Explainable machine learning with routine biomarkers identifies culture-defined bacteremic urosepsis.

Skip to main content An official website of the United States government Here's how you know Here's how you know Official websites use .gov A .gov website belongs to an official government organization in the United States. Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites. Search Log in Dashboard Publications Account settings Log out Search… Search NCBI Primary site navigation Search Logged in as: Dashboard Publications Account settings Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide PERMALINK Copy As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Sci Rep . 2026 Mar 4;16:11982. doi: 10.1038/s41598-026-42178-8 Search in PMC Search in PubMed View in NLM Catalog Add to search Explainable machine learning with routine biomarkers identifies culture-defined bacteremic urosepsis Yuan-Lu Zhang Yuan-Lu Zhang 1 Department of Clinical Laboratory, Fuding Hospital, Fujian University of Traditional Chinese Medicine, Fuding, 355200 Fujian Province China Find articles by Yuan-Lu Zhang 1, # , Dong-Xiao Yu Dong-Xiao Yu 2 Department of Pediatrics, Fuding Hospital, Fujian University of Traditional Chinese Medicine, Fuding, 355200 Fujian Province China Find articles by Dong-Xiao Yu 2, # , Ying-Ying Zheng Ying-Ying Zheng 1 Department of Clinical Laboratory, Fuding Hospital, Fujian University of Traditional Chinese Medicine, Fuding, 355200 Fujian Province China Find articles by Ying-Ying Zheng 1 , Jie Zhang Jie Zhang 1 Department of Clinical Laboratory, Fuding Hospital, Fujian University of Traditional Chinese Medicine, Fuding, 355200 Fujian Province China Find articles by Jie Zhang 1 , Shi-Yan Zhang Shi-Yan Zhang 1 Department of Clinical Laboratory, Fuding Hospital, Fujian University of Traditional Chinese Medicine, Fuding, 355200 Fujian Province China Find articles by Shi-Yan Zhang 1, ✉ , Jinbao Shi Jinbao Shi 3 Department of Nephrology, Fuding Hospital, Fujian University of Traditional Chinese Medicine, Fuding, 355200 Fujian Province China 4 Department of Nephrology, Ningde Hospital of Traditional Chinese Medicine, Ningde, 352100 Fujian Province China Find articles by Jinbao Shi 3, 4, ✉ Author information Article notes Copyright and License information 1 Department of Clinical Laboratory, Fuding Hospital, Fujian University of Traditional Chinese Medicine, Fuding, 355200 Fujian Province China 2 Department of Pediatrics, Fuding Hospital, Fujian University of Traditional Chinese Medicine, Fuding, 355200 Fujian Province China 3 Department of Nephrology, Fuding Hospital, Fujian University of Traditional Chinese Medicine, Fuding, 355200 Fujian Province China 4 Department of Nephrology, Ningde Hospital of Traditional Chinese Medicine, Ningde, 352100 Fujian Province China ✉ Corresponding author. # Contributed equally. Received 2025 Dec 12; Accepted 2026 Feb 24; Collection date 2026. © The Author(s) 2026 Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/ . PMC Copyright notice PMCID: PMC13068939  PMID: 41781502 Abstract Urosepsis is a severe complication of urinary tract infection (UTI) and may lead to organ dysfunction and death. Early identification remains challenging at initial presentation, highlighting the need for improved risk stratification using routinely available data. This single-center retrospective study analyzed clinical data from 182 hospitalized patients with culture-confirmed UTI, including 89 with culture-defined bacteremic urosepsis (concurrent positive blood and urine cultures) and 93 with non-bacteremic UTI. Random Forest (RF), Extreme Gradient Boosting (XGBoost), and multivariable logistic regression (LR) models were developed using routine biomarkers obtained within 0–24 h of the index time; outcomes were assigned using culture results within 48–72 h to minimize information leakage. Discrimination was assessed using the area under the receiver operating characteristic curve (AUC) with bootstrap 95% confidence interval (CI) on a held-out test set. D-dimer was consistently ranked among the top predictors. Compared with non-bacteremic UTI, bacteremic urosepsis showed higher procalcitonin (PCT), C-reactive protein (CRP), and white blood cell count (WBC) and lower albumin (all p < 0.05). On the held-out test set ( n = 37; positives = 18), XGBoost achieved an AUC of 0.886 (95% CI 0.763–0.971), compared with 0.822 (95% CI 0.665–0.938) for RF and 0.822 (95% CI 0.663–0.935) for LR; the AUC difference between XGBoost and RF was not statistically significant (DeLong p = 0.072). Using routine biomarkers available within 24 h, RF and XGBoost demonstrated good discrimination for culture-defined bacteremic urosepsis among inpatients with culture-confirmed UTI. XGBoost yielded a numerically higher AUC than RF, but the difference was not statistically significant in this modest test set. D-dimer, procalcitonin, and albumin emerged as key predictors, supporting the potential utility of routine laboratory indicators for early risk stratification pending external validation. Supplementary Information The online version contains supplementary material available at 10.1038/s41598-026-42178-8. Keywords: Urosepsis, Urinary tract infection, Biomarkers, Machine learning, Random Forest, XGBoost Subject terms: Biomarkers, Diseases, Medical research Introduction Urinary tract infection (UTI) is among the most common bacterial infections worldwide 1 , 2 . A serious subset progresses to urosepsis, a dysregulated host response to a urinary-source infection that can precipitate organ dysfunction and death 3 . Early identification is critical, yet distinguishing at presentation which patients will evolve from localized UTI to systemic infection remains challenging: clinical signs overlap, blood cultures require time to turn positive 4 , and bedside screens such as the quick Sequential Organ Failure Assessment (qSOFA) are designed for prognostic risk stratification rather than definitive diagnosis 5 . Routinely available biomarkers offer complementary information. Procalcitonin (PCT) and C-reactive protein (CRP) index systemic inflammation 6 , 7 . D-dimer reflects coagulation activation and microthrombotic activity 8 . Serum albumin captures metabolic/nutritional reserve and capillary leak 9 . Considered jointly, these biomarkers likely exhibit nonlinear interactions and context-dependent patterns better captured by machine-learning (ML) models 10 . ML methods can integrate heterogeneous routine data and model nonlinearities 