CoChat
Sign Up
Cited by 4Sepsis Diagnosis and TreatmentMachine Learning in HealthcareArtificial Intelligence in Healthcare

Early prediction of bloodstream infections in ICU patients using machine learning methods based on routine laboratory parameters

Open access
Chen, Yanying, Chen, Ziru, Xu, Dongyun, Li, Shibing, Chen, Feixiong, Yu, Xufa, Cai, Shutao, Zeng, Cuilan +5 more·2025·BMC Infectious Diseases
DOI 10.1186/s12879-025-11580-4PMC PMC12482810
Open this paper in CoChat — free accountOpen PDF

Abstract

BACKGROUND: Bloodstream infections (BSIs) are a major cause of morbidity and mortality in Intensive Care Units (ICUs). Although blood cultures remain the diagnostic gold standard, their long turnaround time may hinder early risk warning. Early risk stratification using routinely available laboratory data may facilitate prompt clinical intervention. This study aimed to develop and validate machine learning (ML) models to predict the likelihood of BSIs in ICU patients based on laboratory tests obtained within the first 24 h of admission. METHODS: This was a retrospective, two-center cohort study using data from the Sixth Affiliated Hospital of Sun Yat-sen University (SAH-SYSU) and the Medical Information Mart for Intensive Care IV (MIMIC, v3.0) database. Adult ICU patients (≥ 18 years) with available first-day laboratory results and blood culture data were included. Multiple ML algorithms, including random forest, XGBoost, GBM, LightGBM, and SVM, were trained and validated using tenfold cross-validation. Model performance was assessed via area under the receiver operating characteristic curve (AUROC), calibration curves, Brier score, and decision curve analysis (DCA). Feature selection was conducted using the Boruta algorithm to develop simplified models. External validation was performed across cohorts using shared features. RESULTS: A total of 754 patients from SAH-SYSU (BSI prevalence: 27.7%) and 3,136 patients from MIMIC (BSI prevalence: 14.0%) were included. Tree-based models outperformed linear classifiers. In internal validation, XGBoost achieved the best performance (AUROC = 0.87 in MIMIC, 0.83 in SYSU). Simplified models using Boruta-selected features retained similar predictive performance (p > 0.05). Cross-cohort validation yielded AUROCs of 0.61 (MIMIC → SYSU) and 0.64 (SYSU → MIMIC). A compact four-feature model demonstrated moderate performance (AUROC up to 0.65), supporting feasibility in resource-limited settings. CONCLUSIONS: ML-based models using only routine laboratory tests from the first ICU day can effectively identify patients at increased risk of BSIs. These models offer a rapid, interpretable, and generalizable tool for early clinical decision-making. Future studies should prospectively validate the model and explore its integration into electronic health records to support real-time risk stratification.

AI snapshot

Written from the paper’s full text.

TL;DR

This retrospective two-center study developed and validated machine learning models using routine laboratory tests from the first 24 h of ICU admission to predict bloodstream infection (BSI) risk. Tree-based models (XGBoost, LightGBM) achieved good internal discrimination (AUROC up to 0.87), but performance dropped markedly in cross-cohort external validation (AUROC 0.61–0.65).

Key findings

  • BSI prevalence was 27.7% (209/754) in the SYSU cohort and 14.0% (439/3,136) in MIMIC; BSI patients showed higher lactate (e.g., MIMIC 3.36 vs. 2.71 mmol/L; SYSU 3.50 vs. 2.41 mmol/L) and lower platelets (MIMIC 171.25 vs. 207.75 × 10⁹/L; SYSU 127.34 vs. 180.33 × 10⁹/L) on ICU day 1 (all p < 0.001).
  • In internal tenfold cross-validation, tree-based models outperformed linear classifiers; XGBoost reached AUROC 0.874 (MIMIC, AUPRC 0.791, Brier 0.109) and LightGBM AUROC 0.865 (SYSU, AUPRC 0.848, Brier 0.148), versus AUROCs below ~0.69 for logistic regression/LDA/glmnet.
  • Boruta feature selection (32 features in MIMIC; 12 in SYSU) yielded reduced models with comparable performance (Wilcoxon p > 0.05 vs. full-feature models); SHAP identified lactate, platelet count, INR, and glucose as top predictors.
  • Cross-cohort external validation degraded substantially: AUROC 0.61 (MIMIC → SYSU, XGBoost) and 0.64 (SYSU → MIMIC, LightGBM); a compact four-feature model (lactate, platelets, PLR, eosinophils) gave best AUROC 0.65 (MIMIC → SYSU) and 0.61 (SYSU → MIMIC).

