Development and internal validation of a clinical prediction model for postoperative urinary tract infection in older surgical patients: a retrospective cohort study Authors: Xingyu Sun, Jiming Chen, Lijuan He, Shaohua Wang, Dan Li Published: 2026 ABSTRACT AbstractBackgroundPostoperative urinary tract infection (UTI) is a common complication among older surgical patients and is associated with prolonged hospitalization, increased healthcare utilization, and adverse postoperative outcomes. Early identification of patients at high risk of postoperative UTI remains challenging in routine clinical practice. This study aimed to develop and internally validate a clinically interpretable dynamic postoperative risk prediction model for postoperative UTI in older adult surgical patients using routinely available perioperative and early postoperative data.MethodsThis retrospective cohort study included patients aged ≥ 65 years who underwent inpatient surgery at a tertiary teaching hospital in Southwest China between January 2019 and December 2024. The primary outcome was postoperative UTI occurring during hospitalization or within 30 days after surgery. The model was designed as a dynamic postoperative risk estimation tool rather than a strictly preoperative prediction model. Patients were randomly divided into a training cohort (n = 239) and a validation cohort (n = 103). Candidate predictors were selected a priori based on clinical relevance and existing literature. Least absolute shrinkage and selection operator (LASSO) regression was used for feature selection, followed by multivariable logistic regression to construct the prediction model. Model discrimination, calibration, and clinical utility were assessed using the area under the receiver operating characteristic curve (AUC), calibration plots, and decision curve analysis, respectively. Model interpretability was evaluated using Shapley additive explanations (SHAP).ResultsAmong 239 patients in the training cohort, 48 (20.1%) developed postoperative UTI. Six variables were retained in the final multivariable model: diabetes, ICU admission, urinary catheter duration, recatheterization, length of hospital stay, and readmission within 30 days. A nomogram was constructed to facilitate individualized risk estimation. The model demonstrated good discrimination in the training cohort (AUC 0.886, 95% CI 0.836–0.936) and maintained acceptable performance in the validation cohort (AUC 0.830, 95% CI 0.736–0.924). Calibration plots showed good agreement between predicted and observed risks in both cohorts. Decision curve analysis indicated potential clinical utility across a range of threshold probabilities. SHAP analyses highlighted length of hospital stay and catheter-related variables as the most influential contributors to predicted UTI risk.ConclusionsWe developed and internally validated a clinically interpretable dynamic postoperative risk estimation model for postoperative UTI in older surgical patients using routinely available perioperative and postoperative data. The model demonstrated good discrimination, calibration, and potential clinical utility, and may support in-hospital postoperative risk stratification, catheter management review, intensified symptom surveillance, clinically indicated urine testing or physician reassessment, and reinforcement of modifiable preventive measures during postoperative care. Because several retained predictors are time-dependent postoperative variables, and because the outcome definition reflected routine clinical practice rather than a fully standardized surveillance definition, the model should not be interpreted as a strictly preoperative or early admission-time bedside prediction tool, but rather as a dynamic postoperative risk estimation framework for in-hospital risk stratification. Clinically actionable thresholds and implementation pathways require prospective validation before routine use.Keywords: Postoperative urinary tract infection, Prediction model, Older adults, Nomogram, Perioperative care Introduction Postoperative urinary tract infection (UTI) is one of the most common healthcare-associated infections in surgical patients and represents a substantial source of postoperative morbidity, prolonged hospitalization, and increased healthcare costs [1, 2]. Older adults are particularly vulnerable to postoperative UTI due to age-related physiological changes, multimorbidity, functional impairment, and frequent exposure to invasive procedures such as urinary catheterization [3, 4]. With the rapid aging of the global population, preventing postoperative UTI in older surgical patients has become an increasingly important clinical and public health priority. Previous studies have identified a range of factors associated with postoperative UTI, including advanced age, diabetes, chronic kidney disease, prolonged urinary catheter use, and extended length of hospital stay [5–7]. In older adults, these risk factors often coexist and interact, reflecting a complex interplay between baseline vulnerability, perioperative stress, and postoperative care processes. Despite growing awareness of these associations, most existing studies have focused on individual risk factors or specific surgical populations, limiting their ability to support individualized risk assessment in routine clinical practice [8]. In addition, the clinically relevant prediction time point—preoperative, early postoperative, or dynamic in-hospital assessment—has often not been clearly distinguished. Accurate identification of patients at increased risk of postoperative UTI during the perioperative and postoperative course is essential for guiding targeted preventive strategies, such as optimized catheter management, enhanced surveillance, and timely intervention. Clinical prediction models offer a promising approach to integrate multiple patient- and procedure-related variables into a single, individualized risk estimate [9]. However, existing UTI prediction tools are often derived from heterogeneous populations, lack internal or external validation, or rely on variables not readily available in routine perioperative settings, thereby limiting their clinical applicability [10, 11]. In addition, many previously proposed models provide limited interpretability, which may hinder clinician acceptance and real-world implementation. Older surgical patients represent a distinct and underrepresented population in perioperative risk modeling. Their care is frequently characterized by complex comorbidity profiles, intensive postoperative monitoring, and prolonged hospitalization, all of which may modify UTI risk trajectories compared with younger patients. Moreover, few studies have specifically examined postoperative UTI risk in older adults across a broad range of surgical procedures using routinely collected perioperative data from real-world clinical settings [12]. Therefore, the objective of this study was to develop and internally validate a clinically interpretable dynamic postoperative risk estimation model for postoperative UTI in older adult surgical patients using routinely available perioperative and postoperative variables. By applying penalized regression techniques, conventional multivariable modeling, and model interpretability methods, we aimed to construct a practical tool to support in-hospital postoperative risk stratification and inform targeted preventive strategies in perioperative care. We did not intend this model to serve as a strictly preoperative prediction tool, because some retained predictors may evolve during the postoperative course. Rather, the intended use of the model was dynamic postoperative risk estimation during hospitalization, not merely retrospective explanation after the full clinical course had unfolded. Methods Study design This study was a retrospective observational cohort study designed to develop and internally validate a dynamic postoperative risk prediction model for postoperative urinary tract infection (UTI) in older adult surgical patients. The key elements of the study design, including cohort assembly, outcome definition, predictor selection, model development, validation, and interpretability analyses, were prespecified prior to data analysis and are reported in accordance with the STROBE statement and contemporary recommendations for prediction model development. The intended use-case of the model was dynamic in-hospital postoperative risk estimation rather than strictly preoperative risk prediction. However, this study was conducted using a single-center dataset, and the internal validation strategy was based on a split-sample approach, which