11 . Among tree-based approaches, Random Forest (RF) provides robust performance through bagging and feature subsampling, whereas Extreme Gradient Boosting (XGBoost) often achieves superior discrimination on tabular clinical data 12 . However, comparative evaluations that use only early routine laboratories, are time-anchored to the presentation window, and are grounded in paired urine–blood cultures remain limited 13 . In particular, the relative contributions of inflammation (PCT/CRP, leukocyte indices), coagulation activation (D-dimer), and metabolic reserve (albumin) to early urosepsis identification have not been fully clarified, despite recent biomarker reviews/meta-analyses calling for integrative modeling 14 . To address these gaps, we assembled a consecutive, single-center cohort of inpatients with culture-confirmed UTI, each with paired urine and blood cultures and routine laboratory tests obtained within 0–24 h of the index time. For the primary analysis, we operationalized bacteremic urosepsis as concurrent blood- and urine-culture positivity, and non-bacteremic UTI as urine-culture positive with negative blood cultures. This microbiology-based endpoint was selected to improve specificity for urinary-source bloodstream dissemination in a retrospective setting. Importantly, bacteremia is not synonymous with Sepsis-3; therefore, our outcome reflects culture-defined bacteremic urosepsis rather than all Sepsis-3 sepsis, and culture-negative sepsis is not captured. We developed RF- and XGBoost-based diagnostic models using only routine hematology, coagulation, and chemistry variables, and compared performance by discrimination and clinically relevant classification metrics. We aimed to compare XGBoost and RF for early identification of culture-defined bacteremic urosepsis using routinely available biomarkers, and to characterize key predictors supporting model interpretability and clinical translation. By focusing on readily obtainable tests and head-to-head model comparisons, this work aims to outline an implementable pathway for early risk stratification and to inform prioritization of cultures, monitoring, and antimicrobial decisions. Materials and methods Study design and data source We conducted a retrospective cohort study at Fuding Hospital, Fujian University of Traditional Chinese Medicine. Consecutive inpatients who underwent both blood and urine cultures for a suspected urinary-source infection between May 2022 and April 2024 were screened. Clinical and laboratory data were extracted from the electronic medical record (EMR). Group definitions Bacteremic urosepsis group (n = 89) : patients with concurrent positive blood and urine cultures obtained at presentation (index time) or within the subsequent 48–72 h. Concordance was defined as isolation of the same species from blood and urine cultures. Cases with discordant organisms, polymicrobial cultures, or organisms adjudicated as contaminants according to institutional SOPs were excluded. Non-bacteremic UTI group (n = 93) : patients with a positive urine culture and all blood cultures negative within the same diagnostic window. We used the term “bacteremic urosepsis” to emphasize that the endpoint is microbiology-defined and does not equate to Sepsis-3. A formal Sepsis-3 sensitivity analysis was not performed because complete SOFA component data were unavailable for all patients. Microbiological criteria Urine culture positivity followed laboratory standards: colony count ≥ 10 5 colony-forming units (CFU)/mL for midstream urine, or 10 4 −10 5 CFU/mL with compatible urinary symptoms and/or pyuria, according to institutional microbiology criteria. Blood-culture positivity required growth of a recognized pathogen in at least one blood-culture set, with potential contaminants adjudicated using prespecified criteria. Organisms deemed contaminants according to institutional standard operating procedures (SOPs) (e.g., single-bottle skin commensals without supportive clinical features or repeat isolation) were excluded. Exclusion criteria Patients were excluded if they had severe immunodeficiency (HIV infection or long-term immunosuppressive therapy), end-stage renal disease requiring dialysis or decompensated cirrhosis, active metastatic malignancy, a documented alternative primary source of infection at presentation (e.g., pneumonia or intra-abdominal infection), or systemic antimicrobial exposure within 14 days prior to index time. Index time and window The index time (t₀) was defined as the earliest of blood- or urine-culture order time, and microbiological results within 48–72 h were used to assign group membership. Predictor laboratories were restricted to 0–24 h after t₀ and were obtained prior to availability of culture results, to minimize information leakage. Data collection and variables Clinical variables (age, sex, smoking, alcohol use, and chronic comorbidities) and routine laboratory indices (hematology, coagulation, serum biochemistry, and urinalysis) were retrieved from the electronic medical record. To reflect early identification at presentation, candidate predictors were restricted to test results obtained within 0–24 h of the index time. All measurements followed institutional SOPs for clinical laboratories. Comorbidities were ascertained from documented diagnoses in the medical record at admission. For patients with multiple results within the 0–24 h window, the earliest value was used for model development. Blood and urine specimen collection Blood samples. Peripheral venous blood was collected according to standard phlebotomy procedures into ethylenediaminetetraacetic acid (EDTA) tubes for complete blood counts, sodium-citrate tubes for coagulation assays (including D-dimer), and serum-separator/plain tubes for biochemistry testing. Blood volumes and tube types followed institutional laboratory SOPs. In parallel, 8–10 mL of blood was obtained for blood cultures. Urine samples. Midstream clean-catch urine was collected aseptically as soon as feasible after presentation for urine culture and routine urinalysis. When midstream collection was not feasible (e.g., catheterized patients), urine was obtained per clinical routine and documented accordingly. Bacterial culture and identification Blood culture : Aerobic and anaerobic blood cultures were performed using the BacT/ALERT 3D automated system (bioMérieux, Marcy-l’Étoile, France). Growth in any bottle was considered positive, subject to laboratory adjudication of contaminants per