-DAS category (figure 1). A multivariable logistic regression model containing this fixed predictor set was fitted across imputed datasets, and coefficients were pooled using Rubin’s rules. Previous hospitalisation for BI (OR 6.135, 95% CI 3.421 to 11.000), fever (OR 4.557, 95% CI 2.293 to 9.054), PCT per 0.1 ng/mL (OR 1.173, 95% CI 1.001 to 1.376), WBC (OR 1.091, 95% CI 1.014 to 1.174) and CRP (OR 1.024, 95% CI 1.008 to 1.041) were positively associated with BI. Age (OR 1.017, 95% CI 0.998 to 1.037) and disease duration (OR 1.025, 95% CI 0.986 to 1.066) were retained in the prediction model but showed weak and statistically non-significant associations. C4 hypocomplementemia (OR 0.517, 95% CI 0.294 to 0.911) and intermediate or high SLE-DAS categories also showed negative associations. The nomogram is shown in figure 2. The complete pooled logistic regression coefficients, including the intercept, are provided in online supplemental table S6. Model development and predictor contribution. (A) Rubin’s rules pooled multivariable logistic regression forest plot for the final predictor set. (B) LASSO variable-selection process, including cross-validation and coefficient-profile plots. (C) Predictor contribution ranked by absolute pooled Wald Z values. CRP, C reactive protein; LASSO, Least Absolute Shrinkage and Selection Operator; PCT, procalcitonin; SLE-DAS, Systemic Lupus Erythematosus Disease Activity Score; WBC, white blood cell. Nomogram for identifying bacterial infection among hospitalised patients with SLE. The nomogram estimates the probability of bacterial infection using routinely available clinical and laboratory predictors. SLE-DAS categories indicate low, intermediate and high SLE-DAS categories and should be interpreted within the bacterial infection versus non-infected active SLE differential-diagnosis context. The order of SLE-DAS categories in the nomogram reflects fitted model contribution rather than increasing clinical disease-activity severity. Previous hospitalisation for bacterial infection refers to a documented prior inpatient admission for bacterial infection identifiable in the study hospital electronic medical record system from 2019 onwards. CRP, C reactive protein; PCT, procalcitonin; SLE-DAS, Systemic Lupus Erythematosus Disease Activity Score; WBC, white blood cell. Model performance in the training cohort In the training cohort, the final model achieved an AUC of 0.869 (95% CI 0.833 to 0.905), exceeding CRP (AUC 0.771), PCT (AUC 0.682) and the ESR/CRP ratio (AUC 0.736) (figure 3). Calibration was acceptable, with a Hosmer-Lemeshow test p value of 0.133. Full-workflow bootstrap internal validation with 200 resamples yielded an optimism-corrected AUC of 0.837, corrected Brier score of 0.162, calibration slope of 0.829 and calibration intercept of −0.056 (online supplemental table S2). All 200 bootstrap resamples were completed successfully. Decision curve analysis showed higher net benefit for the final model than for single biomarkers, treat-all strategies or treat-none strategies across a broad range of threshold probabilities. Training-cohort threshold-dependent performance metrics are provided in online supplemental table S4. Predictor selection frequencies in the full-workflow bootstrap analysis are summarised in online supplemental table S8. Model performance in the training cohort. (A) Receiver operating characteristic curves comparing the final model with CRP, PCT and the ESR/CRP ratio. (B) Calibration plot. (C) AUC, sensitivity and specificity for the final model, CRP, PCT and the ESR/CRP ratio. (D) Decision curve analysis. AUC, area under the receiver operating characteristic curve; CRP, C reactive protein; DCA, decision curve analysis; ESR, erythrocyte sedimentation rate; PCT, procalcitonin; ROC, receiver operating characteristic. Comparison with machine learning models In fivefold cross-validation, logistic regression performed similarly to more complex machine learning models. AUCs were 0.844 for logistic regression, 0.844 for random forest, 0.824 for XGBoost, 0.839 for SVM and 0.834 for ANN. Decision curve analysis did not show a clear net benefit advantage for the machine learning models over logistic regression (online supplemental figure S1). Temporal validation In the temporal validation cohort, the final model retained good discrimination, with an AUC of 0.852 (95% CI 0.794 to 0.910) (figure 4). This exceeded CRP (AUC 0.729), PCT (AUC 0.719) and the ESR/CRP ratio (AUC 0.694). Calibration remained acceptable, with a Hosmer-Lemeshow test p value of 0.643, calibration slope of 0.850, calibration intercept of −0.049 and Brier score of 0.144. Decision curve analysis showed preserved net benefit in the temporal validation cohort. Detailed temporal validation performance and exploratory threshold-related metrics are provided in online supplemental table S5. Temporal validation of the final model. (A) Receiver operating characteristic curves comparing the final model with CRP, PCT and the ESR/CRP ratio in the temporal validation cohort. The final model AUC was 0.852, compared with 0.729 for CRP, 0.719 for PCT and 0.694 for the ESR/CRP ratio. (B) Calibration plot showing grouped observed probabilities, the logistic calibration curve and the ideal calibration line; the Hosmer-Lemeshow p value was 0.643, calibration slope was 0.850, calibration intercept was −0.049 and Brier score was 0.144. (C) Decision curve analysis. Values are summarised across five imputed validation datasets. AUC, area under the receiver operating characteristic curve; CRP, C