should be interpreted as an initial internal validation rather than a substitute for external validation. Accordingly, the model was intended to support repeated postoperative reassessment during hospitalization rather than one-time risk prediction at admission. Setting The study was conducted at a tertiary teaching hospital in Southwest China. Consecutive eligible patients were identified from the hospital’s electronic medical record (EMR) system. The study period spanned January 1, 2019, to December 31, 2024, covering patient recruitment, perioperative exposure assessment, postoperative outcome ascertainment, and follow-up. All clinical, laboratory, surgical, anesthetic, and outcome data were routinely collected as part of standard care during hospitalization. Participants Eligibility criteria and cohort assembly Adult patients aged ≥ 65 years who underwent inpatient surgical procedures during the study period were screened for eligibility. Patients were eligible for inclusion if they underwent surgery requiring postoperative inpatient care, had complete perioperative clinical records available in the electronic medical record system, and had no evidence of active urinary tract infection at the time of surgery. Patients were excluded if they had a documented preoperative urinary tract infection or were receiving antimicrobial treatment for UTI before surgery, underwent urological procedures specifically targeting the urinary tract, or had missing outcome data or key perioperative variables required for model development. After applying these criteria, a total of 342 eligible patients were included in the final analysis. For model development and internal validation, the study population was randomly allocated in a 7:3 ratio to a training (derivation) cohort (n = 239) and an independent validation cohort (n = 103) using a computer-generated randomization algorithm without stratification. Follow-up and outcome ascertainment Patients were followed from the date of surgery until hospital discharge and through 30 days postoperatively using inpatient records and readmission data. Postoperative UTI events were identified during hospitalization or within 30 days after surgery. Variables Outcome definition The primary outcome was postoperative urinary tract infection (UTI), defined as a clinically diagnosed UTI occurring after surgery during hospitalization or within 30 days postoperatively. Postoperative UTI events were identified through review of physician documentation in the electronic medical record and were defined according to routine institutional diagnostic practice rather than a formal surveillance protocol. Specifically, a case was considered postoperative UTI when the treating clinician documented a postoperative UTI diagnosis after surgery and initiated antimicrobial treatment, together with compatible urinary symptoms (e.g., dysuria, urgency, frequency, suprapubic discomfort) or systemic features suggestive of infection (e.g., fever or delirium without a more likely alternative source), and at least one supportive laboratory component, including a positive urine culture, pyuria, or bacteriuria. Because this was a retrospective real-world study, uniform application of a standardized surveillance definition such as CDC/NHSN criteria and microbiological confirmation for every case could not be ensured. Candidate predictors Candidate predictors were selected a priori based on clinical relevance and evidence from existing literature and encompassed a broad range of patient-related, clinical, and perioperative characteristics. These included demographic variables (age, sex, and body mass index); functional and comorbidity indicators (frailty status, activities of daily living dependence, ASA physical status classification, Charlson comorbidity index, diabetes, chronic kidney disease, chronic obstructive pulmonary disease, dementia, urinary incontinence, and history of urinary tract infection); preoperative laboratory parameters (serum albumin, hemoglobin, C-reactive protein, and white blood cell count); surgical and perioperative factors (surgery type, emergency surgery, anesthesia type, intraoperative opioid use, surgery duration, estimated blood loss, blood transfusion, and ICU admission); as well as urinary management and postoperative variables (urinary catheterization, catheter duration, recatheterization, urinary retention, appropriate antimicrobial prophylaxis, postoperative broad-spectrum antibiotic use, length of hospital stay, and 30-day readmission). Some retained variables, particularly catheter duration, length of hospital stay, and 30-day readmission, were recognized as postoperative or time-dependent indicators and were therefore interpreted as components of dynamic postoperative risk estimation rather than purely antecedent predictors. Data sources and measurement All variables were extracted from the EMR by trained investigators using a standardized data collection form. Laboratory values were obtained from the hospital laboratory information system. Surgical and anesthetic variables were abstracted from operative and anesthesia records. Outcome ascertainment was based on predefined chart-review rules applied consistently across the training and validation cohorts, although the underlying clinical diagnosis reflected routine institutional practice rather than a formal surveillance algorithm. Outcome assessment and exposure measurement were conducted using identical methods across the training and validation cohorts, ensuring comparability between groups. Bias Several measures were taken to minimize potential sources of bias. Selection bias was reduced by including consecutive eligible patients over a predefined study period. Outcome ascertainment was based on physician-documented diagnoses supported by clinical and laboratory information extracted from the electronic medical record using a standardized data collection procedure. However, because postoperative UTI was identified according to routine clinical practice rather than a fully standardized surveillance definition, some degree of heterogeneity in symptom recognition, urine culture ordering, and documentation practices may have remained, with the potential for ascertainment bias and outcome misclassification. To reduce overfitting and multicollinearity, predictor selection was performed using penalized regression techniques. The use of an independent validation cohort further limited optimism in model performance estimates. Study size The study size was determined by the number of eligible patients meeting the inclusion criteria during the predefined study period. No formal a priori sample size calculation based on hypothesis testing was performed, as this study focused on prediction model development rather than effect estimation. Instead, sample size considerations were evaluated in light of contemporary methodological recommendations for multivariable prediction modeling. In the derivation cohort, 48 postoperative UTI events were observed. After penalized variable selection, the final multivariable model retained six predictors, corresponding to an approximate events-per-variable ratio of 8. We recognize that this should be regarded as borderline rather than ideal for stable model development, particularly in view of the wide confidence intervals observed for some predictors. Accordingly, the present model should be interpreted as an internally validated preliminary model that requires confirmation and updating in larger external datasets. The relatively large candidate predictor set in relation to the number of outcome events should also be considered when interpreting model stability and the potential for residual overfitting. Quantitative variables Continuous variables were analyzed on their original scales. To improve readability and clinical interpretability of regression coefficients, selected continuous predictors with relatively large original units were rescaled for presentation (specifically, blood loss per 100 mL, hemoglobin per 10 g/L, and C-reactive protein per 10 mg/L), so that the corresponding odds ratios would not appear artificially close to 1. Other continuous variables were modeled per 1-unit increase as indicated. No arbitrary categorization of continuous variables was performed unless clinically justified. Statistical methods Model development and confounding control Because the candidate predictor set included both baseline/perioperative variables and evolving postoperative variables, the resulting model was conceptualized as a dynamic postoperative risk estimation model rather than a strictly preoperative prediction model. Baseline