SOPs. Potential contaminants (e.g., single-bottle skin commensals without supportive clinical evidence or repeat isolation) were classified according to prespecified laboratory criteria. Urine culture : Urine was inoculated using a calibrated 1-µL loop and incubated at 35 °C for 24–48 h under appropriate atmospheric conditions. Colony counts were interpreted using institutional criteria (e.g., ≥ 10⁵ CFU/mL for midstream urine, or ≥ 10⁴ CFU/mL with compatible urinary symptoms and/or pyuria). Pathogen identification : Isolates were identified using the VITEK MS automated matrix-assisted laser desorption/ionization time-of-flight (MALDI-TOF) mass spectrometry system (bioMérieux, Marcy-l’Étoile, France). Where required by routine practice, identification was supported by standard biochemical methods. Biomarker assays Inflammatory/infection markers. Procalcitonin (PCT, ng/mL) was measured by electrochemiluminescence (MCL60, Renmai, Nanjing, China). C-reactive protein (CRP, mg/L) was quantified by latex-enhanced immunoturbidimetry (BC-7500CS, Mindray, Shenzhen, China). Hematology and coagulation. Complete blood count parameters—including white blood cells (WBC, ×10 9 /L), absolute neutrophil count (×10 9 /L), absolute lymphocyte count (×10 9 /L), absolute eosinophil count (×10 9 /L), hemoglobin (HGB, g/L), red cell distribution width (RDW, %), and platelets (PLT, ×10 9 /L)—were obtained on an automated hematology analyzer (BC-7500CS, Mindray). Plasma D-dimer (mg/L, fibrinogen equivalent units (FEU)) was determined by immunoturbidimetry (ExC810, Mindray). When multiple measurements were available within 0–24 h, the earliest value was used. Serum biochemistry. Total protein (g/L), albumin (ALB, g/L), uric acid (UA, µmol/L), glucose (GLU, mmol/L), and creatinine (Cr, µmol/L) were measured on an automated chemistry analyzer (AU5800, Beckman Coulter, USA). Urinalysis. Urinary leukocyte esterase and nitrite were assessed using an automated urine analyzer (UF-500i, Sysmex, Japan). Quality control and reporting. All assays followed manufacturer instructions and institutional SOPs. Two-level internal quality controls were run daily; the laboratory participates in external quality-assessment programs. Results were reported with the units indicated above and used for modeling only if obtained within 0–24 h of the index time. Values were extracted directly from the laboratory information system without manual modification. Random Forest, XGBoost, and logistic regression modeling We developed multivariable models to discriminate bacteremic urosepsis from non-bacteremic UTI using routine biomarkers available within 0–24 h of the index time. Three algorithms were evaluated: logistic regression (LR) as a conventional baseline, RF, and XGBoost. All analyses were implemented in Python 3.7 using NumPy 1.21.0, pandas 1.3.0, scikit-learn 0.24.2, and Matplotlib 3.4.2, and XGBoost was implemented with the xgboost package. Computations were performed on a local CPU-only environment (no GPU). Analyses were conducted using a unified scikit-learn pipeline to ensure that preprocessing steps were fitted only on training data. Data preprocessing Missing data. Duplicate admissions were excluded. Records were excluded only if paired culture results (outcome ascertainment) were unavailable or if key predictor panels were missing beyond a prespecified threshold (e.g., > 30% of candidate laboratory variables). For the remaining cohort, sporadic missing predictor values were handled by within-pipeline imputation (median for continuous variables; most frequent for categorical variables) to preserve sample size and prevent information leakage. Encoding and scaling. Binary variables were encoded as 0/1. After imputation, continuous variables were standardized (z-scores) for LR only; tree-based models used unscaled values. Leakage control and time window. The index time (t₀) was defined as the earliest of blood- or urine-culture order time. Only predictors measured within 0–24 h after t₀ were used for model development and evaluation to avoid information leakage. Outcome labels were assigned using culture results within 48–72 h of t₀, and predictors were restricted to values available before culture reporting. Training, validation, and thresholding Data were randomly split into training (80%) and test (20%) sets with a fixed random seed for reproducibility. Within the training set, models were fitted using five-fold cross-validation, and performance was evaluated once on the held-out test set. With the fixed seed, the split yielded a training set of n = 145 (bacteremic urosepsis = 71) and a held-out test set of n = 37 (bacteremic urosepsis = 18). Baseline demographic, clinical, and laboratory characteristics were compared between the training and held-out test sets to assess potential sampling bias. The operating threshold for each model was determined by maximizing Youden’s J (sensitivity + specificity − 1) on the cross-validation validation folds within the training set and then fixed and applied to the held-out test set. Model discrimination was quantified by AUC, and 95% CI were estimated by bootstrap resampling of the held-out test set. AUC comparisons between models were performed using DeLong’s test. We reported threshold-dependent metrics (accuracy, precision, recall, specificity and F1 score) together with the confusion matrix, alongside threshold-independent AUC. Hyperparameters RF was implemented with bootstrap aggregation and feature subsampling at each split (n_estimators = 500; class_weight = “balanced_subsample”; min_samples_leaf = 2; random_state = 42). XGBoost was implemented as gradient-boosted decision trees with shrinkage and row/column subsampling (n_estimators = 600; max_depth = 3; learning_rate = 0.03; subsample = 0.90; colsample_bytree = 0.90; reg_lambda = 1.0; objective = “binary: logistic”; eval_metric = “logloss”; random_state = 42). LR was used as a conventional baseline with L2 regularization and balanced class weights (solver = “liblinear”; max_iter = 500; class_weight = “balanced”; random_state = 42). Hyperparameters were prespecified (no automated tuning) to reduce overfitting risk in a modest-sized dataset. Final models were trained on the full training set and evaluated once on the held-out test set. Model explainability (SHAP) To improve interpretability of the best-performing tree-based model, we conducted post hoc explainability analysis using SHAP (Shapley Additive Explanations). SHAP values were computed with the