Skip to main content

Prediction of carbapenem-resistant gram-negative bacterial bloodstream infection in intensive care unit based on machine learning



Predicting whether Carbapenem-Resistant Gram-Negative Bacterial (CRGNB) cause bloodstream infection when giving advice may guide the use of antibiotics because it takes 2–5 days conventionally to return the results from doctor's order.


It is a regional multi-center retrospective study in which patients with suspected bloodstream infections were divided into a positive and negative culture group. According to the positive results, patients were divided into the CRGNB group and other groups. We used the machine learning algorithm to predict whether the blood culture was positive and whether the pathogen was CRGNB once giving the order of blood culture.


There were 952 patients with positive blood cultures, 418 patients in the CRGNB group, 534 in the non-CRGNB group, and 1422 with negative blood cultures. Mechanical ventilation, invasive catheterization, and carbapenem use history were the main high-risk factors for CRGNB bloodstream infection. The random forest model has the best prediction ability, with AUROC being 0.86, followed by the XGBoost prediction model in bloodstream infection prediction. In the CRGNB prediction model analysis, the SVM and random forest model have higher area under the receiver operating characteristic curves, which are 0.88 and 0.87, respectively.


The machine learning algorithm can accurately predict the occurrence of ICU-acquired bloodstream infection and identify whether CRGNB causes it once giving the order of blood culture.

Peer Review reports


Among the infections of severe patients, the mortality of bloodstream infections is the highest [1]. Recently, with the prevalence of multiple drug-resistant bacteria (MDR) in China, bloodstream infections caused by MDR are becoming more common in critically ill patients [2]. Common clinical MDRs include carbapenem-resistant Enterobacteriaceae (CRE), carbapenem-resistant Acinetobacter baumannii (CRAB), and carbapenem-resistant Pseudomonas aeruginosa (CRPA), which has been on the list of priority pathogens by the World Health Organization [3]. Today, the most common and dangerous MDR in China is Carbapenem-resistant gram-negative bacilli (CRGNB). Empirical antibiotic therapy becomes very difficult once MDR must be considered due to different drug resistance mechanisms, local epidemiology, site of infection, immune status, history of antimicrobial exposure, and MDR colonization records [2].

Blood culture and susceptibility tests, the gold standard of bloodstream infection, guide the empirical antibiotic regimen to targeted use [1]. However, the blood culture needs 1–3 days from sampling to results, depending on the bacterial load in the blood, making it relatively hysteretic to diagnose bloodstream infections [4]. Additionally, this duration may lead to sepsis progression to septic shock and even death which cannot meet the demand of time-racing rescue of sepsis [5]. If there are artificial intelligence predictive tools to remind the incidence of drug-resistant bacteria bloodstream infection when doctors make the order of blood culture, it may help clinicians to determine the antibiotic regimen, which may be helpful to the control of bloodstream infection.

The use of machine learning to predict bloodstream infection of MDR provides a possible solution from another point of view, which is more meaningful for guiding the use of antibiotics. Application of machine learning in MDR-GNB is increasing, including predicting the risk of MDR-GNB infection, predicting whether known infections originate from MDR-GNB, and guiding antibiotic management and prevention and control of MDR-GNB. According to current research, predicting the occurrence of bloodstream infection is feasible [6,7,8,9,10]. Michael Roimiuse and his colleagues used the MIMIC and RHCC databases to predict acquired bloodstream infection in patients with suspected ICU infection using the XGBoost prediction model. The area under the receiver operating characteristic (AUROC) of two centers are 0.89 ± 0.01 and 0.92 ± 0.02, respectively [10]. Four other centers built machine-learning prediction models of bloodstream infection through their central databases, and the AUROC of models varied from 0.77 to 0.82 [6, 8, 9, 11]. No studies predict the occurrence of MDR bloodstream infection, and the guidance for the current situation of MDR bloodstream infection is limited. As the most common and virulent type of MDR, CRGNB are the focus of our study. The antibiotic regimen is determined according to the patient's state, the degree of sepsis shock, the type of bacteria prevalent in the ward once doctors suspect bloodstream infection, and our artificial intelligence can provide some advices for antibiotic use early. Normally, the most commonly used antibiotics were broad-spectrum antibiotics rather than these treat drug-resistant bacteria, including colistin, ceftazidime-avibactam, and tigecycline. This study focuses on predicting whether the pathogenic bacteria of bloodstream infection are MDR bacteria using machine learning algorithms.