characteristics were summarized using means (standard deviations) or medians (interquartile ranges) for continuous variables and counts (percentages) for categorical variables. Group comparisons were performed using the Student’s t test or Mann–Whitney U test for continuous variables and the χ² test or Fisher’s exact test for categorical variables, as appropriate. Candidate predictors were first screened using least absolute shrinkage and selection operator (LASSO) logistic regression with 10-fold cross-validation. The optimal penalty parameter (λ) was selected at the value minimizing the cross-validated error (λ_min). Variables with non-zero coefficients were entered into a multivariable logistic regression model to estimate adjusted odds ratios (ORs) and 95% confidence intervals (CIs). Model validation and performance Model discrimination was assessed using the area under the receiver operating characteristic curve (AUC). Calibration was evaluated using calibration plots comparing predicted and observed risks. In addition to the split-sample derivation/validation strategy, bootstrap resampling was used as an additional internal validation approach to assess optimism in model performance and to generate bias-corrected calibration estimates. Clinical utility was examined using decision curve analysis (DCA) across a range of threshold probabilities. Because no prospectively validated intervention cutoff was available, the DCA was interpreted as exploratory and intended to inform pragmatic threshold ranges for postoperative reassessment rather than to define a single actionable decision threshold. All performance metrics were evaluated in both the training and validation cohorts and interpreted cautiously in view of the modest number of outcome events. We acknowledge that the split-sample validation approach may be statistically inefficient and may overestimate model performance compared with resampling-based or external validation strategies. Missing data and sensitivity Because the analysis relied on routinely collected clinical data, missingness was minimal. A complete-case analysis was performed. Given the low proportion of missing data, no imputation was applied. Model interpretability To enhance transparency and interpretability, Shapley additive explanations (SHAP) were used as a complementary interpretability approach to the multivariable logistic regression model, quantifying the contribution of each predictor to individual risk predictions and ranking global feature importance. SHAP analyses were intended to aid model interpretation rather than replace regression-based effect estimation, and SHAP values should not be interpreted as causal effects. All statistical analyses were performed using R software (version 4.3.1; R Foundation for Statistical Computing, Vienna, Austria), and a two-sided P value < 0.05 was considered statistically significant unless otherwise specified. Results Baseline characteristics of the training and validation sets The baseline characteristics of the study population stratified by the training and validation sets are summarized in Table 1. A total of 342 older adult patients were included, of whom 239 were assigned to the training set and 103 to the validation set. Overall, the two cohorts were well balanced with respect to demographic characteristics, comorbidities, laboratory parameters, and perioperative variables. The mean age was comparable between the training and validation sets (73.23 ± 4.88 vs. 72.69 ± 5.48 years, P = 0.362), and the proportion of female patients was identical in both cohorts (54.4% in each group, P = 1.000). No significant differences were observed in body mass index, frailty status, ADL dependence, ASA class, or Charlson comorbidity index (all P > 0.05). Similarly, the prevalence of major comorbid conditions, including diabetes, chronic kidney disease, chronic obstructive pulmonary disease, dementia, urinary incontinence, and history of urinary tract infection, did not differ significantly between the two groups (all P > 0.05). Preoperative laboratory values, such as serum albumin, hemoglobin, C-reactive protein, and white blood cell count, were also comparable between the training and validation sets. With regard to perioperative characteristics, no statistically significant differences were identified in surgery duration, estimated blood loss, ICU admission, urinary catheterization, catheter duration, recatheterization, length of hospital stay, or 30-day readmission rate between the two cohorts (all P > 0.05). Taken together, these findings indicate a high degree of similarity between the training and validation sets, supporting the adequacy of the data partition and minimizing the risk of selection bias in subsequent model development and validation analyses. Table 1Comparison of baseline characteristics between the training and validation setsVariableTotal (n = 342)Training (n = 239)Validation (n = 103)P valueDemographics Age, years73.07 (5.06)73.23 (4.88)72.69 (5.48)0.362 Sex, female186 (54.4)130 (54.4)56 (54.4)1.000 Body mass index, kg/m²23.58 (3.27)23.54 (3.19)23.65 (3.45)0.770 Frailty status0.64 (0.72)0.63 (0.72)0.67 (0.72)0.619 ADL dependence, yes57 (16.7)40 (16.7)17 (16.5)1.000 ASA class2.91 (0.70)2.92 (0.71)2.90 (0.68)0.871 Charlson comorbidity index2.46 (1.59)2.37 (1.50)2.66 (1.77)0.120Comorbidities Diabetes, yes92 (26.9)64 (26.8)28 (27.2)1.000 Chronic kidney disease, yes64 (18.7)49 (20.5)15 (14.6)0.254 COPD, yes32 (9.4)23 (9.6)9 (8.7)0.956 Dementia, yes28 (8.2)20 (8.4)8 (7.8)1.000 Urinary incontinence, yes77 (22.5)58 (24.3)19 (18.4)0.298 History of UTI, yes106 (31.0)69 (28.9)37 (35.9)0.244Laboratory values Albumin, g/L37.41 (4.41)37.38 (4.37)37.49 (4.49)0.831 Hemoglobin, g/L123.56 (16.18)123.77 (16.16)123.07 (16.28)0.712 C-reactive protein, mg/L4.18 (4.16)4.12 (4.08)4.33 (4.35)0.665 White blood cell count, ×10⁹/L6.65 (1.82)6.75 (1.80)6.44 (1.87)0.145Perioperative characteristics Surgery duration, min169.60 (71.09)170.77 (72.66)166.89 (67.20)0.643 Blood loss, mL296.86 (198.40)300.19 (203.92)289.28 (185.82)0.642 ICU admission, yes72 (21.1)46 (19.2)26 (25.2)0.270 Urinary catheter, yes292 (85.4)209 (87.4)83 (80.6)0.138 Catheter duration, days2.36 (1.79)2.46 (1.82)2.12 (1.70)0.099 Recatheterization, yes16 (4.7)13 (5.4)3 (2.9)0.462 Length of stay, days8.74 (3.40)8.50 (3.33)9.27 (3.48)0.054 Readmission within 30 days, yes26 (7.6)20 (8.4)6 (5.8)0.554Data are presented as mean (standard deviation) for continuous variables and number (percentage) for categorical variables. P values compare the training and validation sets using the Student’s t-test or Mann–Whitney U test for continuous variables and the χ² test or Fisher’s exact test for categorical variables, as appropriate. ADL Activities of daily living, ASA American Society of Anesthesiologists, COPD Chronic obstructive pulmonary disease, ICU Intensive care unit, UTI Urinary tract infection Comparison of baseline characteristics between the training and validation sets Data are presented as mean (standard deviation) for continuous variables and number (percentage) for categorical variables. P values compare the training and validation sets using the Student’s t-test or Mann–Whitney U test for continuous variables and the χ² test or Fisher’s exact test for categorical variables, as appropriate. ADL Activities of daily living, ASA American Society of Anesthesiologists, COPD Chronic obstructive pulmonary disease, ICU Intensive care unit, UTI Urinary tract infection Baseline characteristics according to postoperative UTI status Baseline characteristics of patients in the derivation set stratified by postoperative urinary tract infection (UTI) status are presented in Table 2. Among the 239 patients included in the derivation cohort, 48 (20.1%) developed postoperative UTI, whereas 191 (79.9%) did not. Patients who developed postoperative UTI were significantly older than those without UTI (74.67 ± 4.41 vs. 72.87 ± 4.94 years, P = 0.023). In addition, the UTI group had a higher body mass index (24.38 ± 3.65 vs. 23.33 ± 3.04 kg/m², P = 0.041). Although the proportion of female patients was higher in the UTI group, this difference did not reach statistical significance (66.7% vs. 51.3%, P = 0.081). With respect to comorbidities, patients with postoperative UTI exhibited a significantly higher prevalence of diabetes (41.7% vs. 23.0%, P = 0.015) and chronic kidney disease (33.3% vs. 17.3%, P = 0.024). No significant differences were observed between the two groups in terms of other comorbid conditions, including chronic obstructive pulmonary disease, dementia, urinary incontinence, or history of UTI. Preoperative laboratory parameters were