This retrospective cohort study was conducted at Fuding Hospital, Fujian University of Traditional Chinese Medicine. We consecutively screened hospitalized patients with a diagnosis of UTI and available urine culture results between January and December 2023. For patients with repeated admissions during the study period, only the first eligible admission was retained to avoid within-patient correlation. Records missing the primary outcome (urine culture–based pathogen group) were excluded; missing values in predictors were handled as described in the Data preprocessing section (imputation within the modeling pipeline). The overall study design and workflow are summarized in Figure 1.

Before analysis, the electronic medical record extract was pseudonymized (de-identified) by the hospital information team. All direct personal identifiers (e.g., name, national ID number, phone number, and address) were removed and each record was assigned a unique study code. The code key linking study codes to patient identities was stored separately by the hospital and was not accessible to the research team. Therefore, the analytic dataset contained no direct identifiers and included only variables required for this study.

The study protocol was approved by the Ethics Committee of Fuding Hospital, Fujian University of Traditional Chinese Medicine (Approval No. 2024015). All data were de-identified (anonymized) by the hospital information team prior to analysis, and the linkage key was not accessible to the research team, ensuring compliance with ethical and privacy standards. Due to the retrospective nature of the study, the requirement for written informed consent was waived by the ethics committee. The study was conducted in accordance with applicable local regulations and the Declaration of Helsinki.

Participants were classified into two groups based on urine culture results (midstream clean-catch specimens; catheter specimens were handled according to routine clinical practice, if applicable):

Non-E. coli group (Group 0, n = 150): Patients infected with uropathogens other than E. coli, including Klebsiella pneumoniae, Acinetobacter baumannii, Enterococcus spp., and other bacterial species.

E. coli group (Group 1, n = 158): Patients infected with E. coli identified as the predominant uropathogen.

Inclusion criteria comprised a positive urine culture with bacterial colony counts meeting standard clinical microbiology thresholds: ≥10⁵ colony-forming units (CFU)/mL, or 10⁴–10⁵ CFU/mL accompanied by clinical signs and symptoms consistent with UTIs. Exclusion criteria included: (1) immunocompromised status (e.g., human immunodeficiency virus (HIV) infection or long-term immunosuppressive therapy), (2) severe hepatic or renal failure, (3) disseminated malignancy, (4) concomitant infections outside the urinary tract, or (5) recent antibiotic use within 2 weeks before hospital admission to reduce culture-negative misclassification and biomarker distortion.

Demographic characteristics (age, sex) and laboratory test results were extracted from the electronic medical records. All laboratory analyses were performed in accordance with standard operating procedures for clinical diagnostics and were those obtained at/near the time of urine culture collection (index encounter).

Peripheral venous blood samples were collected under sterile conditions according to routine clinical practice as follows:

2.0 mL in EDTA tubes for complete blood count (CBC);

2.0 mL in citrate tubes for coagulation testing (e.g., D-dimer);

5.0 mL in plain tubes for serum biochemical analyses.

Urine specimens for urinalysis and microbial culture were collected using standard clinical procedures. Midstream clean-catch urine was preferred; when clean-catch was not feasible (e.g., catheterized patients), urine was collected via catheter specimen according to hospital protocol. Samples were obtained in sterile containers and transported to the microbiology laboratory promptly for analysis.

Urine cultures were performed by inoculating 1 μL of well-mixed urine onto 5% sheep blood agar plates using a calibrated loop. Samples were streaked in a standardized pattern and incubated at 35–37°C under aerobic conditions for 18–24 hours, with extended incubation up to 48 hours when needed. Where appropriate, MacConkey agar was used in parallel for differentiation of Gram-negative bacilli. Significant growth was defined as ≥10⁵ CFU/mL, or 10⁴–10⁵ CFU/mL in symptomatic patients. Urine specimens were primarily midstream clean-catch; catheter specimens were processed according to the same laboratory protocol when applicable.

Preliminary identification was based on colony morphology and Gram staining. Definitive species-level identification was conducted using the VITEK MS mass spectrometry system (bioMérieux, France), following the manufacturer’s protocols. Internal quality control was maintained using standard reference strains, such as Escherichia coli ATCC 25922.

Random forest modeling was performed using Python 3.7. The computational environment incorporated the following libraries: NumPy (v1.21.0) for numerical operations, Pandas (v1.3.0) for data manipulation, Scikit-learn (v0.24.2) for machine learning implementation and performance evaluation, and Matplotlib (v3.4.2) for visualization.

All statistical analyses were performed using SPSS software (version 22.0, IBM Corp., Armonk, NY, USA). The distribution of continuous variables was assessed using the Shapiro–Wilk test. Variables conforming to normal distribution were expressed as mean ± standard deviation (SD). Non-normally distributed variables were reported as median with interquartile range (IQR) and analyzed using the Mann–Whitney U test. Categorical variables were compared using the chi-square (χ²) test.

To identify independent predictors of E. coli infection, univariate logistic regression was first conducted. Variables with a p-value < 0.20 were subsequently entered into a multivariable logistic regression model using a backward stepwise elimination strategy. Adjusted odds ratios (ORs) and corresponding 95% confidence intervals (CIs) were reported. All statistical tests were two-tailed, and a P value < 0.05 was considered statistically significant.