Study characteristics

The database in our study is from the general intensive care unit (ICU) database (SHZJU-ICU) of the second affiliated hospital of the Zhejiang University School of Medicine, a large academic teaching hospital in southeast China with 3800 beds in four districts of Hangzhou. General ICU has set up four wards in three of these districts, with 26 beds, 40 beds, and 10 beds to treat critically ill patients independently. Data from the three wards were shared in 2021. A total of approximately 18000 critically ill patients were included in this database, included demographics, vital signs, clinical examination, medication orders, clinical diagnosis, and medical documents.

It was a regional multi-center retrospective study that included all critically ill patients with blood cultures from 2015.01 to 2021.12. In this study, each blood culture is taken as a time point, and the patient data before the time point are collected to form a complete data. The patients included in the study collected information consisting of demography, diagnosis and complications, vital signs, and laboratory indicators within one day before blood culture sampling. The antibiotic use records and clinical status data were collected two weeks before sampling. Any positive blood culture was marked as a positive patient during hospitalization, but removed the repetitive data with multiple blood culture and only remained the first positive data regardless of whether the subsequent results are positive or not. Only the first blood culture data entry was left in patients with repeated negative blood culture. This study divided patients with suspected bloodstream infections into culture-positive and negative groups. According to the positive results, patients were divided into the CRGNB group and other groups. The machine learning algorithm was used to predict whether the blood culture was positive and whether the pathogen was CRGNB. Figure 1 shows the flow chart. This study has passed the ethics approval of second affiliated Hospital of Zhejiang University school of medicine, and the approval number is IRB-2016–1511.

Fig. 1
figure 1

The logical flow of model prediction and flow chart of our study. CR-GNB Carbapenem-resistant gram-negative bacteria, BSI Bloodstream infection, BC Blood culture

Variable definition

Inclusion and exclusion criteria: all patients admitted to the ICU were included in our study. Exclusion criteria: patients with a hospital stay of fewer than 48 h, Blood culture results taken within 48 h after admission into ICU. We retained the first positive blood culture regardless of whether the subsequent results are positive or not and removed others repeatedly. The history of all antibiotic use refers to the records of antibiotics used within two weeks; all laboratory data are taken from the past 24 h of blood culture sampling. If there are duplicate data, take the average. Invasive catheterization defined as the catheterization of invading blood vessels, including central venous catheter, dialysis catheter, catheter of extracorporeal membrane oxygenation, Swan-Ganz, and Pulse index Continuous Cardiac Output. The carbapenems used in our center includes meropenem, imipenem-cilastatin sodium hydrate, ertapenem, biapenem.

Bacterial detection

Determining and interpreting the minimum inhibitory concentration (MIC) were consistent with the Clinical and Laboratory Standards Institute (CLSI) standards [12, 13]. Carbapenem resistance was defined as reduced susceptibility with a minimum inhibitory concentration of ≥ 2 mg/L for imipenem or meropenem. The definition of tigecycline MIC follows the criteria of the European Committee on Antimicrobial Susceptibility Testing (EUCAST), MIC > 2 mg/L [14]. A laboratory physician conducted the drug susceptibility test with analysis instruments (VITEK2 AST-GN16 France). The broth dilution method determined the strains with mediated sensitivity to tigecycline. We use the blood culture instrument and media system Provided by bioMerieux BacT/ALERT® 3D with Mycobacteria Indication.