largely comparable between groups. However, patients with UTI had a lower preoperative white blood cell count compared with those without UTI (6.29 ± 1.57 vs. 6.87 ± 1.84 × 10⁹/L, P = 0.045), while no significant differences were found in serum albumin, hemoglobin, or C-reactive protein levels. Regarding perioperative characteristics, most surgical and anesthetic variables did not differ significantly between the two groups. Notably, patients who developed postoperative UTI had a significantly longer duration of urinary catheterization (3.21 ± 1.87 vs. 2.28 ± 1.77 days, P = 0.001) and a substantially prolonged length of hospital stay (11.56 ± 3.49 vs. 7.73 ± 2.82 days, P < 0.001). Furthermore, the 30-day readmission rate was significantly higher among patients with UTI than among those without UTI (18.8% vs. 5.8%, P = 0.009). Overall, these findings suggest that older age, higher body mass index, selected comorbidities, prolonged urinary catheterization, and extended hospitalization are more frequently observed among patients who develop postoperative UTI, supporting their consideration as candidate predictors for subsequent multivariable modeling. Table 2Baseline characteristics in the derivation set according to postoperative UTI statusVariableNon-UTI (n = 191)UTI (n = 48)P valueDemographics Age, years72.87 (4.94)74.67 (4.41)0.023 Sex, female98 (51.3)32 (66.7)0.081 Body mass index, kg/m²23.33 (3.04)24.38 (3.65)0.041Frailty / functional status Frailty status0.60 (0.72)0.73 (0.71)0.276 ADL dependence, yes30 (15.7)10 (20.8)0.526 ASA class2.88 (0.71)3.04 (0.71)0.172 Charlson comorbidity index2.28 (1.43)2.71 (1.71)0.079Comorbidities Diabetes, yes44 (23.0)20 (41.7)0.015 Chronic kidney disease, yes33 (17.3)16 (33.3)0.024 COPD, yes17 (8.9)6 (12.5)0.630 Dementia, yes15 (7.9)5 (10.4)0.778 Urinary incontinence, yes44 (23.0)14 (29.2)0.486 History of UTI, yes52 (27.2)17 (35.4)0.346Preoperative laboratory values Albumin, g/L37.50 (4.43)36.88 (4.15)0.377 Hemoglobin, g/L123.08 (15.36)126.52 (18.96)0.188 C-reactive protein, mg/L4.00 (4.05)4.58 (4.21)0.385 White blood cell count, ×10⁹/L6.87 (1.84)6.29 (1.57)0.045Surgical / perioperative characteristics Surgery type0.160 General63 (33.0)13 (27.1) Gynecologic24 (12.6)3 (6.2) Orthopedic43 (22.5)9 (18.8) Other31 (16.2)8 (16.7) Urologic16 (8.4)6 (12.5) Vascular14 (7.3)9 (18.8) Emergency surgery, yes24 (12.6)5 (10.4)0.873 General anesthesia, yes142 (74.3)33 (68.8)0.548 Opioid use, yes128 (67.0)33 (68.8)0.955 Surgery duration, min169.61 (73.50)175.42 (69.75)0.621 Blood loss, mL288.88 (193.79)345.17 (237.00)0.087 Transfusion, yes22 (11.5)8 (16.7)0.472 ICU admission, yes36 (18.8)10 (20.8)0.915 Urinary catheter, yes163 (85.3)46 (95.8)0.086 Catheter duration, days2.28 (1.77)3.21 (1.87)0.001 Recatheterization, yes8 (4.2)5 (10.4)0.179 Urinary retention, yes31 (16.2)12 (25.0)0.229 Appropriate prophylaxis, yes146 (76.4)32 (66.7)0.229 Postoperative broad-spectrum antibiotics, yes28 (14.7)4 (8.3)0.361 Preoperative pyuria, yes23 (12.0)10 (20.8)0.179 Preoperative bacteriuria, yes12 (6.3)6 (12.5)0.249 Length of stay, days7.73 (2.82)11.56 (3.49)< 0.001 Readmission within 30 days, yes11 (5.8)9 (18.8)0.009Data are presented as mean (standard deviation) for continuous variables and number (percentage) for categorical variables. P values compare the UTI and non-UTI groups using the Student’s t-test or Mann–Whitney U test for continuous variables and the χ² test or Fisher’s exact test for categorical variables, as appropriate. ADL Activities of daily living, ASA American Society of Anesthesiologists, COPD Chronic obstructive pulmonary disease, ICU Intensive care unit, UTI Urinary tract infection Baseline characteristics in the derivation set according to postoperative UTI status Data are presented as mean (standard deviation) for continuous variables and number (percentage) for categorical variables. P values compare the UTI and non-UTI groups using the Student’s t-test or Mann–Whitney U test for continuous variables and the χ² test or Fisher’s exact test for categorical variables, as appropriate. ADL Activities of daily living, ASA American Society of Anesthesiologists, COPD Chronic obstructive pulmonary disease, ICU Intensive care unit, UTI Urinary tract infection Feature selection using LASSO regression The candidate predictors selected by least absolute shrinkage and selection operator (LASSO) regression are presented in Table 3. Using LASSO logistic regression with 10-fold cross-validation, the optimal penalty parameter was determined at the value that minimized the cross-validated error (λmin). At this penalty level, a total of 22 variables with non-zero coefficients were retained as candidate predictors for postoperative urinary tract infection. The selected variables encompassed a broad range of patient-related, clinical, and perioperative characteristics. Several perioperative and postoperative factors demonstrated relatively larger coefficients, including recatheterization, ICU admission, diabetes, 30-day readmission, urinary catheterization, and emergency surgery, suggesting a stronger contribution to postoperative UTI risk within the penalized model. In addition, indicators reflecting disease burden and hospitalization course, such as length of hospital stay, chronic kidney disease, catheter duration, and Charlson comorbidity index, were also retained.Demographic and functional variables, including female sex, age, body mass index, ADL dependence, and ASA class, were selected despite having comparatively smaller coefficients, indicating potential additive predictive value when considered jointly with other clinical factors. Furthermore, several laboratory and treatment-related variables, such as preoperative white blood cell count, C-reactive protein, hemoglobin level, intraoperative blood loss, appropriate antimicrobial prophylaxis, and postoperative broad-spectrum antibiotic use, were retained with low-magnitude coefficients, reflecting their possible contribution to risk discrimination in a multivariable predictive context. Importantly, as a penalized regression approach, LASSO was applied to reduce dimensionality and mitigate multicollinearity among candidate variables rather than to assess statistical significance. Therefore, the retained predictors represent variables with potential relevance for risk prediction and were subsequently considered for further multivariable modeling and nomogram construction, rather than being interpreted as independent risk factors. Table 3Candidate predictors with non-zero coefficients selected by LASSO regression for postoperative UTIVariableLASSO coefficientRecatheterization0.885ICU admission−0.880Diabetes0.878Readmission within 30 days0.866Urinary catheter use0.640Emergency surgery−0.461Length of hospital stay, days0.400Chronic kidney disease0.369Appropriate antimicrobial prophylaxis−0.297Female sex0.204Catheter duration, days0.201Preoperative white blood cell count, ×10⁹/L−0.093Charlson comorbidity index0.083ADL dependence0.067ASA class0.056Body mass index, kg/m²0.055Postoperative broad-spectrum antibiotics−0.041Age, years0.038Urinary retention0.028Preoperative C-reactive protein, mg/L0.022Preoperative hemoglobin, g/L0.016Blood loss, mL0.001Variables were selected using Least absolute shrinkage and selection operator (LASSO) logistic regression with 10-fold cross-validation. The tuning parameter (λ) was chosen at the value that minimized the cross-validated error (λmin). Only variables with non-zero coefficients at λmin are shown. Coefficients represent log-odds and indicate the direction and relative contribution of each predictor to postoperative urinary tract infection risk. LASSO regression is a feature selection method and does not provide estimates of statistical significance; therefore, these coefficients should not be interpreted as odds ratios or inferential effects. ADL Activities of daily living, ASA American Society of Anesthesiologists, ICU Intensive care unit, UTI Urinary tract infection Candidate predictors with non-zero coefficients selected by LASSO regression for postoperative UTI Variables were selected using Least absolute shrinkage and selection operator (LASSO) logistic regression with 10-fold cross-validation. The tuning parameter (λ) was chosen at the value that minimized the cross-validated error (λmin). Only variables with non-zero coefficients at λmin are shown. Coefficients represent log-odds and indicate the direction and relative contribution of each predictor to postoperative urinary tract infection risk. LASSO regression is a feature selection method and does not provide estimates of statistical significance; therefore, these coefficients should