TreeExplainer for the XGBoost model and were used solely for model interpretation, without influencing model training, feature selection, threshold determination, or performance evaluation. To avoid information leakage, SHAP analyses were performed on the held-out test set predictions only and reported as explanatory outputs rather than as a model-selection criterion. SHAP values were calculated on the model output scale (log-odds) and represent additive contributions of each feature to the prediction relative to the model baseline. Global importance was summarized by mean absolute SHAP value (summary plot). For case-based explanation, we generated SHAP waterfall plots for one correctly classified non-bacteremic UTI case and one correctly classified bacteremic urosepsis case, selected with predicted probabilities close to the median of their respective groups. All SHAP visualizations were generated after final model fitting and were intended to facilitate clinical interpretation of how routine biomarkers jointly influenced individual risk estimates. Evaluation metrics Model performance assessment Confusion matrix We computed true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN) to summarize classification outcomes. From the confusion matrix, we additionally derived specificity = TN/(TN + FP) and accuracy = (TP + TN)/(TP + TN+FP + FN). Receiver operating characteristic (ROC) curve and AUC ROC curves depict the relationship between the false-positive rate (FPR) and true-positive rate (TPR) across decision thresholds. The AUC quantifies overall discriminative ability, with values closer to 1.0 indicating superior performance. AUC 95% confidence interval (CI) were estimated by bootstrap resampling of the held-out test set. AUC differences between models were compared using DeLong’s test. For contextual comparison, we also computed ROC curves and AUCs for selected individual biomarkers on the held-out test set using the same bootstrap procedure for 95% CIs. Precision, recall, and F1 score Precision = TP/(TP + FP). Recall (Sensitivity) = TP/(TP + FN). F1 score = 2 × (Precision × Recall)/(Precision + Recall). Unless otherwise specified, the operating threshold for point estimates was selected by Youden’s J on validation folds and applied to the test set. Threshold-dependent metrics were calculated on the held-out test set using the fixed operating threshold defined within the training cross-validation. Statistical analysis Data were analyzed using SPSS version 22.0 for descriptive statistics and between-group comparisons, whereas model development and performance evaluation were conducted in Python (as described above). Normality of continuous variables was assessed with the Shapiro–Wilk test. For normally distributed data, results are presented as mean ± standard deviation (Mean ± SD) and compared between groups using independent-samples t tests. For non-normally distributed data, results are summarized as median (interquartile range, IQR) and compared using the Mann–Whitney U test. Categorical variables were compared using the chi-square (χ²) test. To assess potential confounding by coronary artery disease (CAD), we conducted CAD-stratified analyses and fitted multivariable LR models including CAD status (ascertained from documented clinical diagnoses in the medical record), age and sex as prespecified covariates, and a D-dimer×CAD interaction term. Interaction analyses and multiple pairwise comparisons were considered exploratory and were interpreted cautiously. A two-sided p < 0.05 was considered statistically significant. Ethics statement This study was approved by the Medical Ethics Committee of Fuding Hospital, Fujian University of Traditional Chinese Medicine (Approval No. FDSYY-2024015). All patient data were de-identified, and procedures complied with institutional and national regulations as well as the Declaration of Helsinki. Owing to the retrospective design and the use of routinely collected clinical data with minimal risk, the requirement for written informed consent was waived by the Ethics Committee. No interventions were performed and patient care was not affected by this study. Results Shapiro-Wilk normality testing Shapiro–Wilk testing indicated that hemoglobin, total protein, and albumin were approximately normally distributed in both groups ( p > 0.05); these variables were compared using independent-samples t tests. All other continuous variables were non-normally distributed ( p < 0.05) and were compared using the Mann–Whitney U test, with results reported as median (IQR). Baseline characteristics of the bacteremic urosepsis and non-bacteremic UTI cohorts There were no significant between-group differences in sex, smoking, alcohol use, hypertension, malignancy, or diabetes (all p > 0.05) (Table 1 ). CAD was more frequent in the non-bacteremic UTI group than in the bacteremic urosepsis group (18.3% vs. 6.7%; p = 0.019). Median ages were 69.0 and 66.0 years, respectively, and did not differ significantly ( p = 0.118). Table 1. Baseline characteristics of the bacteremic urosepsis and non-bacteremic urinary tract infection cohorts [n (%), median (IQR)]. Variable Category Non-bacteremic UTI ( n = 93) Bacteremic urosepsis ( n = 89) χ 2 /Z p -value Sex Female 54 (58.1) 55 (61.8) 0.264 0.617 Male 39 (41.9) 34 (38.2) Smoking No 86 (92.5) 84 (94.4) 0.269 0.604 Yes 7 (7.5) 5 (5.6) Alcohol use No 84 (90.3) 84 (94.4) 1.055 0.304 Yes 9 (9.7) 5 (5.6) Hypertension No 38 (40.9) 36 (40.4) 0.003 0.955 Yes 55 (59.1) 53 (59.6) CAD No 76 (81.7) 83 (93.3) 5.484 0.019 Yes 17 (18.3) 6 (6.7) Malignancy No 80 (86.0) 77 (86.5) 0.009 0.923 Yes 13 (14.0) 12 (13.5) Diabetes No 63 (67.7) 56 (62.9) 0.467 0.494 Yes 30 (32.3) 33 (37.1) Age (years) Median (IQR) 69.0 (57.0–81.0) 66.0 (53.0–74.0) −1.561 0.118 Range 1.0–92.0 1.0–88.0 Open in a new tab Notes: Categorical variables were compared using the chi-square ( χ 2 ) test; continuous variables with non-normal distribution were compared using the Mann–Whitney U test (reported as Z). Malignancy, Non-metastatic malignancy history. Two-sided p < 0.05 was considered statistically significant. Abbreviations: UTI, urinary tract infection; IQR, interquartile range; CAD, coronary artery disease. Laboratory findings in the bacteremic urosepsis vs. Non-bacteremic UTI cohorts As summarized in Table 2 , patients with bacteremic urosepsis exhibited significantly higher levels of D-dimer, white blood cell count, neutrophils, C-reactive protein, procalcitonin, blood glucose, and serum creatinine