Statistical methods

We use R v.3.6.3 (R Core Team, 2020) and RStudio v.1.4.1029 (RStudio Team, 2020) with related packages to accomplish our data analysis [15]. We use median and quartile spacing presentation for numerical variables with non-normal distribution and use mean and standard deviation for normal distribution data. Extreme values and outliers will be deleted. Then, variables with missing values exceeding 40% were excluded, and cases with missing variables exceeding 50% were removed. Variables with missing values ranging from 10 to 40% were addressed through multiple imputations, while those with less than 10% missing data were filled via simple interpolation [16, 17]. Sensitivity analysis is used to evaluate the stability of multiple interpolation. Multivariate logistic regression and three machine learning algorithms, decision tree, random forest, SVM, and XGBoost, were selected to establish models with corresponding R packages such as "glm," "rpart," "randomForest," "xgboost," and "rattle" [18, 19]. All samples were grouped into a 70% training set, a 15% validation set, and a 15% test set, which was is a relatively common proportion of distribution [8, 20, 21]. The decision tree model produces different branches by calculating the characteristics of independent variables and divides the data into subsets with similar features to achieve classification [19]. Other algorithms are all classification algorithms based on the decision tree. Random forest is an independent but comprehensive decision of hundreds of decision trees with higher accuracy than the decision tree. XGBoost, an improved version of the Gradient Boosting algorithm, builds numerous interrelated decision trees and is highly efficient and flexible [18]. Group comparisons were assessed using an independent sample T-test, a chi-square test, and multivariate logistic regression. Multivariate logistic regression used the step-by-step decreasing method to adjust the parameters to reduce the variable collinearity. The odds ratio (OR) and 95% confidence interval (CI) were calculated to evaluate the association strength. Statistical significance was assigned to a P value of less than 0.05. The evaluation parameters included sensitivity, specificity, positive predictive value, negative predictive value, and AUROC curves. We evaluate the bias of the prediction model through PROBAST framework and write the prediction model through the TRIPOD scheme to ensure the structural integrity of the research [22, 23].


Approximately 18,000 patients were included in the SHZJU-ICU database, of which 8652 effective blood cultures, and 2375 patients were considered as bloodstream infection during hospitalization. According to the exclusion criteria, 952 patients with positive blood cultures and 1422 negative patients were included (Fig. 1).

There was no statistical difference in gender and age between the two groups. The incidence of bloodstream infection was higher in patients with infectious diseases (19.9% vs. 7.4%, P < 0.001, OR = 3.10) and internal diseases (6.2% vs. 2.7%, P < 0.001, OR = 2.41), and there was no significant difference in other diseases, including trauma, cerebrovascular accident, heart disease, and surgical disease. The positive rate of blood culture was higher in patients with multiple organ dysfunction syndrome, including acute kidney injury (26.1% vs. 16.0%, P < 0.001, OR = 1.85) and respiratory failure with mechanical ventilation (72.1% vs. 57.5%, P < 0.001, OR = 1.91). Immunosuppressant use (48.8% vs. 28.1%, P < 0.001, OR = 2.44) and a history of multiple antibiotics are also high-risk factors for bloodstream infection. All laboratory markers before blood culture are significant and specific high-risk factors can be found in Table 1. The hospitalization time and the stay in the ICU of patients with bloodstream infections were more prolonged than that of negative patients. Meanwhile, the mortality was higher. Multivariate analysis showed that the heart rate (P < 0.001, OR = 1.98), temperature (P < 0.001, OR = 1.92), procalcitonin (P < 0.001, OR = 1.97), and lactic acid (P = 0.013, OR = 1.22) before blood culture, a carbapenem (P < 0.001, OR = 2.89) and glycopeptide use history (P < 0.001, OR = 1.98) were the high factors of bloodstream infection (Table 2). The effectiveness of prediction model shows that the random forest model has the best prediction ability, with AUROC being 0.86, followed by the XGBoost prediction model. The best prediction model accuracy is 77.7%, the highest specificity is 83.6%, and the sensitivity is poor, with a maximum of 70.5% (Table 3) (Fig. 2).

Table 1 Baseline characteristic and variables of bloodstream infection in ICU
Table 2 Parameters in the multivariable logistic regression model of bloodstream infection in ICU
Table 3 Test set evaluation result of machine learning and multiple logistic regression model for bloodstream infection and CRGNB bacteremia
Fig. 2
figure 2

AUROC of the test set for four machine learning algorithm and multivariable logistic regression model in bloodstream prediction model and CRGNB bloodstream prediction model

Among patients with positive blood cultures, there were 418 patients in the CRGNB group and 534 in the non-CRGNB group. Among the 418 cases of CRGNB bloodstream infection, CRKP and CRAB accounted for 168 and 187 cases respectively, with CRPA of 54 cases, CRE of 9 cases (supplement Table 1). The incidence of CRGNB bloodstream infection in trauma patients was higher, and there was no statistical difference in other diseases. Mechanical ventilation, history of glucocorticoid use, immunosuppressive condition, invasive catheterization, quinolone, carbapenem, and cephalosporin use history are high-risk factors for CRGNB bloodstream infection (Table 4). In multivariate analysis, the main high-risk factors were mechanical ventilation, invasive catheterization, and carbapenem use history (Table 5). Sankey Diagram can visually show the relationship between different species of CRGNB and high-risk factors in Fig. 3.