not be interpreted as odds ratios or inferential effects. ADL Activities of daily living, ASA American Society of Anesthesiologists, ICU Intensive care unit, UTI Urinary tract infection Univariable and multivariable logistic regression analyses Univariable and multivariable logistic regression analyses for postoperative urinary tract infection (UTI) are summarized in Table 4 and visualized in Fig. 1. In univariable analyses, diabetes was associated with increased odds of postoperative UTI (OR 2.39, 95% CI 1.22–4.64; P = 0.010), as were readmission within 30 days (OR 3.78, 95% CI 1.43–9.75; P = 0.006) and chronic kidney disease (OR 2.39, 95% CI 1.16–4.83; P = 0.016). Longer length of hospital stay (per day: OR 1.48, 95% CI 1.32–1.69; P < 0.001) and longer catheter duration (per day: OR 1.31, 95% CI 1.11–1.56; P = 0.002) were also positively associated with UTI risk. Higher body mass index showed a modest association (per 1 kg/m²: OR 1.11, 95% CI 1.00–1.23; P = 0.042), and older age was associated with increased odds (per year: OR 1.08, 95% CI 1.01–1.15; P = 0.024). Preoperative white blood cell count demonstrated an inverse association (OR 0.83, 95% CI 0.69–0.99; P = 0.047). Several additional variables showed nonsignificant trends (e.g., recatheterization, female sex, urinary catheter use) in univariable testing (Table 4). Variables entered into the multivariable model were selected using least absolute shrinkage and selection operator (LASSO) regression (λ = lambda.min), and effect estimates from the adjusted model are presented in Table 4; Fig. 1. In the multivariable model, diabetes remained an independent risk factor (adjusted OR 5.45, 95% CI 1.94–16.88; P = 0.002), as did recatheterization (adjusted OR 7.08, 95% CI 1.08–47.52; P = 0.040). Increased length of hospital stay was independently associated with higher odds of postoperative UTI (per day: adjusted OR 1.80, 95% CI 1.48–2.29; P < 0.001), and longer catheter duration also remained significant (per day: adjusted OR 1.47, 95% CI 1.07–2.05; P = 0.019). By contrast, ICU admission was associated with lower observed odds of postoperative UTI in the adjusted model (adjusted OR 0.18, 95% CI 0.03–0.73; P = 0.025). Readmission within 30 days showed a borderline association after adjustment (adjusted OR 4.90, 95% CI 1.00–26.02; P = 0.053) (Table 4; Fig. 1). For readability and interpretability, selected continuous predictors were rescaled for display in Table 4 (blood loss per 100 mL, hemoglobin per 10 g/L, and C-reactive protein per 10 mg/L), whereas other continuous variables were modeled per 1-unit increase as indicated. Table 4Univariable and multivariable logistic regression analyses for postoperative urinary tract infectionVariableUnivariable OR95% CIP valueAdjusted OR95% CIP valueRecatheterization2.660.77–8.380.1007.081.08–47.520.040ICU admission1.130.50–2.420.7550.180.03–0.730.025Diabetes2.391.22–4.640.0105.451.94–16.880.002Readmission within 30 days3.781.43–9.750.0064.901.00–26.020.053Urinary catheter use3.951.13–25.040.0674.000.43–81.460.279Emergency surgery0.810.26–2.090.6840.310.04–1.610.195Length of hospital stay, days1.481.32–1.69< 0.0011.801.48–2.29< 0.001Chronic kidney disease2.391.16–4.830.0162.150.66–7.070.201Appropriate prophylaxis0.620.31–1.250.1670.550.18–1.700.296Female sex1.900.99–3.760.0581.930.70–5.660.216Catheter duration, days1.311.11–1.560.0021.471.07–2.050.019Preoperative WBC, ×10⁹/L0.830.69–0.990.0470.800.59–1.060.136Charlson comorbidity index1.200.98–1.470.0821.230.90–1.690.193ADL dependence1.410.61–3.060.3971.800.47–6.620.381ASA class1.370.87–2.150.1731.280.64–2.600.488Body mass index, kg/m²1.111.00–1.230.0421.090.93–1.270.286Postoperative broad-spectrum antibiotics0.530.15–1.440.2570.730.12–3.400.701Age, per year1.081.01–1.150.0241.060.96–1.180.223Urinary retention1.720.78–3.610.1611.360.39–4.530.624Preoperative CRP, per 10 mg/L1.030.95–1.100.3931.090.97–1.210.095Preoperative hemoglobin, per 10 g/L1.010.99–1.030.1891.031.00–1.070.080Blood loss, per 100 mL1.0011.000–1.0030.0901.0021.000–1.0050.054Odds ratios (ORs) and 95% Confidence intervals (CIs) were estimated using univariable and multivariable logistic regression analyses. Variables included in the multivariable model were selected based on Least absolute shrinkage and selection operator (LASSO) regression (λ = lambda.min). Continuous variables were analyzed per 1-unit increase unless otherwise specified; blood loss was expressed per 100 mL increase, preoperative hemoglobin per 10 g/L increase, and preoperative C-reactive protein per 10 mg/L increase to improve readability and interpretability of effect estimates. Bold values indicate statistical significance (P < 0.05). UTI Urinary tract infection, ICU Intensive care unit, WBC White blood cell count, ADL Activities of daily living, ASA American Society of Anesthesiologists Univariable and multivariable logistic regression analyses for postoperative urinary tract infection Odds ratios (ORs) and 95% Confidence intervals (CIs) were estimated using univariable and multivariable logistic regression analyses. Variables included in the multivariable model were selected based on Least absolute shrinkage and selection operator (LASSO) regression (λ = lambda.min). Continuous variables were analyzed per 1-unit increase unless otherwise specified; blood loss was expressed per 100 mL increase, preoperative hemoglobin per 10 g/L increase, and preoperative C-reactive protein per 10 mg/L increase to improve readability and interpretability of effect estimates. Bold values indicate statistical significance (P < 0.05). UTI Urinary tract infection, ICU Intensive care unit, WBC White blood cell count, ADL Activities of daily living, ASA American Society of Anesthesiologists Forest plot of univariable and multivariable logistic regression analyses for postoperative urinary tract infection. Legend: The figure displays odds ratios (ORs) with 95% Confidence intervals (CIs) from univariable (blue) and multivariable (red) logistic regression models evaluating factors associated with postoperative Urinary tract infection (UTI). The vertical dashed line indicates the null value (OR = 1). For readability and interpretability, selected continuous variables were rescaled as follows: blood loss per 100 mL, hemoglobin per 10 g/L, and C-reactive protein (CRP) per 10 mg/L; this rescaling was performed solely for presentation so that effect estimates would be easier to interpret. Other continuous predictors are expressed per 1-unit increase as indicated. Confidence intervals extending beyond the plotting range are truncated and denoted by arrows. Abbreviations: ADL, Activities of daily living; ASA, American Society of Anesthesiologists; CRP, C-reactive protein; ICU, Intensive care unit; WBC, White blood cell count; UTI, Urinary tract infection Development of the prediction nomogram A nomogram was constructed based on the final multivariable model to facilitate individualized dynamic postoperative risk estimation during the in-hospital course (Fig. 2). The nomogram assigns point values to each predictor (diabetes, ICU admission, catheter duration, recatheterization, length of stay, and readmission within 30 days), which are summed to yield total points and mapped to the predicted probability of postoperative UTI (Fig. 2). A clinical nomogram was constructed from the final multivariable logistic regression model to estimate an individual patient’s dynamic postoperative risk of urinary tract infection (UTI) during the in-hospital course and to support pragmatic risk-stratified postoperative assessment. For each predictor (diabetes, ICU admission, catheter duration in days, recatheterization, length of hospital stay in days, and readmission within 30 days), locate the patient’s value on the corresponding axis and project upward to the Points scale to assign points. Sum the points across predictors to obtain Total Points, then project downward to the Linear Predictor and Risk scales to derive the predicted probability of postoperative UTI. Abbreviations: ICU, intensive care unit; UTI, urinary tract infection. Model performance, calibration, and clinical utility In the training cohort, the model demonstrated good discrimination (AUC 0.886, 95% CI 0.836–0.936; Fig. 3A), and its discriminative ability remained acceptable in the validation cohort (AUC 0.830, 95% CI 0.736–0.924; Fig. 4A). Calibration plots showed overall agreement between predicted and observed risks, with the apparent and bias-corrected curves closely tracking the ideal reference line in both cohorts (Figs. 3B and 4B). The bias-corrected curves were derived using bootstrap resampling to reduce optimism in internal performance