reactive protein; ESR, erythrocyte sedimentation rate; PCT, procalcitonin; ROC, receiver operating characteristic. Sensitivity analyses Complete-case temporal validation was restricted to admissions with complete values for all final model predictors; because PCT was missing in 49 validation admissions, this analysis included 151 admissions (72 BI and 79 NA). Complete-case validation yielded an AUC of 0.840 (95% CI 0.774 to 0.906), calibration slope of 0.899, calibration intercept of 0.266, calibration-in-the-large of 0.326, Brier score of 0.162 and Hosmer-Lemeshow p value of 0.090. Predictor-only imputation in the temporal validation cohort yielded an AUC of 0.853 (95% CI 0.796 to 0.911), calibration slope of 0.966, calibration intercept of 0.025, calibration-in-the-large of 0.045, Brier score of 0.143 and Hosmer-Lemeshow p value of 0.616. An additional analysis using 20 imputed datasets selected the same final predictor set as the primary analysis, with training AUC 0.860 and temporal validation AUC 0.853. A first-admission-only sensitivity analysis could not be performed because no deidentified patient-level identifier was available. Detailed sensitivity analysis results are summarised in online supplemental table S9A,B. Further analysis of SLE-DAS Across low, intermediate and high SLE-DAS categories, predicted BI probability was higher in observed BI patients than in observed NA patients. Within the BI group, however, predicted probability did not increase monotonically across low, intermediate and high SLE-DAS categories. Restricted cubic spline analysis showed an overall association between continuous SLE-DAS and BI (P overall<0.001), with evidence that the association departed from linearity (P non-linear=0.014). The OR decreased as SLE-DAS increased and remained below one at higher SLE-DAS values (online supplemental figure S3). Discussion This retrospective study developed and temporally validated a clinical prediction model for identifying the presence of BI among hospitalised patients with SLE, using NA SLE as the comparator. The final model combined age, disease duration, previous hospitalisation for BI, fever, WBC, CRP, PCT, C4 hypocomplementemia and ordinal SLE-DAS category. It outperformed individual inflammatory biomarkers in the training cohort and retained good discrimination, calibration and net benefit in a temporal validation cohort. These findings support a multivariable approach to a common and clinically consequential differential-diagnosis problem in SLE. Previous hospitalisation for BI and fever were among the strongest clinical predictors associated with BI. Previous infection-related hospitalisation may capture cumulative susceptibility, including prior severe infection, recurrent healthcare exposure, organ damage, persistent immune dysfunction and treatment intensity.2–4 30 It was ascertained from the institutional electronic medical record from 2019 onwards without an additional fixed recency window. Because its predictive value may depend on recency, future validation should evaluate prespecified look-back windows or time-weighted definitions. Disease duration was also retained in the final model, although its association was weak and statistically non-significant. Rather than implying a direct causal effect, disease duration may reflect cumulative disease burden, prior treatment exposure, damage accrual or repeated healthcare contact. Fever is nonspecific in SLE, but its association with BI in this model supports systematic infection assessment rather than attributing fever solely to disease activity. Inflammatory markers contributed meaningful information but were insufficient as stand-alone discriminators. CRP may increase with BI, but it may also be influenced by serositis, arthritis or severe systemic inflammation.1 5 31 PCT is more infection-oriented in many clinical contexts, and recent SLE studies have continued to evaluate PCT, alone or combined with complement-related markers, for distinguishing BI from lupus flare.32 33 However, PCT may remain low in early, localised or mild infection and can also be affected by renal dysfunction or, rarely, severe lupus activity.34–36 Modelling CRP, PCT and WBC as continuous variables within a broader clinical model better reflects the heterogeneous presentation of hospitalised SLE. Peripheral lymphocyte subset variables were evaluated as candidate immune-status markers but were not retained in the final pooled prediction model. Exploratory baseline comparisons showed variable patterns in selected lymphocyte subset markers between BI and NA patients, which may reflect immune dysregulation, treatment-related lymphocyte perturbation, infection-related immune activation or disease activity context.37 38 Recent machine learning work in lupus nephritis has also examined peripheral lymphocyte subsets for co-infection prediction, supporting their potential exploratory value while underscoring the need for validation before routine use.19 However, these variables showed insufficient selection stability for inclusion in the final model. Therefore, lymphocyte subset findings should be interpreted as exploratory and hypothesis-generating rather than as part of the final prediction model specification. C4 hypocomplementemia and intermediate or high SLE-DAS categories showed negative associations with BI in the multivariable model. These findings may appear discordant with studies in which high disease activity predicts subsequent infection. The difference lies in the estimand and comparator. Previous infection-risk studies