Table 4 Baseline characteristic and variables of CRGNB Bacteremia in ICU
Table 5 Parameters in the multivariable logistic regression model of CRGNB Bacteremia in ICU
Fig. 3
figure 3

The relationship between different species of CRGNB and high-risk factors including mechanical ventilation, invasive catheterization, and carbapenem use history

In the CRGNB prediction model analysis, the SVM and random forest model have higher AUROC curves, which are 0.88 and 0.87, respectively (Table 3). In the model performance, the sensitivity is significantly improved compared to the blood flow infection prediction model, and the sensitivity of random forest model is 89.1. The SVM model has the highest overall accuracy, reaching 75%.


This study mainly presents the following clinical scenario: when clinicians suspect bloodstream infection and make blood culture orders, artificial intelligence predicts the positive rate of blood culture and the probability caused by CRGNB culture result based on previous data. There are some clinical studies on the predictive model of positive blood culture, but no research of CRGNB. This study shows that the artificial intelligence algorithm has the potential to predict the occurrence of nosocomial CRGNB bloodstream infection accurately. Limited by the longer blood culture cycle and fewer choices of antibiotics, it is necessary to predict CRGNB bloodstream infection earlier so that the empirical antimicrobial regimen can be transformed into the target antimicrobial regimen more quickly and the use of unnecessary broad-spectrum antibiotics can be reduced [24].

In this study, the AUROC of our bloodstream infection prediction model was 0.86. It was similar to the previously published bloodstream infection prediction model, and the AUROC was between 0.82 and 0.926 [6,7,8,9,10]. The judgment of the outcome of the bloodstream infection prediction model is based on the results of blood culture, and this outcome index is very consistent with the construction of the binary machine learning algorithm which is better than traditional multivariate logistic model based on multi-dimensional data analysis. But the same problem with these studies is that blood culture is not reliable. Blood culture can be regarded as bloodstream infection, but bloodstream infection is not necessarily positive blood culture. Before blood culture breaks through its own defects, false negative is almost inevitable. Although there are many blood culture results in our study, we only choose the results of the first sampling in patients. Follow-up blood culture may lead to misleading due to the use of various antibiotics. In addition, continuous negative results may lead to a significant increase in negative samples, resulting in inadequate model fitting. These factors are often ignored in other studies, resulting in false high accuracy.

On this basis, the accuracy of predicting bloodstream infection caused by CRGNB was excellent, with an AUROC of 0.88. There are very few studies on this perspective. The researchers conducted a retrospective study of patients with hematologic diseases in a tertiary hospital in Barcelona. Typically, 3235 episodes of neutropenia and 180 infections (5.6%) were recorded in 349 patients. The machine learning algorithm predicted the incidence of MDR-GNB infection with an AUROC of 0.79 [25]. Our research has a more significant amount of data and a precise prediction result, which is of great significance in improving the accuracy of the prediction model.

For ICU patients suspected of bloodstream infection, the positive rate of blood culture in patients with severe sepsis before the use of antibiotics was 31.4% to 50.6% [26, 27]. The positive rate of blood culture was approximately 40%. A prospective study of patients in medical wards of 31 centers in Italy found that the incidence of MDR-GNB bloodstream infection was 48.2%. High-risk factors included advanced age, previous hospitalization history, and history of antibiotic use [28]. It seems to confirm the current epidemic trend of drug-resistant bacteria that CRGNB accounts for nearly half of the patients with positive blood culture in our central, similar to the international epidemiological data. The isolation rate of CRGNB in hospital-acquired infections is generally high. Statistics worldwide show that the isolation rate of CRGNB in Southeast Asia varies from 26 to 65% [27]. Our data do not fully represent the epidemiological characteristic of CRGNB bloodstream infection in our center, which is related to the exclusion criteria of this study. We removed bloodstream infections from community sources as much as possible. At the same time, it is difficult to achieve no antibiotic exposure if bloodstream infection occurs during hospitalization in the ICU, which reduces the detection rate of sensitive bacteria and adds to the detection rate of MDR in the study. Therefore, the data of our center only show ICU-related CRGNB bloodstream infection to a certain extent. There were KPC-2 in 96.7% genotype of the CRKP through epidemic survey in Zhejiang, China. We did not carry out genotype testing for CRPA and CRAB, so we did not present genotype in the article [29,30,31]. Our study provides a reference for nosocomial prevention and control, identifies patients with high-risk drug-resistant bacteria infection and implements more accurate contact isolation and hand hygiene to reduce costs and increase efficiency [32]. The prediction model provides some preliminary references for antibiotic management. With the increase of the amount of data and model optimization, the prediction model can provide more convincing guidance and suggestions.