assessment. Decision-curve analysis further indicated potential clinical utility, as the model provided greater net benefit than “treat-all” or “treat-none” strategies across a range of clinically relevant threshold probabilities in both the training and validation cohorts (Figs. 3C and 4C). Because no empirically validated cutoff was available, these DCA findings were interpreted as supportive of pragmatic risk-stratified postoperative reassessment rather than a single definitive intervention threshold. Discrimination, calibration, and decision-curve performance of the postoperative UTI prediction model in the training cohort. (A) Receiver operating characteristic (ROC) curve evaluating model discrimination for postoperative urinary tract infection (UTI) in the training cohort, with the area under the curve (AUC) and its 95% confidence interval (CI) annotated. The diagonal dashed line indicates no discrimination. (B) Calibration plot in the training cohort comparing predicted probabilities with observed event rates. The “apparent” curve reflects calibration in the training data, and the “bias-corrected” curve represents bootstrap-based optimism-adjusted calibration; the dashed line denotes perfect calibration. (C) Decision curve analysis (DCA) in the training cohort showing the net benefit of the full model across a range of risk thresholds, compared with default strategies of treating all (“All”) or treating none (“None”); curves for individual predictors are displayed for reference. The DCA was interpreted as supportive of pragmatic risk-stratified postoperative reassessment rather than as defining a single validated intervention cutoff. Abbreviations: AUC, area under the curve; CI, confidence interval; DCA, decision curve analysis; ROC, receiver operating characteristic; UTI, urinary tract infection. (A) Receiver operating characteristic (ROC) curve evaluating model discrimination for postoperative urinary tract infection (UTI) in the validation cohort, with the area under the curve (AUC) and its 95% confidence interval (CI) annotated. The diagonal dashed line indicates no discrimination. (B) Calibration plot in the validation cohort comparing predicted probabilities with observed event rates. The “apparent” curve reflects calibration in the validation data, and the “bias-corrected” curve represents bootstrap-based optimism-adjusted calibration; the dashed line denotes perfect calibration. (C) Decision curve analysis (DCA) in the validation cohort showing the net benefit of the full model across a range of risk thresholds, compared with default strategies of treating all (“All”) or treating none (“None”); curves for individual predictors are displayed for reference. The DCA was interpreted as supportive of pragmatic risk-stratified postoperative reassessment rather than as defining a single validated intervention cutoff. Abbreviations: AUC, area under the curve; CI, confidence interval; DCA, decision curve analysis; ROC, receiver operating characteristic; UTI, urinary tract infection. Model interpretability using SHAP Model interpretability was further examined using SHAP (Shapley additive explanations) methods (Fig. 5). The SHAP summary plot illustrated the direction and magnitude of each predictor’s contribution to individual predictions (Fig. 5A), while global feature importance (mean absolute SHAP values) ranked length of stay as the most influential predictor, followed by ICU admission, catheter duration, and diabetes, with readmission within 30 days and recatheterization contributing to a lesser extent (Fig. 5B). Overall, these interpretability analyses were consistent with the regression results and served as a complementary explanation layer for the fitted model, highlighting urinary management and healthcare exposure intensity (length of stay and catheterization-related variables) as major contributors to predicted postoperative UTI risk (Table 4; Figs. 1 and 5), rather than as evidence of causal effects. SHAP-based interpretation of the postoperative UTI prediction model. Legend: (A) SHAP (Shapley additive explanations) summary (beeswarm) plot showing the distribution of feature contributions across individuals. Each dot represents one patient. The x-axis denotes the SHAP value (impact on the model output), where positive values increase the predicted probability of postoperative urinary tract infection (UTI) and negative values decrease it. Dot color reflects the corresponding feature value (low to high). Features are ordered by overall importance. (B) Bar plot of global feature importance quantified by the mean absolute SHAP value (mean |SHAP|), indicating each predictor’s average contribution magnitude to model predictions. SHAP values are presented as an interpretability aid complementary to the regression model and should not be interpreted as causal effects. Abbreviations: SHAP, Shapley additive explanations; UTI, Urinary tract infection; ICU, Intensive care unit Discussion In this retrospective cohort study of older adults undergoing inpatient surgery at a tertiary hospital in Southwest China (2019–2024), we developed and internally validated a clinically interpretable dynamic postoperative risk estimation model for postoperative UTI. Using routinely available perioperative data and a transparent modeling pipeline (LASSO feature selection followed by multivariable logistic regression), the final model retained six predictors—diabetes, ICU admission, catheter duration, recatheterization, length of hospital stay, and 30-day readmission—and demonstrated good discrimination in both the training and validation cohorts (AUC 0.886 and 0.830, respectively). Importantly, this model is best interpreted as a dynamic in-hospital postoperative risk stratification tool rather than a strictly preoperative or immediate postoperative bedside prediction model. Calibration curves suggested overall agreement between predicted and observed risks, and decision curve analysis indicated potential net benefit across clinically relevant thresholds. SHAP-based interpretability analyses were concordant with regression findings and highlighted hospitalization intensity and urinary management factors as dominant contributors to predicted risk, supporting the model’s face validity and potential utility in perioperative quality improvement. Nevertheless, these findings should be interpreted with appropriate caution because the derivation cohort contained a modest number of outcome events, which may limit coefficient stability and inflate apparent performance. Although penalized variable selection was used, this does not fully eliminate the risk of overfitting when the event count is limited relative to the dimensionality of the initial candidate predictor set. Interpretation of key predictors and comparison with prior evidence Our findings reinforce the central role of urinary catheter exposure in postoperative UTI risk. Both catheter duration and recatheterization emerged as strong predictors in the adjusted model. This is consistent with guideline statements emphasizing that indwelling catheterization is a major modifiable driver of catheter-associated UTI and that minimizing catheter use and duration is among the most effective preventive strategies [13–15]. Empirical perioperative data similarly support that longer catheter dwell time increases postoperative UTI risk across surgical settings [16, 17], and procedure-specific syntheses (e.g., orthopedic arthroplasty) suggest indwelling catheterization increases UTI events compared with avoidance or early removal strategies [18]. Clinically, these results underscore that catheter management is not merely a documentation variable but a high-leverage target for prevention—particularly in older adults, where bladder dysfunction, limited mobility, and perioperative delirium risk can complicate urinary management decisions. Diabetes was another robust predictor (adjusted OR > 5), aligning with longstanding evidence that hyperglycemia and diabetes confer susceptibility to infections, including postoperative and catheter-associated UTIs [19, 20]. In older adults, diabetes may represent both impaired host defenses and a marker of broader vascular and functional vulnerability that increases exposure to invasive monitoring and extended hospitalization. Chronic kidney disease showed an elevated univariable association but did not retain independent significance after adjustment, which may reflect correlation with other markers of illness burden (e.g., length of stay, ICU admission) and reduced statistical power once collinearity is addressed by penalized selection. Length of hospital stay was the most influential feature