generally examined future infection susceptibility, whereas our model estimated the probability of concurrent BI among hospitalised patients with SLE, using NA disease as the comparator. Because the NA comparator group was defined by active, non-infected SLE, low complement and higher ordinal SLE-DAS categories were enriched in the disease activity-predominant state. In this setting, SLE-DAS provides context about disease activity predominance rather than a simple linear marker of infection probability.20 21 23 Restricted cubic spline analysis supported this interpretation by showing a nonlinear association between SLE-DAS and BI probability. Patients with BI and concomitant active SLE were included in the BI group because the model was designed to identify BI when infection and disease activity coexist. Future studies incorporating dynamic disease activity assessment, treatment changes and infection evidence grading may refine this classification. The nearly identical ORs for intermediate and high SLE-DAS categories should not be interpreted as evidence of a graded increase in infection probability from intermediate to high disease activity. Rather, once disease activity exceeded the low category, ordinal SLE-DAS mainly indicated a disease activity-predominant comparator state in this differential-diagnosis model. Full-workflow bootstrap internal validation yielded a more conservative optimism-corrected estimate, consistent with the additional uncertainty captured by repeating multiple imputation, LASSO selection and model refitting within each resample.39 The model nevertheless retained good discrimination in temporal validation. Logistic regression performed similarly to random forest, XGBoost, SVM and ANN models, while offering greater interpretability and easier clinical translation through a nomogram and web calculator. This is consistent with recent lupus-related machine learning work showing that immune and laboratory variables can be used for infection-related prediction, although model transportability and incremental benefit over interpretable regression approaches require further validation.19 Interpretability is particularly important for a bedside differential-diagnosis tool based on routinely available clinical variables. This study has several strengths. The comparator group consisted of non-infected patients with active SLE, which reflects the most difficult clinical scenario. Candidate predictors covered history, symptoms, inflammatory markers, complement status and disease activity, and all are routinely accessible in hospital practice. We also used full-workflow bootstrap internal validation, decision curve analysis, temporal validation, missing-data sensitivity analyses and exploratory SLE-DAS analyses, and provided full model coefficients, a nomogram and a web-based calculator to support reproducibility and clinical interpretability. Several limitations should be noted. First, this was a single-centre retrospective study with hospitalisation as the unit of analysis. Without a deidentified patient-level identifier, repeated admissions could not be modelled with within-patient clustering or restricted to first admissions; their impact cannot be quantified, and within-patient correlation may reduce effective independence and underestimate statistical uncertainty, while overlap between study periods could introduce some optimism into temporal validation. Second, BI was defined using a retrospective composite reference standard. Treatment response was clinically adjudicated rather than based on prespecified numerical criteria, so diagnostic certainty may have varied and some incorporation of clinical judgement is possible. Third, including concomitant active SLE in the BI group matched the prediction target but complicates interpretation of disease activity measures. Fourth, PCT was missing in 26.3% of training and 24.5% of temporal validation admissions. Although sensitivity analyses showed similar discrimination, PCT unavailability may limit implementation. Fifth, previous hospitalisation for BI was ascertainable only from the study-hospital electronic medical record from 2019 onwards; earlier or outside-hospital admissions may have been missed, and available look-back duration varied by index date. Its transportability and optimal recency window require external evaluation. Finally, despite full-workflow bootstrap, same-centre temporal validation and missing-data sensitivity analyses, residual optimism and centre-specific effects may remain. Outcome-informed imputation was used for retrospective evaluation; implementation requires complete predictors or a prespecified missing-input strategy. Multicentre external validation and prospective impact studies are required. Conclusion This study developed and temporally validated a clinical prediction model that identifies the presence of BI among hospitalised patients with SLE. The model showed good discrimination, acceptable calibration and clinical net benefit. It may support structured assessment of BI probability, but it should not replace comprehensive clinical judgement. SLE-DAS should be interpreted as disease activity context within this differential-diagnosis model and considered alongside infection-oriented clinical and laboratory information. Protocol and registration No formal protocol was prepared before data analysis. This retrospective study was retrospectively registered on OSF after completion of data collection and analysis for transparency. Registration DOI: 10.17605/OSF.IO/MZAC5. Supplementary material