Methods

  • Retrospective cohort study of adult (≥ 18 years) ICU patients with ≥ 24 h stay, first-day laboratory data, and ≥ 1 blood culture within 7 days of ICU admission, from MIMIC-IV v3.0 (n = 3,136) and the Sixth Affiliated Hospital of Sun Yat-sen University (n = 754); outcome defined by blood culture positivity.
  • Laboratory parameters (CBC, liver/renal function, electrolytes, coagulation, blood gas, plus derived PLR, NLR, LGR) from the first 24 h; missing values imputed with missRanger, with variable missingness thresholds (40% MIMIC, 20% SYSU) and harmonized feature sets.
  • Eight ML algorithms trained via the mlr3 framework with tenfold cross-validation (plus Boruta feature reduction and cross-cohort external validation), evaluated by AUROC, AUPRC, F1, sensitivity/specificity, Brier score, calibration curves, SHAP, and decision curve analysis.

Limitations

  • Stated: differing missing-data thresholds (40% MIMIC vs. 20% SYSU) may have introduced selection bias and reduced cohort comparability; performance declined in external validation due to case-mix, practice, and laboratory platform differences (recalibration suggested); the model does not predict BSI onset timing or causative pathogens.
  • Evident from the text: internal performance relied on tenfold cross-validation within cohorts without a held-out test set, and external AUROCs of 0.61–0.65 indicate limited real-world transferability; prospective validation is still needed (as the authors note).

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1186/s12879-025-11580-4

Cites

22 of 22 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

17 of 27 cited sources checked · 3 supported · 0 flagged · accuracy not yet measured

  • Premisecitation

    Machine learning models have demonstrated significant potential in predicting BSIs by analyzing readily available clinical and laboratory data.

    “Machine learning models have demonstrated significant potential in predicting BSIs by analyzing readily available clinical and laboratory data.”
    Quote in the paperSubstantial claim4 of 5 sources checked
    • cites [10] Liporaci F, Carlotti D, Carlotti A. A machine learning model for the early diagnosis of bloodstream infection in patients admitted to the pediatric intensive care unit. PLoS ONE. 2024;19(5):e0299884. https://doi.org/10.1371/journal.pone.0299884. source
    • cites [11] Zhou T, Ren Z, Ma Y, et al. Early identification of bloodstream infection in hemodialysis patients by machine learning. Heliyon. 2023;9(7):e18263. https://doi.org/10.1016/j.heliyon.2023.e18263. source
    • cites [12] Zhang F, Wang H, Liu L, Su T, Ji B. Machine learning model for the prediction of gram-positive and gram-negative bacterial bloodstream infection based on routine laboratory parameters. BMC Infect Dis. 2023;23(1):675. https://doi.org/10.1186/s12879-023-08602-4. source
    • cites [13] Zoabi Y, Kehat O, Lahav D, Weiss-Meilik A, Adler A, Shomron N. Predicting bloodstream infection outcome using machine learning. Sci Rep. 2021;11(1):20101. https://doi.org/10.1038/s41598-021-99105-2. source

      The Abstract and Discussion report that machine-learning models used electronic medical records, including demographics and blood tests, to predict outcomes in patients with BSI. The Introduction explicitly distinguishes this from predicting who will develop BSI: the study included patients with confirmed infections. Thus the data and predictive potential are supported, but predicting BSIs themselves is not. (LLM, conf=0.88)

      Experimental — accuracy not yet measured
    • cites [14] Roimi M, Neuberger A, Shrot A, Paul M, Geffen Y, Bar-Lavie Y. Early diagnosis of bloodstream infections in the intensive care unit using machine-learning algorithms. Intensive Care Med. 2020;46(3):454–62. https://doi.org/10.1007/s00134-019-05876-8. source
  • Premisecitation

    Diagnostic delays in BSIs exacerbate the risk of multiorgan failure.