than those with non-bacteremic UTI (all p < 0.05), which is consistent with greater systemic inflammation and physiological stress. In contrast, hemoglobin, total protein, and albumin were significantly lower in the bacteremic urosepsis group (all p < 0.05), suggesting that hypoalbuminemia may be associated with greater disease severity. Platelet and lymphocyte counts were also significantly lower in bacteremic urosepsis (both p < 0.05), whereas eosinophil counts differed between groups ( p < 0.001) (Table 2 ). Table 2. Comparison of laboratory indices between the non-bacteremic UTI and bacteremic urosepsis groups [median (IQR) or mean ± SD]. Variable Non-bacteremic UTI( n = 93) Bacteremic Urosepsis ( n = 89) t/Z p - value Hemoglobin (g/L) 110.61 ± 27.16 102.68 ± 22.79 2.128 0.035 D-dimer (mg/L) 2.35 (1.10–2.35) 2.83 (1.51–2.83) −4.361 < 0.001 Platelets (10 9 /L) 213.0 (154.0–266.0) 156.00 (115.5–232.0) −3.526 < 0.001 RDW (%) 13.8 (13.0–15.2) 13.7 (13.0–15.0) −0.590 0.555 Eosinophils (10 9 /L) 0.03 (0.01–0.13) 0.10 (0.00–0.20) −4.134 < 0.001 WBC (10 9 /L) 9.35 (6.36–13.15) 13.43 (8.18–16.88) −3.447 0.001 Neutrophils (10 9 /L) 7.45 (4.49–10.80) 12.09 (6.58–15.43) −3.970 < 0.001 Lymphocytes (10 9 /L) 1.01 (0.67–1.57) 0.76 (0.54–1.23) −2.777 0.005 CRP (mg/L) 67.73 (24.84–97.40) 118.41 (54.50–155.64) −4.316 < 0.001 Procalcitonin (ng/mL) 0.42 (0.10–3.99) 15.56 (2.04–30.76) −7.064 < 0.001 Glucose (mmol/L) 7.19 (5.75–8.30) 8.09 (6.27–10.13) −2.134 0.033 Uric acid (umol/L) 278.00 (209.50 −375.50) 321.00 (184.00 −414.50) −0.422 0.673 Creatinine (umol/L) 81.00 (62.50–112.00) 99.00 (70.47–153.50) −2.111 0.035 Total protein (g/L) 62.54 ± 7.86 59.80 ± 8.04 2.323 0.021 Albumin (g/L) 34.46 ± 4.80 31.41 ± 4.92 4.231 < 0.001 Open in a new tab Notes: All continuous values were rounded to two decimals; therefore, in some variables the reported median and upper quartile may coincide after rounding. Variables with non-normal distributions were compared using the Mann–Whitney U test (reported as Z); normally distributed variables were compared using the independent-samples t test. Two-sided p < 0.05 was considered statistically significant. IQR, interquartile range; SD, standard deviation; RDW, red cell distribution width; WBC, white blood cell count; CRP, C-reactive protein. Single-marker performance, model discrimination, and classification performance We first evaluated the discriminative performance of individual biomarkers on the held-out test set ( n = 37; bacteremic urosepsis = 18). Among single markers, PCT showed the highest AUC (0.880, 95% CI 0.744–0.975), followed by absolute neutrophil count (AUC 0.766, 95% CI 0.591–0.903) and WBC (AUC 0.713, 95% CI 0.523–0.867) (Supplementary Table S1 ). Given the modest test-set size, single-marker estimates were interpreted cautiously. We further constructed a multivariable LR model using the same candidate predictors as the ML models. LR achieved an AUC of 0.822 (95% CI 0.663–0.935), comparable to RF (AUC 0.822, 95% CI 0.665–0.938), while XGBoost achieved the highest AUC of 0.886 (95% CI 0.763–0.971) (Table 3 ). Table 3. Comparative performance of models on the held-out test set ( n = 37; bacteremic urosepsis = 18). Metric LR (95% CI) RF (95% CI) XGBoost (95% CI) Accuracy 0.730 (0.595–0.865) 0.784 (0.649–0.919) 0.811 (0.676–0.919) Precision 1.000 (1.000–1.000) 0.812 (0.600–1.000) 0.923 (0.733–1.000) Recall 0.444 (0.211–0.684) 0.722 (0.500–0.909) 0.667 (0.429–0.867) F1 Score 0.615 (0.348–0.812) 0.765 (0.571–0.903) 0.774 (0.560–0.909) AUC 0.822 (0.663–0.935) 0.822 (0.665–0.938) 0.886 (0.763–0.971) Open in a new tab Notes: LR, logistic regression; RF, random forest; XGBoost, extreme gradient boosting; AUC, area under the receiver operating characteristic curve; CI, confidence interval. 95% CI were estimated by bootstrap resampling (2000 iterations) on the held-out test set. Classification metrics were computed on the test set using operating thresholds selected by Youden’s J within 5-fold cross-validation on the training set and then applied to the test set. DeLong test comparing XGBoost vs. RF: p = 0.072. DeLong pairwise comparisons were interpreted as exploratory given the modest sample size. To formally compare ROC curves, we applied DeLong’s test on the same test set. The AUC difference between XGBoost and RF did not reach statistical significance ( p = 0.072), and XGBoost was not significantly different from LR ( p = 0.301). In single-marker comparisons, PCT showed no significant difference compared with XGBoost ( p = 0.929), whereas XGBoost significantly outperformed WBC ( p = 0.039) (Supplementary Table S2 ). Pairwise DeLong comparisons were considered exploratory and were not adjusted for multiple testing. Using operating thresholds selected by Youden’s J within training cross-validation and applied to the test set, we further compared classification performance across models (Table 3 ). LR showed high precision but low recall at the selected operating threshold, indicating reduced positive-case capture in this setting. The training and test sets were broadly comparable in baseline characteristics, including outcome prevalence (49.0% vs. 48.6%; p = 0.973), with only minor differences observed for tumor history, platelet count, and albumin (Supplementary Table S3 ). These differences were considered unlikely to materially affect the direction of model comparisons. CAD stratified and adjusted sensitivity analyses We performed CAD-stratified and CAD-adjusted sensitivity analyses for D-dimer. In the CAD-negative stratum, D-dimer was significantly higher in bacteremic urosepsis than in non-bacteremic UTI ( p = 8.66 × 10⁻⁶), whereas in the CAD-positive stratum the same direction was observed but was not statistically significant ( p = 0.456) due to the small sample size (bacteremic urosepsis n = 6) ( Supplementary Table S4 ). In multivariable logistic regression adjusting for age and sex, D-dimer remained positively associated with bacteremic urosepsis (per 1 mg/L FEU increase, adjusted OR 1.22, 95% CI 1.00–1.48; p = 0.048), and the D-dimer×CAD interaction was not significant ( p = 0.206) (Supplementary Table S5 ). These CAD-related analyses were considered exploratory and interpreted alongside the primary models. Patient-level explainability (SHAP) Using SHAP to interpret the XGBoost model, D-dimer, platelet count, PCT, age, RDW, creatinine, WBC, and albumin were the most influential predictors on the held-out test set (Fig. 1 ). Local waterfall explanations further illustrated