The accuracy of the bloodstream infection prediction model and AUROC is not very high in our research. This may be related to the low positive rate of blood culture, and there are many false negative cases and false positive due to contaminations. Due to varying sampling times and early antibiotic exposure, many septic shock patients were considered as bloodstream infections but with negative blood cultures. These patients are classified as negative cases, increasing the uncertainty of predictive model classification [33]. There are false positive and false negative, which is related to the level of hospitals, the timing of blood culture and the history of the use of antibiotics. Although there are next generation sequencing (NGS) and polymerase chain reaction (PCR) results as a supplement [34], the status of blood culture as a gold standard cannot be shaken in the short term. Whether we can include PCR or NGS results to enhance the diagnosis of bloodstream infection needs more literature support. With the continuous improvement of detection technology, the diagnosis of bloodstream infection is more accurate, and the prediction model will be more accurate. This study is a retrospective study of a single database; as a regional multi-center database, the data source is relatively simple, and the prediction model may have the risk of poor external adaptability. There are errors, biases and deficiencies in the clinical variables in the retrospective study. In particular, continuous variables, such as the recording of heart rate, temperature and blood pressure may lead to inaccurate data interpretation due to errors in analytical methods, which may affect the accuracy of model construction. It was a similar limitation in all research by the MIMIC database. Currently, the overall accuracy of artificial intelligence prediction model of CRGNB bloodstream infection is not very high, and the distance to guide the use of clinical antibiotics requires more scenario application practice, which must be confirmed by prospective and even randomized controlled trial research. At the same time, our model also has the problem of external validation, because the incidence of bloodstream infection is different in different MDR epidemic areas. Low and high epidemic areas of CRGNB may not be able to share the same model, which can only be achieved by adjusting the parameters or even modifying the model framework or even reconstructing a model.


Machine learning algorithm can accurately predict the occurrence of ICU-acquired bloodstream infection and identify whether CRGNB causes it. This can provide more references for clinicians to make antibiotic decisions.

Availability of data and materials

The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.



Carbapenem-Resistant Gram-Negative Bacterial


Multiple Drug-Resistant Bacteria


Carbapenem-Resistant Enterobacteriaceae


Carbapenem-Resistant Acinetobacter Baumannii


Carbapenem-resistant Pseudomonas aeruginosa


Carbapenem-resistant Klebsiella Pneumoniae


Intensive Care Unit


Minimum Inhibitory Concentration


Clinical and Laboratory Standards Institute


Odds Ratio


Confidence Interval


Area Under the Receiver Operating Characteristic


Next Generation Sequencing


Polymerase Chain Reaction


  1. Kollef MH, Torres A, Shorr AF, Martin-Loeches I, Micek ST. Nosocomial Infection. Crit Care Med. 2021;49(2):169–87.

    Article  PubMed  Google Scholar 

  2. Timsit JA-O, Ruppé E, Barbier F, Tabah A, Bassetti M. Bloodstream infections in critically ill patients: an expert statement. Intensive Care Med. 2020;46(2):266–84.

    Article  PubMed  PubMed Central  Google Scholar 

  3. Willyard C. The drug-resistant bacteria that pose the greatest health threats. Nature. 2017;543(7643):15.

    Article  CAS  PubMed  Google Scholar 

  4. Ransom EM, Alipour Z, Wallace MA, Burnham CA. Evaluation of Optimal Blood Culture Incubation Time To Maximize Clinically Relevant Results from a Contemporary Blood Culture Instrument and Media System. J Clin Microbiol. 2021;59(3):e02459-20.

    Article  PubMed  PubMed Central  Google Scholar 

  5. Singer M, Deutschman CS, Seymour CW, et al. The Third International Consensus Definitions for Sepsis and Septic Shock (Sepsis-3). JAMA. 2016;315(8):801–10.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  6. Boerman AW, Schinkel M, Meijerink L, et al. Using machine learning to predict blood culture outcomes in the emergency department: a single-centre, retrospective, observational study. BMJ Open. 2022;12(1):e053332.