    “These diagnostic delays not only exacerbate the risk of multiorgan failure but also contribute to longer hospital stays and higher healthcare burdens [8, 9].”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites [8] Tang F, Yuan H, Li X, Qiao L. Effect of delayed antibiotic use on mortality outcomes in patients with sepsis or septic shock: a systematic review and meta-analysis. Int Immunopharmacol. 2024;129:111616. https://doi.org/10.1016/j.intimp.2024.111616. source
    • cites [9] Sterling SA, Miller WR, Pryor J, Puskarich MA, Jones AE. The impact of timing of antibiotics on outcomes in severe sepsis and septic shock: a systematic review and meta-analysis. Crit Care Med. 2015;43(9):1907–15. https://doi.org/10.1097/CCM.0000000000001142. source

      The Introduction and Discussion address delays in antibiotic administration for severe sepsis and septic shock, not delays in diagnosing bloodstream infections. The Discussion mentions organ dysfunction but does not link diagnostic delay to an increased risk of multiorgan failure. Searches for diagnostic delay, bloodstream infection, and organ failure found no passage making that link. (LLM, conf=0.76)

      Experimental — accuracy not yet measured
  • Premisecitation

    Approximately 47.3% of ICU patients with BSIs receive inadequate therapy within the first 24 h of symptom onset.

    “However, in some BSIs patients with less typical presentations delayed treatment is common: approximately 47.3% of ICU patients with BSIs receive inadequate therapy within the first 24 h of symptom onset [5–7].”
    Quote in the paperSubstantial claim3 of 3 sources checked
    • cites [5] Kollef MH, Shorr AF, Bassetti M, et al. Timing of antibiotic therapy in the ICU. Crit Care. 2021;25(1):360. https://doi.org/10.1186/s13054-021-03787-z. source

      The Conclusions (no page given) discuss delayed appropriate therapy in bloodstream infections, and the Abstract says ICU antibiotics are often not optimally administered. Neither passage reports that 47.3% of ICU patients with BSIs receive inadequate therapy within 24 hours of symptom onset, or links delays to less typical presentations. (LLM, conf=0.83)

      Experimental — accuracy not yet measured
    • cites [6] Weiss SL, Fitzgerald JC, Balamuth F, et al. Delayed antimicrobial therapy increases mortality and organ dysfunction duration in pediatric sepsis. Crit Care Med. 2014;42(11):2409–17. https://doi.org/10.1097/CCM.0000000000000509. source

      The Results report that 29 of 130 pediatric sepsis patients received inappropriate initial antimicrobial therapy, but do not give a rate for ICU patients with bloodstream infections. Searches for the 47.3% figure and inadequate therapy within 24 hours found no supporting passage in our copy. (LLM, conf=0.75)

      Experimental — accuracy not yet measured
    • cites [7] Zasowski EJ, Bassetti M, Blasi F, et al. A systematic review of the effect of delayed appropriate antibiotic treatment on the outcomes of patients with severe bacterial infections. Chest. 2020;158(3):929–38. https://doi.org/10.1016/j.chest.2020.03.087. source

      The Introduction says many patients with severe bacterial infections experience delays in receiving appropriate antibiotics, and the Discussion reports delayed therapy in 27% of patients in its mortality analysis. Neither passage gives a 47.3% estimate for ICU patients with bloodstream infections or establishes a 24-hour window from symptom onset. (LLM, conf=0.88)

      Experimental — accuracy not yet measured
  • Premisecitation

    Most reported ML-based BSI prediction models integrate a combination of vital signs, clinical scores, and laboratory values.