that these variables acted in combination to increase or decrease the predicted probability of bacteremic urosepsis (Supplementary Figures S1 –S2), supporting clinically interpretable, patient-level reasoning. Model-based feature-importance rankings for RF and XGBoost are provided in Supplementary Figures S3 –S4 and showed similar top-feature patterns. Fig. 1. Open in a new tab Global SHAP summary plot for the XGBoost model on the held-out test set ( n = 37; bacteremic urosepsis = 18). Each dot represents one patient. The x-axis shows the SHAP value (additive contribution to the model output in log-odds), where positive values increase and negative values decrease the predicted probability of bacteremic urosepsis. Dot color indicates the feature value (low to high). Features are ranked by mean absolute SHAP value (global importance). Performance evaluation of ML models Confusion-matrix analysis On the held-out test set ( n = 37; bacteremic urosepsis = 18), the RF confusion matrix showed 16 true negatives, 3 false positives, 13 true positives, and 5 false negatives. The XGBoost model showed 18 true negatives, 1 false positive, 12 true positives, and 6 false negatives (Figs. 2 and 3 ). Overall, XGBoost reduced false positives (1 vs. 3) and achieved slightly higher overall accuracy (30/37 vs. 29/37), corresponding to higher specificity and precision at the selected operating threshold, while RF produced one fewer false negative (5 vs. 6), indicating a slightly higher recall (sensitivity) for detecting bacteremic urosepsis. Fig. 2. Open in a new tab Confusion matrix of the Random Forest model on the held-out test set. The operating threshold was selected using Youden’s J within 5-fold cross-validation on the training set and then applied to the test set. Fig. 3. Open in a new tab Confusion matrix of the XGBoost model on the held-out test set. The operating threshold was selected using Youden’s J within 5-fold cross-validation on the training set and then applied to the test set. ROC curves and AUC assessment As illustrated in Figs. 4 and 5 , both models demonstrated good discrimination on the held-out test set, with XGBoost achieving a numerically higher AUC than Random Forest (0.886 vs. 0.822). However, ROC-curve comparison using DeLong’s test showed that this AUC difference did not reach statistical significance ( p = 0.072). Given the modest size of the held-out test set, AUC comparisons should be interpreted cautiously, and the two models showed broadly comparable discrimination. Consistent with the feature-importance (Supplementary Figures S3 –S4) and SHAP analyses (Fig. 1 ), D-dimer ranked among the most influential predictors in the XGBoost model, supporting the contribution of coagulation-related signals in multivariable risk stratification when combined with inflammatory and host-response biomarkers. Fig. 4. Open in a new tab Receiver operating characteristic (ROC) curve of the Random Forest model on the held-out test set. Thin dotted lines indicate the upper and lower 95% confidence limits obtained by bootstrap resampling ( n = 2000). AUC is reported with 95% CI. Fig. 5. Open in a new tab Receiver operating characteristic (ROC) curve of the XGBoost model on the held-out test set. Thin dotted lines indicate the upper and lower 95% confidence limits obtained by bootstrap resampling ( n = 2000). AUC is reported with 95% CI. Discussion Leveraging RF and XGBoost on routine laboratories obtained within 24 h of presentation, we developed models to identify culture-defined bacteremic urosepsis among inpatients with culture-confirmed UTI and observed good discrimination on held-out data. Because this endpoint is microbiology-defined, it does not fully equate to Sepsis-3, and our findings should be interpreted as early risk stratification for bacteremia within UTI rather than comprehensive sepsis adjudication 15 . Our results are consistent with reports showing strong performance of tree-based methods on tabular clinical data 16 , 17 , including recent Scientific Reports implementations of interpretable/ML-based sepsis-related risk prediction 18 . Across models, top predictors included D-dimer, platelet count, PCT, CRP, creatinine, and leukocyte indices, mapping to coagulation activation, systemic inflammation, and early organ dysfunction in sepsis pathophysiology 7 , 19 – 23 . LR provided a linear baseline with reasonable discrimination, albeit numerically lower than XGBoost in this dataset. Thus, model choice and threshold selection should be guided by clinical priorities, such as minimizing false positives versus maximizing sensitivity for bacteremic urosepsis. Among predictors, D-dimer ranked highly but should be interpreted cautiously. Although it plausibly reflects sepsis-associated coagulation activation and immunothrombosis, it is highly non-specific and may be elevated due to age, cardiovascular disease, malignancy, venous thromboembolism, recent surgery/trauma, and impaired renal clearance 24 . Accordingly, D-dimer alone is insufficient for diagnosis and is better viewed as a risk/severity marker. In our models, its utility derived from multivariable integration with inflammatory markers (PCT/CRP), leukocyte indices, albumin, and renal-function measures, rather than any single marker in isolation. Given the baseline imbalance in CAD, we performed CAD-stratified and CAD-adjusted sensitivity analyses; D-dimer remained associated after adjustment, and the D-dimer×CAD interaction was not significant ( p = 0.206). Albumin also contributed to prediction and likely reflects illness severity and fluid shifts rather than a causal driver. Mechanistically, the relevance of coagulation–inflammation coupling in sepsis and sepsis-induced coagulopathy is supported by contemporary syntheses 25 , 26 , including a recent Scientific Reports update highlighting neutrophil extracellular trap (NET)-associated immunothrombosis pathways in sepsis-induced coagulopathy 27 . Clinical integration and time-critical use. Because predictors are routinely available early after initial evaluation, the model could be implemented as an automated laboratory information system (LIS)/EMR risk score triggered when the first laboratory panel returns. A prespecified operating threshold could support earlier reassessment, timely cultures, antimicrobial optimization, and closer monitoring, while preserving clinician oversight. At the current stage, this should be viewed as a decision-support prototype requiring external validation, calibration, and