in SHAP analyses and remained independently associated with UTI in the multivariable model. This association is biologically and operationally plausible: longer hospitalization increases exposure time to catheterization, staff contacts, and healthcare-associated pathogens [21, 22]. However, it also introduces an important interpretive nuance: postoperative UTI can itself prolong hospitalization, creating a bidirectional relationship and potential temporal ambiguity. Thus, while length of stay is valuable for risk prediction—particularly as a time-updated indicator of evolving postoperative course—it should not be interpreted as a causal risk factor in the same way as preoperative comorbidities or catheter dwell time. ICU admission was inversely associated with postoperative UTI after adjustment, although this finding should be interpreted cautiously. Similar inverse associations have been reported in certain observational contexts when ICU admission functions as a proxy for higher-intensity monitoring, stricter catheter bundles, and earlier device removal protocols, or when confounding by indication, selection factors, competing risks, or different urine testing patterns alter the observed association [23, 24]. In our setting, this finding is most appropriately interpreted as a context marker of care pathways rather than a biologically protective exposure. If ICU admission in this setting captured differences in catheter stewardship or infection-prevention practices, these care-process differences may partly explain the observed association; however, this interpretation remains speculative and requires prospective evaluation. The inclusion of 30-day readmission as a retained predictor warrants particularly cautious interpretation. Because readmission occurs after discharge and may itself reflect delayed recognition or consequences of postoperative infection, it should not be interpreted as a true antecedent predictor for early bedside risk assessment. Rather, in the present model, readmission is more appropriately viewed as a downstream indicator of the postoperative course that may improve retrospective risk estimation but limits immediate clinical deployability at the time of admission or surgery. For this reason, the present model should be interpreted primarily as a tool for dynamic in-hospital postoperative reassessment rather than for admission-time bedside decision-making. This consideration further supports our interpretation of the current model as a dynamic postoperative risk estimation framework. Future studies should develop and externally validate simplified early-use models that exclude temporally downstream variables such as readmission. Model performance and clinical implications The model’s discrimination was strong in the training cohort and remained acceptable in the validation cohort, suggesting acceptable internal validity within the same institution, but not necessarily transportability or generalizability to other clinical settings. In the context of prediction modeling, discrimination alone is insufficient for clinical adoption; calibration and decision-analytic performance are equally important. The favorable calibration patterns and decision curve results suggest that the model could support threshold-based actions rather than indiscriminate interventions. Importantly, however, because some retained predictors are only available during the evolving postoperative course, the model is best viewed as a dynamic in-hospital postoperative risk stratification tool rather than an admission-time bedside calculator. In practical terms, the model may be used to support repeated postoperative reassessment rather than automated intervention and to identify patients who may warrant a more structured review pathway, including closer reassessment of catheter necessity, intensified surveillance for urinary or systemic symptoms, earlier urine testing when clinically indicated, physician or multidisciplinary reassessment, and reinforcement of modifiable preventive measures such as glycemic review and timely catheter removal where feasible. To translate the DCA findings into a pragmatic clinical framework, we now propose preliminary illustrative risk categories for potential future implementation, in which patients with predicted risk < 10% may be considered lower risk, those with predicted risk 10% to < 20% may be considered intermediate risk, and those with predicted risk ≥ 20% may be considered higher risk. Under such a framework, lower-risk patients may continue standard postoperative management. Patients in the intermediate-risk range may undergo more frequent nursing review of catheter necessity, closer monitoring for new urinary symptoms or unexplained fever/delirium, and a lower threshold for urine testing or physician reassessment when clinically appropriate. Patients in the higher-risk range may additionally warrant earlier multidisciplinary review of modifiable risk factors, including catheter removal feasibility, glycemic review, and intensified postoperative surveillance, while still avoiding indiscriminate testing or treatment in the absence of clinical indication. These proposed categories are intended only to illustrate potential workflow integration rather than to represent prospectively validated decision cutoffs or treatment thresholds, and they should therefore be interpreted cautiously pending external and prospective validation. Under this pragmatic framework, a higher predicted risk would not mandate treatment by itself, but would trigger a more deliberate postoperative review process focused on catheter stewardship, symptom surveillance, clinically indicated testing, and reassessment of modifiable contributors to infection risk. Our use of SHAP provides an additional interpretability layer, which can enhance clinician trust and facilitate discussion of which variables contributed most to a patient’s predicted risk within the fitted model, consistent with modern expectations for transparent prediction model reporting and use [25–27]. However, these SHAP-based explanations are complementary to the regression model and should not be interpreted as establishing causal relationships or underlying mechanisms. Strengths This study has several strengths. First, it focuses on an understudied but high-risk population—older surgical patients—whose infection risk reflects both baseline vulnerability and perioperative care processes. Second, predictors were derived from routinely collected variables, supporting feasibility in real-world perioperative workflows without requiring specialized biomarkers. Third, we used a transparent internal validation strategy combining split-sample evaluation with bootstrap-based optimism correction and assessed discrimination, calibration, and decision-analytic utility, aligning with reporting expectations for prediction model development and validation studies. Finally, by incorporating SHAP-based interpretability, we provide clinically intuitive explanations that align with catheter-associated infection prevention frameworks while maintaining model transparency. Limitations and considerations likely to be scrutinized Several limitations of this study merit careful consideration, as they may influence model interpretability, generalizability, and future clinical implementation. First, this was a retrospective, single-center study conducted at a tertiary hospital in Southwest China; therefore, local practice patterns—including catheter insertion and removal protocols, antimicrobial prophylaxis strategies, urine culture thresholds, and infection-control bundles—may differ substantially across institutions, potentially limiting model transportability. In addition, the use of a split-sample internal validation strategy may be statistically inefficient and may overestimate model performance, as it does not fully account for between-center heterogeneity. Accordingly, the current model should be interpreted as an internally validated model rather than a generalizable prediction tool. External validation and, where necessary, recalibration in independent cohorts are essential before clinical implementation. Second, postoperative UTI was defined according to routine clinical diagnosis supported by symptoms, laboratory findings, physician documentation, and treatment initiation, which reflects real-world practice but does not fully correspond to a uniform surveillance definition such as CDC/NHSN criteria. This may have introduced heterogeneity in outcome ascertainment, particularly because older adults often present atypically and urine culture practices may vary across clinicians and care settings. In