    “However, most reported ML-based BSI prediction models integrate a combination of vital signs, clinical scores, and laboratory values.”
    Quote in the paperSubstantial claim4 of 5 sources checked
    • cites [10] Liporaci F, Carlotti D, Carlotti A. A machine learning model for the early diagnosis of bloodstream infection in patients admitted to the pediatric intensive care unit. PLoS ONE. 2024;19(5):e0299884. https://doi.org/10.1371/journal.pone.0299884. source

      The Methods (Model development; page not available) say this paper’s model combines clinical data, vital signs, and laboratory data. The Discussion describes several other BSI prediction models with differing inputs, but does not establish that most reported models combine all three categories or use clinical scores. (LLM, conf=0.84)

      Experimental — accuracy not yet measured
    • cites [11] Zhou T, Ren Z, Ma Y, et al. Early identification of bloodstream infection in hemodialysis patients by machine learning. Heliyon. 2023;9(7):e18263. https://doi.org/10.1016/j.heliyon.2023.e18263. source

      The Methods section ('Predictors and outcome measures'; page unavailable) says this paper’s ML models used vital signs and laboratory measurements. It does not say that most reported ML-based BSI models use that combination, and its listed predictors do not include clinical scores. (LLM, conf=0.86)

      Experimental — accuracy not yet measured
    • cites [12] Zhang F, Wang H, Liu L, Su T, Ji B. Machine learning model for the prediction of gram-positive and gram-negative bacterial bloodstream infection based on routine laboratory parameters. BMC Infect Dis. 2023;23(1):675. https://doi.org/10.1186/s12879-023-08602-4. source

      The Background describes several ML-based bacteremia models but does not identify their inputs as a combination of vital signs, clinical scores, and laboratory values. The Methods > Variables section describes this paper’s own model as using routine laboratory blood tests. Searches for the claimed combination found no passage establishing what most reported models use. (LLM, conf=0.72)

      Experimental — accuracy not yet measured
    • cites [13] Zoabi Y, Kehat O, Lahav D, Weiss-Meilik A, Adler A, Shomron N. Predicting bloodstream infection outcome using machine learning. Sci Rep. 2021;11(1):20101. https://doi.org/10.1038/s41598-021-99105-2. source

      The Methods section (Variable and feature selection; page unavailable) lists laboratory results among the inputs to this paper’s own BSI outcome models, while the Discussion compares those models with a clinical score. Neither section says that most reported ML-based BSI models combine vital signs, clinical scores, and laboratory values; searches for that combination found no supporting passage in our copy. (LLM, conf=0.72)

      Experimental — accuracy not yet measured
    • cites [14] Roimi M, Neuberger A, Shrot A, Paul M, Geffen Y, Bar-Lavie Y. Early diagnosis of bloodstream infections in the intensive care unit using machine-learning algorithms. Intensive Care Med. 2020;46(3):454–62. https://doi.org/10.1007/s00134-019-05876-8. source
  • Findingown data

    XGBoost AUROC was 0.87 in MIMIC internal validation.

    “In internal validation, XGBoost achieved the best performance (AUROC = 0.87 in MIMIC, 0.83 in SYSU).”
    Quote in the paperSubstantial claim
  • Classified as probably not a substantial claim. 18
  • May combine two claims. 13

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Recent studies have explored various ML algorithms, such as logistic regression, random forest, support vector machines, and gradient boosting, to assess the risk of BSIs in real time.