threshold recalibration before deployment; importantly, clinical impact of sepsis prediction models depends on implementation and workflow integration 28 . The “0–24 h” window reflects the earliest routine tests after presentation rather than a recommendation to delay care; the score can be computed within hours and updated with repeat labs. Future work should develop time-updated models using serial biomarkers over shorter actionable horizons (e.g., 2–6 h, 6–12 h) and evaluate downstream outcomes (intensive care unit (ICU) admission, organ failure progression, mortality) and clinical impact. Post hoc explainability using SHAP provides patient-level auditability by decomposing each individual prediction into additive feature attributions on the model-output scale 29 (i.e., per-biomarker contributions to the predicted risk). However, these attributions describe how the model uses observed associations in the data rather than causal effects; therefore, SHAP findings should be interpreted with clinical context and confirmed through external validation in independent cohorts 30 . Limitations Include the single-center design and modest sample size, which may limit generalizability. Performance was assessed on a single held-out split, so estimates may be unstable; repeated resampling and external validation are needed to better quantify robustness and mitigate overfitting. Calibration was not formally assessed and should be evaluated and potentially recalibrated in larger external cohorts before implementation. The microbiology-based endpoint may exclude culture-negative sepsis, and incomplete SOFA components precluded ΔSOFA-based sensitivity analyses. We also did not evaluate clinical outcomes; therefore, our findings primarily support diagnostic risk stratification rather than prognostication. Conclusion. XGBoost yielded a numerically higher AUC than RF, but the difference was not statistically significant. Both models showed good discrimination using routine biomarkers, with D-dimer, PCT, and albumin among the most informative predictors. ML models based on routine laboratory tests may support earlier risk stratification for culture-defined bacteremic urosepsis among inpatients with UTI, pending external validation. Prospective, multicenter evaluation—including calibration, threshold recalibration, and assessment of downstream clinical impact—is required before broader implementation. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 (493.5KB, docx) Supplementary Material 2 (13.1KB, docx) Supplementary Material 3 (12.8KB, docx) Supplementary Material 4 (16.3KB, docx) Supplementary Material 5 (16.3KB, docx) Supplementary Material 6 (12KB, docx) Acknowledgements We express our gratitude to the staff of the Department of Clinical Laboratory, Fuding Hospital, Fujian University of Traditional Chinese Medicine for their dedication and assistance in data collection and analysis. Abbreviations ALB Albumin AUC Area under the receiver operating characteristic curve CAD Coronary artery disease CFU Colony-forming units CI Confidence interval Cr Creatinine CRP C-reactive protein EDTA Ethylenediaminetetraacetic acid EMR Electronic medical record FEU Fibrinogen equivalent units FN False negative FP False positive FPR False-positive rate GLU Glucose HGB Hemoglobin HIV Human immunodeficiency virus IQR Interquartile range LIS Laboratory information system LR Logistic regression MALDI-TOF Matrix-assisted laser desorption/ionization time-of-flight ML Machine learning PCT Procalcitonin PLT Platelet count qSOFA Quick Sequential Organ Failure Assessment RDW Red cell distribution width RF Random forest ROC Receiver operating characteristic SD Standard deviation Sepsis-3 Third international consensus definitions for sepsis and septic shock SHAP SHapley additive exPlanations SOFA Sequential organ failure assessment TN True negative TP True positive TPR True-positive rate UA Uric acid UTI Urinary tract infection WBC White blood cell count XGBoost Extreme gradient boosting Author contributions Y-L. Z. D-X. Y. and Y-Y. Z. Conceptualization, Methodology, Formal analysis, Investigation, Data Curation, Writing - Original Draft Preparation, Visualization. Contributed equally to this work. J. Z. Software, Validation, Formal analysis, Data Curation, Writing - Review & Editing, Visualization. S-Y. Z. and J. S. Conceptualization, Resources, Data Curation, Supervision, Project Administration, Funding Acquisition, Writing - Review & Editing, Correspondence. Funding Financial support from Project on Clinical Research of Fujian University of Traditional Chinese Medicine, China, Grant/Award Number: XB2024107. Data availability The raw datasets analyzed during this study are available from the corresponding author on reasonable request. Declarations 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. Yuan-Lu Zhang and Dong-Xiao Yu contributed equally to this work. Contributor Information Shi-Yan Zhang, Email: [email protected]. Jinbao Shi, Email: [email protected]. References 1. Al Lawati, H., Blair, B. M. & Larnard, J. Urinary Tract Infections: Core Curriculum 2024. Am. J. Kidney Dis. 83 , 90–100. 10.1053/j.ajkd.2023.08.009 (2024). [ DOI ] [ PubMed ] [ Google Scholar ] 2. Global burden of bacterial antimicrobial resistance. in 2019: a systematic analysis. Lancet 399 , 629–655. 10.1016/s0140-6736(21)02724-0 (2022). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 3. Tandogdu, Z. et al. Urosepsis 30-day mortality, morbidity, and their risk factors: SERPENS study, a prospective, observational multi-center study. World J. Urol. 42 , 314. 10.1007/s00345-024-04979-2 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 4. Gotzens, J. et al. Blood Cultures Time-to-Positivity as an Antibiotic Stewardship Tool in Immunocompromised Children with Gram-Negative Bacteraemia. Antibiot. (Basel) . 14 10.3390/antibiotics14080847 (2025). [ DOI ] [ PMC free article ] [ PubMed ] 5. Guarino, M. et al. Update on Sepsis and Septic Shock in Adult Patients: Management in the Emergency Department. J Clin Med 12, (2023). 10.3390/jcm12093188 (2023). [ DOI ] [ PMC free article ] [ PubMed ] 6. Zaki, H. A. et al. Accuracy of procalcitonin for diagnosing sepsis in adult patients admitted to the emergency department: a systematic review and meta-analysis. Syst. Rev. 13 , 37. 10.1186/s13643-023-02432-w (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 7. Norman-Bruce, H. et al. Diagnostic test accuracy of procalcitonin and C-reactive protein for predicting invasive and serious bacterial infections in young febrile infants: a systematic review and meta-analysis. Lancet Child. Adolesc. Health . 