A RF classifier (scikit-learn) was trained to distinguish between E. coli (Group = 1) and non–E. coli (Group = 0) UTIs. Missing predictor values were imputed using a median strategy within a pipeline. Imputation parameters were estimated from the training set only, and then applied to the held-out test set to prevent information leakage. The dataset was split once using a stratified 70/30 train-test split (random_state = 42).

Hyperparameter tuning was conducted on the training set using RandomizedSearchCV with 5-fold stratified cross-validation, optimizing ROC-AUC. The candidate search space included: n_estimators ∈ {300, 500, 800, 1200}, max_depth ∈ {None, 3, 5, 7, 10, 15}, min_samples_split ∈ {2, 5, 10, 20}, min_samples_leaf ∈ {1, 2, 3, 5, 8}, and max_features ∈ {“sqrt”, “log2”, 0.3, 0.5, 0.8}. To address potential class imbalance, the class_weight parameter was set to “balanced”. The final RF hyperparameters selected by RandomizedSearchCV were: n_estimators = 300, max_depth = 15, min_samples_split = 10, min_samples_leaf = 8, max_features = log2, with class_weight = “balanced” and random_state = 42.

To contextualize the RF model and reduce the risk of model-specific findings, we benchmarked RF against commonly used baseline classifiers trained on the same feature set and data split. These baselines included (i) regularized logistic regression (L2-penalized), (ii) support vector machine (SVM), and (iii) gradient boosting (tree-based boosting). For all models, missing predictors were imputed using the same median-imputation strategy within a training-only pipeline to prevent information leakage; continuous features were standardized for linear models (logistic regression/SVM). Hyperparameters were tuned using stratified cross-validation within the training set, and final performance was reported on the held-out stratified test set using the same discrimination (ROC-AUC, PR-AUC) and calibration metrics (Brier score; calibration intercept/slope).

To interpret the RF model, we used SHAP to quantify each feature’s contribution to the predicted probability of E. coli at the individual level. SHAP values were computed using TreeExplainer (TreeSHAP) following the framework described by Lundberg et al (Lundberg et al., 2020), which provides consistent additive attributions for tree-based models and can reflect nonlinear and interaction-driven effects. Global importance was summarized as the mean absolute SHAP value computed on the held-out test set, and signed SHAP summary (beeswarm) plots were used to visualize directionality and heterogeneity. SHAP analyses were conducted using the Python SHAP package (v0.45.1).

The Shapiro–Wilk test revealed that most continuous variables, including CRP, ALT, AST, PCT, and D-dimer, deviated from a normal distribution in both groups (all P < 0.05). Only HGB) and ALB in the non-E. coli group demonstrated normal distribution (P = 0.309 and 0.773, respectively). Accordingly, non-parametric tests (e.g., Mann–Whitney U test) were employed for intergroup comparisons of non-normally distributed variables.

A total of 308 bacterial isolates were identified from patients with UTIs. As shown in Figure 2, E. coli was the most prevalent pathogen, comprising 158 isolates (51.3%), followed by Enterococcus spp. (57, 18.5%) and Klebsiella spp. (24, 7.8%). Other detected organisms included non-fermenting Gram-negative bacilli (20, 6.5%), Streptococcus spp. (14, 4.5%), Staphylococcus spp. (2, 0.7%), and other members of the Enterobacteriaceae family (33, 10.7%). A detailed breakdown of bacterial species and their respective proportions is presented in Table 1.

A total of 308 patients were included: 150 in the non-E. coli group and 158 in the E. coli group. No significant difference was observed in median age between the groups (68.0 vs. 66.0 years, P = 0.301). However, gender distribution differed significantly, with a higher proportion of females in the E. coli group (74.3% vs. 44.7%, P < 0.001) (Table 2).

Compared to the non -E. coli group, the E. coli group exhibited significantly higher values of HGB (P = 0.004), LYM (P = 0.003), ALT (P = 0.001), and ALB (P = 0.021). In contrast, RDW (P = 0.016) and D-dimer levels (P = 0.014) were lower in the E. coli group (Table 3).

As shown in Table 4, univariate logistic regression analysis identified eight variables significantly associated with the presence of E. coli UTI, including HGB, RDW, LYM, ALT, ALB, BUN, sex, and urinary WBC (grade 1). These variables were subsequently entered into a multivariate logistic regression model to determine independent predictors (Table 5).

The multivariable analysis revealed three statistically significant independent predictors of E. coli UTIs: LYM (OR = 1.690, P = 0.016); sex (male) (OR = 0.299, P < 0.001); Urinary WBC (Grade 1) (OR = 3.100, P = 0.002) (Table 5).

These results indicate that females are more likely to develop E. coli-associated UTIs. Additionally, elevated LYM and the presence of urinary WBC (grade 1) significantly increased the risk of E. coli UTIs.

Pearson correlation heatmaps were constructed separately for the non-E. coli group (Group 0) and the E. coli group (Group 1) to examine inter-variable associations (Figures 3A, B).