    “Recent studies have explored various ML algorithms, such as logistic regression, random forest, support vector machines, and gradient boosting, to assess the risk of BSIs in real time [10–14].”
    Quote in the paperNot a substantial claim4 of 5 sources checked
    • cites [10] Liporaci F, Carlotti D, Carlotti A. A machine learning model for the early diagnosis of bloodstream infection in patients admitted to the pediatric intensive care unit. PLoS ONE. 2024;19(5):e0299884. https://doi.org/10.1371/journal.pone.0299884. source

      The Methods (Data processing; Model training and testing, page not provided) describe logistic regression and gradient boosting models for classifying patients with or without BSI. The Discussion (page not provided) reports a random-forest model for predicting central line-associated BSI. Our copy does not identify a support vector machine model, and it describes automatic evaluation as something the trained model could do, rather than real-time risk assessment that these studies performed. (LLM, conf=0.86)

      Experimental — accuracy not yet measured
    • cites [11] Zhou T, Ren Z, Ma Y, et al. Early identification of bloodstream infection in hemodialysis patients by machine learning. Heliyon. 2023;9(7):e18263. https://doi.org/10.1016/j.heliyon.2023.e18263. source

      The Methods and Introduction describe using logistic regression, random forest, support vector machines, and XGboost to predict bloodstream infection in hemodialysis patients. The paper aims at early prediction, but it does not report real-time risk assessment; its models were evaluated on patient datasets. (LLM, conf=0.86)

      Experimental — accuracy not yet measured
    • cites [12] Zhang F, Wang H, Liu L, Su T, Ji B. Machine learning model for the prediction of gram-positive and gram-negative bacterial bloodstream infection based on routine laboratory parameters. BMC Infect Dis. 2023;23(1):675. https://doi.org/10.1186/s12879-023-08602-4. source

      The Background (page unavailable) describes studies using random forest, logistic regression, and support vector machines to predict bloodstream infections or bacteremia. It does not identify gradient boosting among the algorithms, and it does not establish real-time risk assessment. The Discussion (page unavailable) says this paper’s model distinguishes gram-positive from gram-negative bacteremia before pathogen test results are available, rather than assessing whether a patient has a BSI. (LLM, conf=0.87)

      Experimental — accuracy not yet measured
    • cites [13] Zoabi Y, Kehat O, Lahav D, Weiss-Meilik A, Adler A, Shomron N. Predicting bloodstream infection outcome using machine learning. Sci Rep. 2021;11(1):20101. https://doi.org/10.1038/s41598-021-99105-2. source

      The Methods (Development of the models) describe a gradient-boosting model, and the Methods (Variable and feature selection) say its inputs were available when a blood sample was sent for culture, apart from gram-stain information in the inclusive model. But the Introduction specifies that this study predicted poor outcomes among patients with confirmed BSI, not the risk of developing BSI. Our copy does not show this study using logistic regression, random forests, or support vector machines. (LLM, conf=0.88)

      Experimental — accuracy not yet measured
    • cites [14] Roimi M, Neuberger A, Shrot A, Paul M, Geffen Y, Bar-Lavie Y. Early diagnosis of bloodstream infections in the intensive care unit using machine-learning algorithms. Intensive Care Med. 2020;46(3):454–62. https://doi.org/10.1007/s00134-019-05876-8. source

  • May combine two claims.

    Premisecitation

    Diagnostic delays in BSIs contribute to longer hospital stays and higher healthcare burdens.

    “These diagnostic delays not only exacerbate the risk of multiorgan failure but also contribute to longer hospital stays and higher healthcare burdens [8, 9].”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites [8] Tang F, Yuan H, Li X, Qiao L. Effect of delayed antibiotic use on mortality outcomes in patients with sepsis or septic shock: a systematic review and meta-analysis. Int Immunopharmacol. 2024;129:111616. https://doi.org/10.1016/j.intimp.2024.111616. source
    • cites [9] Sterling SA, Miller WR, Pryor J, Puskarich MA, Jones AE. The impact of timing of antibiotics on outcomes in severe sepsis and septic shock: a systematic review and meta-analysis. Crit Care Med. 2015;43(9):1907–15. https://doi.org/10.1097/CCM.0000000000001142. source

      The Introduction says severe sepsis and septic shock are associated with health care costs, but does not link those costs to diagnostic delays. The Discussion addresses delayed recognition and antibiotic timing; neither it nor the searched passages report that diagnostic delays in bloodstream infections lengthen hospital stays or increase healthcare burden (searched: diagnostic delays, length of stay, hospital costs). (LLM, conf=0.76)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    BSI prevalence in the MIMIC cohort was 14.0%.