    Article  PubMed  PubMed Central  Google Scholar 

  7. Ripoli A, Sozio E, Sbrana F, et al. Personalized machine learning approach to predict candidemia in medical wards. Infection. 2020;48(5):749–59.

    Article  PubMed  Google Scholar 

  8. Bhavani SV, Lonjers Z, Carey KA, et al. The Development and Validation of a Machine Learning Model to Predict Bacteremia and Fungemia in Hospitalized Patients Using Electronic Health Record Data. Crit Care Med. 2020;48(11):e1020–8.

    Article  PubMed  PubMed Central  Google Scholar 

  9. 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.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  10. 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.

    Article  PubMed  Google Scholar 

  11. Mahmoud E, Al Dhoayan M, Bosaeed M, Al Johani S, Arabi YM. Developing Machine-Learning Prediction Algorithm for Bacteremia in Admitted Patients. Infect Drug Resist. 2021;14:757–65.

    Article  PubMed  PubMed Central  Google Scholar 

  12. Humphries RM, Abbott AN, Hindler JA. Understanding and Addressing CLSI Breakpoint Revisions: a Primer for Clinical Laboratories. J Clin Microbiol. 2019;57(6):e00203-19.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  13. Humphries RA-OX, Bobenchik AM, Hindler JA, Schuetz AA-OX. Overview of Changes to the Clinical and Laboratory Standards Institute Performance Standards for Antimicrobial Susceptibility Testing, M100, 31st Edition. J Clin Micribiol. 2021;59(12):e0021321.

    Article  Google Scholar 

  14. Giske CA-O, Turnidge JA-O, Cantón RA-O, Kahlmeter G. Update from the European Committee on Antimicrobial Susceptibility Testing (EUCAST). J Clin Microbiol. 2022;60(3):e0027621.

    Google Scholar 

  15. Vuorre M, Crump MJC. Sharing and organizing research products as R packages. Behav Res Methods. 2021;53(2):792–802.

    Article  PubMed  Google Scholar 

  16. Blazek K, van Zwieten A, Saglimbene V, Teixeira-Pinto A. A practical guide to multiple imputation of missing data in nephrology. Kidney Int. 2021;99(1):68–74.

    Article  PubMed  Google Scholar 

  17. Patrician PA. Multiple imputation for missing data. Res Nurs Health. 2002;25(1):76–84.

    Article  PubMed  Google Scholar 

  18. Tianqi C, Guestrin C. XGBoost: A Scalable Tree Boosting System. ACM PRESS; 2016. p. 785–94.

    Google Scholar 

  19. Sanchez-Pinto LN, Luo Y, Churpek MM. Big Data and Data Science in Critical Care. Chest. 2018;154(5):1239–48.

    Article  PubMed  PubMed Central  Google Scholar 

  20. Dong JA-O, Feng T, Thapa-Chhetry B, et al. Machine learning model for early prediction of acute kidney injury (AKI) in pediatric critical care. Crit Care. 2021;25(1):288.

    Article  PubMed  PubMed Central  Google Scholar 

  21. Zhang Y, Hu J, Hua T, Zhang J, Zhang Z, Yang M. Development of a machine learning-based prediction model for sepsis-associated delirium in the intensive care unit. Sci Rep. 2023;13(1):12697.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  22. Moons KGM, Wolff RF, Riley RD, et al. PROBAST: A Tool to Assess Risk of Bias and Applicability of Prediction Model Studies: Explanation and Elaboration. Ann Intern Med. 2019;170(1):W1–33.

    Article  PubMed  Google Scholar 

  23. Collins GS, Reitsma JB, Altman DG, Moons KG. Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD): the TRIPOD statement. Br J Surg. 2015;102(3):148–58.

    Article  CAS  PubMed  Google Scholar 

  24. Sheu CC, Chang YT, Lin SY, Chen YH, Hsueh PR. Infections Caused by Carbapenem-Resistant Enterobacteriaceae: An Update on Therapeutic Options. Front Microbiol. 2019;10:80.

    Article  PubMed  PubMed Central  Google Scholar 

  25. Garcia-Vidal C, Puerta-Alcalde P, Cardozo C, et al. Machine Learning to Assess the Risk of Multidrug-Resistant Gram-Negative Bacilli Infections in Febrile Neutropenic Hematological Patients. Infect Dis Ther. 2021;10(2):971–83.