In the non-E. coli group (Group 0), most variable pairs demonstrated weak correlations (|r| < 0.3). Notable moderate correlations were observed between CRP and neutrophils (r = 0.61), between HDL and CHO (r = 0.56), and RDW and HGB (r = 0.56).

In contrast, Group 1 exhibited more pronounced inter-variable relationships. A strong correlation was observed between PCT and AST (r = 0.62), while a moderate correlation was noted between PCT and GLU (r = 0.53).

These distinct patterns suggest that the systemic inflammatory and metabolic responses may differ between E. coli and non-E. coli UTIs.

The confusion matrix for the random forest model on the test set is shown in Figure 4. The model correctly identified 30 patients in the non-E. coli group and 29 in the E. coli group. Misclassifications included 15 false positives (non-E. coli cases incorrectly predicted as E. coli) and 19 false negatives. This distribution reflects a relatively balanced classification performance, though slight misclassification bias toward both classes was observed.

Because the final predictor dimensionality was limited (P = 19) relative to the sample size (n = 308), the model was not exposed to high-dimensional one-hot expansion. Model complexity was further controlled by cross-validated hyperparameter tuning within the training set and evaluated on a held-out test set.

On the held-out test set (n = 93), the RF model achieved an overall accuracy of 0.63. Class-wise precision/recall/F1 were 0.61/0.67/0.64 for the non–E. coli group and 0.66/0.60/0.63 for the E. coli group (Supplementary Table S2).

The model showed moderate discrimination with a ROC-AUC of 0.66 (95% CI 0.54–0.77) and a PR-AUC of 0.66 (95% CI 0.54–0.81). Calibration was acceptable, with a Brier score of 0.233 (95% CI 0.210–0.255) and a calibration slope of 1.130 (intercept 0.124) (Table 6).

At the default threshold of 0.50, the model achieved an accuracy of 0.65 (95% CI 0.55–0.74) with sensitivity 0.62 (0.49–0.76) and specificity 0.67 (0.52–0.80) (Table 7). Using the Youden threshold determined on the training set (0.476) yielded sensitivity 0.65 (0.52–0.79) and specificity 0.60 (0.44–0.74).

The diagnostic performance of the RF model was further evaluated using the ROC curve and the precision–recall (P–R) curve (Figures 5, 6). As shown in Figure 5, the ROC curve remained above the diagonal reference line across most thresholds, indicating discriminative ability beyond chance. The corresponding ROC-AUC, with the 95% confidence interval, is reported in Table 6. As shown in Figure 6, the P–R curve was generally above the baseline precision determined by the positive-class prevalence, supporting clinically meaningful precision–recall trade-offs. The corresponding PR-AUC (average precision), with 95% confidence interval, is summarized in Table 6.

The RF model’s performance was benchmarked against several standard baselines, including regularized logistic regression, SVM, and gradient boosting, using the same stratified held-out evaluation. The RF model demonstrated comparable performance to these baselines, with ROC-AUC values ranging from 0.59 to 0.66 and PR-AUC values from 0.61 to 0.66 (Supplementary Table S3). Notably, the discrimination remained moderate across all methods, justifying the decision to report conservative performance estimates derived from stratified evaluation rather than potentially optimistic non-stratified splits.

Further stability checks were conducted to compare the top features selected by each model. The overlap of the top-10 features, measured by the Jaccard index, and the rank correlation of feature importance were assessed to evaluate the consistency of the selected biomarkers across different models. Supplementary Figure S1 presents the Jaccard index of top-10 features across the models, while Supplementary Figure S2 shows the Spearman rank correlation between the importance rankings of these features.

The model showed acceptable calibration on the held-out test set (Brier score = 0.233; calibration intercept = 0.124; calibration slope = 1.130; Table 6). The calibration curve is shown in Figure 7.

Model interpretability was assessed using SHAP. Global SHAP ranking (mean |SHAP| on the test set) identified Sex, LYM, and ALT as the most influential predictors contributing to the model’s output for E. coli classification (Figure 8).

The SHAP beeswarm plot further illustrated both directionality and inter-individual heterogeneity (Figure 9). Positive SHAP values indicate an increased predicted probability of E. coli (Group = 1), whereas negative values indicate a decreased probability. For Sex (0 = female, 1 = male), higher values (male) were predominantly associated with negative SHAP values, while lower values (female) were more often associated with positive SHAP values, suggesting that female sex increased and male sex decreased the model-predicted probability of E. coli in this cohort. Wider horizontal dispersion for certain variables indicates greater heterogeneity in effects and possible interaction patterns at the patient level.

This study is limited by its single-center, retrospective design, and external validation is needed to confirm generalizability across diverse patient populations. Although multiple models (Random Forest, logistic regression, SVM, and gradient boosting) were compared, the findings should be interpreted within the context of the dataset and applied classifiers. Future research should assess the stability of biomarkers across different algorithms and evaluate model performance in external cohorts. Additionally, prospective validation and temporal testing are needed to assess the model’s generalizability and clinical applicability over time.