    “A total of 754 patients from SAH-SYSU (BSI prevalence: 27.7%) and 3,136 patients from MIMIC (BSI prevalence: 14.0%) were included.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The best-performing model, XGBoost, achieved an AUPRC of 0.791 and an F1-score of 0.682 without reweighting or resampling.

    “As the best-performing model (XGBoost) achieved an AUPRC of 0.791 and an F1-score of 0.682 without reweighting or resampling”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Models using only the 4 common important predictors (lactate, platelet count, PLR, and eosinophil counts) yielded a best AUROC of 0.65 (LightGBM: MIMIC → SYSU) and 0.61 (LightGBM: SYSU → MIMIC).

    “These minimal-feature models yielded a best AUROC of 0.65 (LightGBM: MIMIC → SYSU) and 0.61 (LightGBM: SYSU → MIMIC)”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    In the MIMIC-IV cohort, the best-performing model (XGBoost) achieved AUPRC = 0.791, sensitivity = 0.573, specificity = 0.967, F1-score = 0.682, and Brier score = 0.109.

    “in the MIMIC-IV cohort, the best-performing model (XGBoost) achieved AUPRC = 0.791, sensitivity = 0.573, specificity = 0.967, F1-score = 0.682, and Brier score = 0.109.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    In the SYSU cohort, the top model (LightGBM) reached AUPRC = 0.848, sensitivity = 0.763, specificity = 0.830, F1-score = 0.733, and Brier score = 0.148.

    “In the SYSU cohort, the top model (LightGBM) reached AUPRC = 0.848, sensitivity = 0.763, specificity = 0.830, F1-score = 0.733, and Brier score = 0.148.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    MIMIC-IV version 3.0 is a large, publicly available database comprising data from over 65,000 patients admitted to the ICUs of the Beth Israel Deaconess Medical Center in Boston, MA.

    “The Medical Information Mart for Intensive Care (MIMIC-IV version 3.0, hereafter referred to as MIMIC) is a large, publicly available database comprising data from over 65,000 patients admitted to the ICUs of the Beth Israel Deaconess Medical Center in Boston, MA [15].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [15] Johnson AEW, Bulgarelli L, Shen L, et al. Author correction: MIMIC-IV, a freely accessible electronic health record dataset. Sci Data. 2023;10(1):219. https://doi.org/10.1038/s41597-023-02136-9. source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductcitation

    This retrospective study was conducted in accordance with the TRIPOD guidelines.

    “This retrospective study was conducted in accordance with the TRIPOD (Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis) guidelines [16].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [16] Moons KG, Altman DG, Reitsma JB, et al. Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD): explanation and elaboration. Ann Intern Med. 2015;162(1):W1-73. https://doi.org/10.7326/M14-0698. source

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    The reliance on complex and multi-source clinical data may hinder widespread clinical adoption, especially in resource-limited or emergency settings.

    “The reliance on complex and multi-source clinical data may hinder widespread clinical adoption, especially in resource-limited or emergency settings.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductown data

    Ten-fold cross-validation was performed within the MIMIC and SYSU datasets to evaluate internal performance.

    “Ten-fold cross-validation was performed within the MIMIC and SYSU datasets to evaluate internal performance.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductown data

    Models trained on features common to both cohorts were externally validated in the other cohort to assess model transferability across populations.

    “we identified a subset of laboratory features common to both cohorts and trained models using these variables in one cohort while externally validating them in the other, in order to assess model transferability across populations.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    Variable sets were harmonized across datasets to ensure comparability.

    “Variable sets were subsequently harmonized across datasets to ensure comparability.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    The decline in external performance may reflect differences in patient case mix, data collection procedures, clinical management protocols, and laboratory testing platforms between institutions.

    “which may reflect differences in patient case mix, data collection procedures, clinical management protocols, and laboratory testing platforms between institutions.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Study design and conductreasoning

    Different thresholds for missing data were applied when selecting variables in the two datasets (40% for MIMIC vs. 20% for SYSU), reflecting substantial differences in data completeness.

    “First, different thresholds for missing data were applied when selecting variables in the two datasets (40% for MIMIC vs. 20% for SYSU), reflecting substantial differences in data completeness.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationcitation

    Thrombocytopenia in BSI patients may reflect early immune-thrombotic interactions implicated in the pathogenesis of bloodstream infections and sepsis.

    “which may reflect early immune-thrombotic interactions implicated in the pathogenesis of bloodstream infections and sepsis [21, 22].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [21] Audia S, Mahevas M, Samson M, Godeau B, Bonnotte B. Pathogenesis of immune thrombocytopenia. Autoimmun Rev. 2017;16(6):620–32. https://doi.org/10.1016/j.autrev.2017.04.012. source
    • cites [22] Engelmann B, Massberg S. Thrombosis as an intravascular effector of innate immunity. Nat Rev Immunol. 2013;13(1):34–45. https://doi.org/10.1038/nri3345. source

  • May combine two claims.

    Study design and conductreasoning

    The approach of harmonizing variable sets and applying the same preprocessing pipeline may have introduced selection bias and affected direct comparability between cohorts.

    “this approach may have introduced selection bias and affected direct comparability between cohorts.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductnone given

    The model is not designed to predict the exact timing of BSI onset or to identify the causative pathogens, which limits its scope of clinical application.

    “Finally, the model is not designed to predict the exact timing of BSI onset or to identify the causative pathogens, which limits its scope of clinical application.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    Model calibration was evaluated using calibration curves and Brier scores.

    “model calibration was evaluated using calibration curves and Brier scores.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductown data

    Decision curve analysis was performed to estimate the potential clinical net benefit of using the ML models to predict BSIs compared with conventional multivariable logistic regression models.

    “decision curve analysis (DCA) was performed to estimate the potential clinical net benefit of using the ML models to predict BSIs, compared with conventional multivariable logistic regression models.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductreasoning

    Harmonization and identical preprocessing pipelines were used to minimize the impact of potential bias from the missing data threshold approach.

    “harmonization and identical preprocessing pipelines were used to minimize its impact.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    These models offer a rapid, interpretable, and generalizable tool for early clinical decision-making.

    “These models offer a rapid, interpretable, and generalizable tool for early clinical decision-making.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Blood cultures remain the diagnostic gold standard for BSI.

  • May combine two claims.

    Interpretationown data

    The observed laboratory differences indicate early hemodynamic disturbances and possible immune dysregulation in BSI patients.

    “indicating early hemodynamic disturbances and possible immune dysregulation.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    These results suggest that early BSI risk prediction based on routine data is feasible and may support timely clinical decisions in the ICU.

    “These results suggest that early BSI risk prediction based on routine data is feasible and may support timely clinical decisions in the ICU.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductown data

    The Boruta algorithm was applied for feature selection in each cohort to identify important predictors of BSI.

    “we applied the Boruta algorithm for feature selection in each cohort to identify important predictors of BSI.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    The predominance of these predictors underscores their biological relevance and corroborates results from feature selection.

    “This finding underscores their biological relevance and corroborates results from feature selection.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

    Interpretationnone given

    The decline in external performance likely reflects inter-institutional differences in patient characteristics, data distributions, and clinical practice patterns.

    “This decline in external performance likely reflects inter-institutional differences in patient characteristics, data distributions, and clinical practice patterns.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    The statistically significant laboratory differences suggest that critically ill patients with BSI already manifest distinct clinical patterns upon ICU admission, which could be harnessed by machine learning models for early prediction.

    “These statistically significant laboratory differences suggest that critically ill patients with BSI already manifest distinct clinical patterns upon ICU admission, which could be harnessed by machine learning models for early prediction.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductown data

    SHAP analysis was performed on the best-performing model to enhance interpretability and identify influential predictors.