addition, microbiological confirmation was not uniformly available for every case, and some degree of outcome misclassification is therefore possible. Such misclassification may have influenced regression coefficients and model performance, potentially reducing the validity, reproducibility, and transportability of the model. Future studies should externally validate the model using standardized, guideline-aligned diagnostic definitions and, where feasible, microbiologically confirmed endpoints. Third, several influential predictors—most notably length of hospital stay, catheter duration, and readmission within 30 days—are postoperative or time-dependent variables shaped by the evolving clinical course and care processes. Their inclusion introduces temporal ambiguity, raises the possibility of reverse causation and data leakage, and limits interpretation of the model as a strictly preoperative or early postoperative prediction tool, because postoperative UTI itself may influence hospitalization duration and subsequent healthcare utilization. Accordingly, the present model should be interpreted as a dynamic postoperative risk estimation model rather than a purely antecedent bedside prediction tool. Future studies should prioritize the development of simplified early prediction models restricted to variables available before or at the time of surgery. Fourth, although penalized regression and bootstrap-based internal validation were used to mitigate overfitting, the number of outcome events in the derivation cohort was modest (48 events for six retained predictors), corresponding to a borderline effective events-per-variable ratio. This may have contributed to model instability, overfitting risk, and the wide confidence intervals observed for some predictors, particularly recatheterization and 30-day readmission. Larger datasets with more outcome events will be important to improve coefficient stability, support model updating, and enable more reliable external validation. It is also possible that a more parsimonious model, based on fewer clinically stable predictors, could achieve similar predictive performance with improved interpretability and robustness; however, this would require formal comparison in larger development and validation datasets. Fifth, residual confounding cannot be excluded, as several potentially relevant determinants—including perioperative glycemic control, urinary tract obstruction, catheter insertion technique, catheter material, staffing factors, adherence to catheter bundles, perioperative delirium, mobility, and detailed antimicrobial timing—were not available in the electronic medical record and may partially explain observed associations. This limitation is particularly relevant to the inverse association observed for ICU admission, which may reflect unmeasured differences in care pathways, surveillance intensity, catheter practices, or indication-related selection rather than a true protective effect. Sixth, missing data were handled using complete-case analysis, which is reasonable when missingness is limited but may introduce bias if data are not missing at random; future studies should report missingness patterns in detail and consider multiple imputation with sensitivity analyses. Finally, model validation relied on a split-sample internal validation strategy, which may be less statistically efficient than resampling-based approaches; therefore, external validation across independent institutions and healthcare systems remains essential before clinical deployment. In addition, although we propose a preliminary pragmatic risk-stratified framework to illustrate possible workflow integration, the corresponding risk thresholds have not been prospectively validated and should not yet be used as definitive clinical decision cutoffs. Future directions Future research should (i) externally validate and recalibrate the model across multiple centers and healthcare systems with different clinical practices and patient populations; (ii) assess model transportability across settings with varying catheter management protocols, antimicrobial strategies, and diagnostic criteria for UTI; (iii) prospectively evaluate clinically actionable risk thresholds and determine whether risk-stratified postoperative management pathways improve patient outcomes; (iv) develop and validate simplified early prediction variants for bedside use, restricted to variables available before or at the time of surgery and excluding temporally downstream postoperative indicators such as readmission and length of stay; (v) perform model updating, revision, and comparison with more parsimonious alternative models using larger multicenter datasets to improve robustness, interpretability, and generalizability; (vi) test whether model-guided catheter stewardship interventions reduce UTI rates without increasing urinary retention or other adverse outcomes; and (vii) incorporate standardized, guideline-aligned infection definitions and microbiologically confirmed endpoints, together with microbiology/pathogen profiles where available, to strengthen outcome validity, reproducibility, and support tailored antimicrobial stewardship. Prospective impact studies are ultimately required to confirm that the model improves decision-making and patient outcomes beyond conventional care pathways. Conclusions Using routinely collected perioperative and postoperative variables, we developed and internally validated an interpretable dynamic postoperative risk estimation model for postoperative UTI in older surgical patients. The model showed good discrimination, acceptable calibration, and potential clinical utility, while highlighting catheter-related exposures and hospitalization intensity as major drivers of risk. Given the retrospective single-center design and the inclusion of postoperative course indicators, the model should not be interpreted as a strictly preoperative bedside decision support tool. Rather, it may serve as a preliminary framework for dynamic in-hospital postoperative risk stratification and for triggering pragmatic review pathways focused on catheter management, intensified surveillance, and reassessment of modifiable postoperative risk factors. External validation and, if needed, model recalibration in independent cohorts, together with prospective evaluation of clinically actionable thresholds and development of simplified early-use models, are essential before broader clinical implementation. Acknowledgements Not applicable. Authors’ contributions Xingyu Sun and Jiming Chen contributed equally to this work and share first authorship. Xingyu Sun and Jiming Chen were responsible for study conception and design, data acquisition, statistical analysis, and drafting of the initial manuscript. Lijuan He participated in data collection, data curation, and interpretation of the results. Shaohua Wang provided critical input on pathological considerations and contributed to the interpretation of clinical findings. Dan Li contributed to study supervision, methodological guidance, and critical revision of the manuscript for important intellectual content. All authors reviewed and approved the final manuscript and agreed to be accountable for all aspects of the work. Funding This work was supported by the Natural Science Foundation of Southwest Medical University (Grant No. 00031786). Data availability The datasets analyzed during the current study are available from the corresponding author on reasonable request. Declarations Ethics approval and consent to participate This study was reviewed and approved by the Ethics Committee of The Affiliated Traditional Chinese Medicine Hospital, Southwest Medical University (approval No.BY2026019). Given the retrospective cohort design and the use of existing electronic medical records, the requirement for written informed consent was waived by the ethics committee. All patient data were anonymized prior to analysis, and no identifiable personal information was accessed or reported. The study was conducted in accordance with the Declaration of Helsinki and relevant institutional and national ethical guidelines for research involving human participants. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Xingyu Sun and Jiming Chen contributed equally to this work and share first authorship Shaohua Wang and Dan Li share corresponding authorship. Contributor Information Shaohua Wang, Email: wangshaohua20006@163.com. Dan Li, Email: Lidan0918@163.com. References Associated Data Data Availability Statement The datasets analyzed during the current study are available from the corresponding author on reasonable request.