8 , 358–368. 10.1016/s2352-4642(24)00021-x (2024). [ DOI ] [ PubMed ] [ Google Scholar ] 8. Iba, T., Helms, J. & Levy, J. H. Sepsis-induced coagulopathy (SIC) in the management of sepsis. Ann. Intensive Care . 14 , 148. 10.1186/s13613-024-01380-5 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 9. Saravi, B. et al. Capillary leak and endothelial permeability in critically ill patients: a current overview. Intensive Care Med. Exp. 11 , 96. 10.1186/s40635-023-00582-8 (2023). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 10. Mușat, F. et al. Machine Learning Models in Sepsis Outcome Prediction for ICU Patients: Integrating Routine Laboratory Tests-A Systematic Review. Biomedicines 12 10.3390/biomedicines12122892 (2024). [ DOI ] [ PMC free article ] [ PubMed ] 11. Yadgarov, M. Y. et al. Early detection of sepsis using machine learning algorithms: a systematic review and network meta-analysis. Front. Med. (Lausanne) . 11 , 1491358. 10.3389/fmed.2024.1491358 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 12. Hollmann, N. et al. Accurate predictions on small data with a tabular foundation model. Nature 637 , 319–326. 10.1038/s41586-024-08328-6 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 13. Su, M., Guo, J., Chen, H. & Huang, J. Developing a machine learning prediction algorithm for early differentiation of urosepsis from urinary tract infection. Clin. Chem. Lab. Med. 61 , 521–529. 10.1515/cclm-2022-1006 (2023). [ DOI ] [ PubMed ] [ Google Scholar ] 14. Zhang, N., Liu, Y., Yang, C. & Li, X. Review of the Predictive Value of Biomarkers in Sepsis Mortality. Emerg. Med. Int. 2024 (2715606). 10.1155/2024/2715606 (2024). [ DOI ] [ PMC free article ] [ PubMed ] 15. Singer, M. et al. The Third International Consensus Definitions for Sepsis and Septic Shock (Sepsis-3). Jama 315 , 801–810. 10.1001/jama.2016.0287 (2016). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 16. Yang, Z., Cui, X. & Song, Z. Predicting sepsis onset in ICU using machine learning models: a systematic review and meta-analysis. BMC Infect. Dis. 23 , 635. 10.1186/s12879-023-08614-0 (2023). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 17. Sun, J., Cheng, K. & Xie, Y. Urinary Tract Infections Detection with Molecular Biomarkers. Biomolecules 14 10.3390/biom14121540 (2024). [ DOI ] [ PMC free article ] [ PubMed ] 18. Liu, Z., Shu, W., Li, T., Zhang, X. & Chong, W. Interpretable machine learning for predicting sepsis risk in emergency triage patients. Sci. Rep. 15 , 887. 10.1038/s41598-025-85121-z (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 19. Zhang, J., Song, C. & Hu, Z. The value of D-dimer-albumin ratio as a prognostic biomarker in critically ill patients with sepsis: A retrospective single-center study. Heliyon 10 , e39057. 10.1016/j.heliyon.2024.e39057 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 20. Qu, X. et al. Prognostic value of baseline plasma D-dimer levels in sepsis: a prospective cohort study. Pract. Lab. Med. 46 , e00498. 10.1016/j.plabm.2025.e00498 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 21. Shi, J. et al. Procalcitonin and C-reactive protein as diagnostic biomarkers in COVID-19 and Non-COVID-19 sepsis patients: a comparative study. BMC Infect. Dis. 24 10.1186/s12879-023-08962-x (2024). [ DOI ] [ PMC free article ] [ PubMed ] 22. Al-Ismaili, Z., Palijan, A. & Zappitelli, M. Biomarkers of acute kidney injury in children: discovery, evaluation, and clinical application. Pediatr. Nephrol. 26 , 29–40. 10.1007/s00467-010-1576-0 (2011). [ DOI ] [ PubMed ] [ Google Scholar ] 23. Turgman, O., Schinkel, M. & Wiersinga, W. J. Host Response Biomarkers for Sepsis in the Emergency Room. Crit. Care . 27 , 97. 10.1186/s13054-023-04367-z (2023). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 24. Franchini, M., Focosi, D., Pezzo, M. P. & Mannucci, P. M. How we manage a high D-dimer. Haematologica 109 , 1035–1045. 10.3324/haematol.2023.283966 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 25. Kato, H. et al. Efficacy and safety of recombinant human soluble thrombomodulin in patients with sepsis-induced disseminated intravascular coagulation - A meta-analysis. Thromb. Res. 226 , 165–172. 10.1016/j.thromres.2023.05.009 (2023). [ DOI ] [ PubMed ] [ Google Scholar ] 26. Stark, K. & Massberg, S. Interplay between inflammation and thrombosis in cardiovascular pathology. Nat. Rev. Cardiol. 18 , 666–682. 10.1038/s41569-021-00552-1 (2021). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 27. Zaiema, S., Shalaby, N. A., Mohamed, T. H., Bayoumy, A. A. M. & Galal, R. Sepsis-induced coagulopathy: recent insights on the role and clinical application of neutrophil extracellular trap formation. Sci. Rep. 16 , 624. 10.1038/s41598-025-31771-y (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 28. Boussina, A. et al. Impact of a deep learning sepsis prediction model on quality of care and survival. NPJ Digit. Med. 7 , 14. 10.1038/s41746-023-00986-6 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 29. Hur, S. et al. Comparison of SHAP and clinician friendly explanations reveals effects on clinical decision behaviour. NPJ Digit. Med. 8 , 578. 10.1038/s41746-025-01958-8 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 30. Bilodeau, B., Jaques, N., Koh, P. W. & Kim, B. Impossibility theorems for feature attribution. Proc. Natl. Acad. Sci. U S A . 121 , e2304406120. 10.1073/pnas.2304406120 (2024). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials Supplementary Material 1 (493.5KB, docx) Supplementary Material 2 (13.1KB, docx) Supplementary Material 3 (12.8KB, docx) Supplementary Material 4 (16.3KB, docx) Supplementary Material 5 (16.3KB, docx) Supplementary Material 6 (12KB, docx) Data Availability Statement The raw datasets analyzed during this study are available from the corresponding author on reasonable request. Articles from Scientific Reports are provided here courtesy of Nature Publishing Group ACTIONS View on publisher site PDF (1.9 MB) Cite Collections Permalink PERMALINK Copy RESOURCES Similar articles Cited by other articles Links to NCBI Databases Cite Copy Download .nbib .nbib Format: AMA APA MLA NLM Add to Collections Create a new collection Add to an existing collection Name your collection * Choose a collection Unable to load your collection due to an error Please try again Add Cancel Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top