    “we performed SHapley Additive exPlanations (SHAP) analysis on the best-performing model.”
    Quote in the paperNot a substantial claim
29 of 29

  • The paper relies on this without support.

    Blood cultures often require extended incubation periods, typically 24 to 72 h, to yield definitive results.

    “Traditional diagnostic methods, particularly blood cultures, often require extended incubation periods, typically 24 to 72 h, to yield definitive results.”

  • The paper relies on this without support.

    Blood cultures remain the diagnostic gold standard for BSI.

    “Although blood cultures remain the diagnostic gold standard, their long turnaround time may hinder early risk warning.”

  • The paper relies on this without support.

    Platelet-to-lymphocyte ratio, neutrophil-to-lymphocyte ratio, and lactate-to-glucose ratio are associated with inflammation, metabolic stress, and infection severity.

    “These have been previously reported to be associated with inflammation, metabolic stress, and infection severity.”
3 of 3

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Abstract understates the SYSU XGBoost AUROC

    Major

    The abstract reports an internal-validation AUROC of 0.83 for XGBoost in SYSU, whereas Table 3 reports 0.863 for the full-feature SYSU XGBoost model.

  • XGBoost is called the best SYSU model despite a higher LightGBM AUROC

    Minor

    For full-feature internal validation in SYSU, Table 3 gives LightGBM an AUROC of 0.865 and XGBoost an AUROC of 0.863. Thus XGBoost is not the highest-AUROC model there.

  • SYSU linear-model AUROCs are not all below 0.680

    Minor

    The text says all SYSU linear-model AUROCs were below 0.680, but Table 3 reports 0.681 for the full-feature glmnet model.

3 of 3

Cites

Flagged first, then by citations

Cited by

Still being checked.

Authors
Chen, Y., Chen, Z., Xu, D., Li, S., Chen, F., Yu, X., Cai, S., Zeng, C., Ye, X., Yang, J., Liu, J., Lin, J., & Xiao, S.
Affiliations
  • Sun Yat-sen University (CN)
  • Guangdong Academy of Medical Sciences (CN)
Corresponding author
  • Liu, Jun
  • Xiao, Siyu
ORCID
Venue
BMC Infectious Diseases
Publisher
Springer Science and Business Media LLC
Published
2025-09-29
Volume
25
Issue
1
Pages
1196-1196
ISSN
1471-2334 (ISSN-L)
DOI
10.1186/s12879-025-11580-4
PMC
PMC12482810
OpenAlex
W4414601401
Semantic Scholar
c024dce8a4f167cba58db3fdbf701725d2892205
Cited by
4
References
22
Citations by year
2026: 4
FWCI
2.47
Type
journal article
Language
en
Open access
Licence
Locations
  • BMC Infectious Diseases (publishedVersion, cc-by-nc-nd)
  • PubMed (publishedVersion)
  • Europe PMC (PubMed Central) (submittedVersion, cc-by-nc-nd)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
  • PubMed Central (submittedVersion, cc-by-nc-nd)
Funders
Sun Yat-sen University
Topics
  • Health Sciences › Medicine › Epidemiology › Sepsis Diagnosis and Treatment
  • Physical Sciences › Computer Science › Artificial Intelligence › Machine Learning in Healthcare
  • Health Sciences › Health Professions › Health Information Management › Artificial Intelligence in Healthcare
Keywords
bloodstream infection, intensive care unit patients, early prediction, routine laboratory parameters, risk stratification, machine learning prediction models, XGBoost, Boruta feature selection, blood culture, resource-limited settings, electronic health records
MeSH terms
Machine Learning, Random Forest, Boosting Machine Learning Algorithms, Predictive Learning Models, Classification Algorithms, Prediction Algorithms, Adult, Aged, Female, Humans, Intensive Care Units, Male, Middle Aged, Retrospective Studies, ROC Curve, Sepsis

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-23.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-23.

Cited sources are checked by gpt-6-sol.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review