    Article  PubMed  PubMed Central  Google Scholar 

  26. Cheng MP, Stenstrom R, Paquette K, et al. Blood Culture Results Before and After Antimicrobial Administration in Patients With Severe Manifestations of Sepsis: A Diagnostic Study. Ann Intern Med. 2019;171(8):547–54.

    Article  PubMed  Google Scholar 

  27. Scheer CS, Fuchs C, Gründling M, et al. Impact of antibiotic administration on blood culture positivity at the beginning of sepsis: a prospective clinical cohort study. Clin Microbiol Infect. 2019;25(3):326–31.

    Article  CAS  PubMed  Google Scholar 

  28. Falcone M, Tiseo G, Dentali F, et al. Predicting resistant etiology in hospitalized patients with blood cultures positive for Gram-negative bacilli. Eur J Intern Med. 2018;53:21–8.

    Article  PubMed  Google Scholar 

  29. Liu L, Lou N, Liang Q, et al. Chasing the landscape for intrahospital transmission and evolution of hypervirulent carbapenem-resistant Klebsiella pneumoniae. Sci Bull. 2023;68(23):3027–47.

    Article  Google Scholar 

  30. Li P, Liang Q, Liu W, et al. Convergence of carbapenem resistance and hypervirulence in a highly-transmissible ST11 clone of K. pneumoniae: An epidemiological, genomic and functional study. Virulence. 2021;12(1):377–88.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  31. Gu D, Dong N, Zheng Z, et al. A fatal outbreak of ST11 carbapenem-resistant hypervirulent Klebsiella pneumoniae in a Chinese hospital: a molecular epidemiological study. Lancet Infect Dis. 2017;18(1):37–46.

    Article  CAS  PubMed  Google Scholar 

  32. Liang Q, Zhao Q, Xu X, Zhou Y, Huang M. Early Prediction of Carbapenem-resistant Gram-negative Bacterial Carriage in Intensive Care Units Using Machine Learning. J Glob Antimicrob Resist. 2022;29:225–31.

    Article  CAS  PubMed  Google Scholar 

  33. De Corte T, Van Hoecke S, De Waele J. Artificial Intelligence in Infection Management in the ICU. Crit Care. 2022;26(1):79.

    Article  PubMed  PubMed Central  Google Scholar 

  34. Nutman A, Tellapragada C, Giske CG, Yahav D. New evidence for managing Gram-negative bloodstream infections. Curr Opin Infect Dis. 2021;34(6):599–610.

    Article  CAS  PubMed  Google Scholar 

Download references


Thanks for all the medical staff of ICU for their contribution to nosocomial infection prevention and control; Thanks for the nosocomial infection department for providing professional reference to the details of the study.


Young Innovative Talents Support Program with No. 2022485651 in the Health Department of Zhejiang Province. The funding bodies played no role in the design of the study and collection, analysis, and interpretation of data and in writing the manuscript.

Author information

Authors and Affiliations



LQQ designed the research scheme, analysed the characteristics of the data, constructed the prediction model, and drafted the manuscript. DS extracted and collated clinical data through the database, built models, and collected prospective research data. CJ and XYS determined clinical variables, analysed the importance of model variables, and completed clinical data in prospective studies. CXY established the database and was responsible for the data integrity of analytical data. HM and XZJ supervised the research process, provided clinical reference, and revised the paper. All authors read and approved the final manuscript.

Corresponding author

Correspondence to Man Huang.

Ethics declarations

Ethics approval and consent to participate

Our research involving human data have been performed in accordance with the Declaration of Helsinki. Our research was approved by the ethics committee of the Second Affiliated Hospital of Zhejiang University School of Medicine, and the approval number is IRB-2016–1511. Our data are derived from the special database and had the informed consent form of the subjects, and sensitive information including name, ID card, contact number and so on are deleted through desensitization program. All informed consent was obtained from the patient once admission into ICU and the informed consent we provide can be found in the Supplementary files.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.

Additional information

Publisher’s Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary Information

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, 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 changes were made. 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 The Creative Commons Public Domain Dedication waiver ( applies to the data made available in this article, unless otherwise stated in a credit line to the data.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Liang, Q., Ding, S., Chen, J. et al. Prediction of carbapenem-resistant gram-negative bacterial bloodstream infection in intensive care unit based on machine learning. BMC Med Inform Decis Mak 24, 123 (2024).

Download citation

  • Received:

  • Accepted:

  • Published:

  • DOI: