Front. Cell. Infect. Microbiol. Frontiers in Cellular and Infection Microbiology Front. Cell. Infect. Microbiol. 2235-2988 Frontiers Media S.A. 10.3389/fcimb.2025.1667176 Original Research Early diagnosis and prognostic prediction of secondary bloodstream infections caused by Acinetobacter baumannii in critically ill patients by machine-learning algorithms ChenHengxin 1 † Writing – review & editing Writing – original draft GanWenjia 1 † Writing – review & editing Data curation ZhouXianling 2 Data curation Formal analysis Writing – original draft Methodology LiuPingjuan 1 Data curation Software Writing – original draft DingTangdan 1 Writing – original draft Formal analysis XuHongxu 1 Validation Resources Writing – original draft ChenPeisong 1 * Methodology Project administration Software Writing – review & editing Resources ChenYili 1 * Writing – original draft Project administration Funding acquisition Methodology Writing – review & editing Department of Laboratory Medicine, The First Affiliated Hospital of Sun Yat-sen University, Guangzhou, Guangdong, China Department of Clinical Immunology, The Third Affiliated Hospital of Sun Yat-sen University, Guangzhou, Guangdong, China Correspondence: Yili Chen, chenyli3@mail.sysu.edu.cn; Peisong Chen, chps@mail3.sysu.edu.cn

These authors have contributed equally to this work

 08 01 2026 2025 15 1667176 16 07 2025 08 12 2025 28 11 2025 Copyright © 2026 Chen, Gan, Zhou, Liu, Ding, Xu, Chen and Chen. 2026 Chen, Gan, Zhou, Liu, Ding, Xu, Chen and Chen https://creativecommons.org/licenses/by/4.0/ This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms. Background

Secondary bloodstream infections (sBSI) caused by Acinetobacter baumannii (AB) are a major threat to patient safety in the Intensive Care Unit (ICU) due to their prevalence and severity. Developing accurate predictive models is crucial for enhancing clinical decision-making and improving patient outcomes. This study aimed to leverage machine learning (ML) to create a diagnostic model for predicting the risk of AB-sBSI in ICU patients and a prognostic model for assessing the associated 30-day mortality risk.

 Methods

The multicenter, retrospective study enrolled 4,267 ICU patients with AB isolated from non-blood sites. Of these, 337 patients developed bloodstream infection. The analysis included 70 patients with confirmed AB secondary bloodstream infection (AB-sBSI) and 76 age and sex matched controls with non AB-sBSI. For 30-day mortality assessment, the AB-sBSI patients were categorized into non-survivors (n=39) and survivors (n=31). Demographic, microbiological, and laboratory data encompassing hematological, coagulation, and inflammatory markers were analyzed. Fourteen machine learning models were evaluated using the Deepwise and Beckman Coulter DxAI platforms with five-fold cross-validation. Model performance was assessed using five standard metrics, and the DeLong test was applied for AUC comparison. After data preprocessing, patients were enrolled to form an external validation cohort.

 Results

The AB-sBSI risk diagnosis model, constructed with 11 features, identified red cell distribution width as the most significant predictor. The AdaBoost model outperformed both comparative models (Linear Discriminant Analysis, Logistic Regression, LinearSVC) and the conventional biomarker C-reactive protein (AUC = 0.66), with AUCs of 0.937 in training and 0.786 in validation. For 30-day mortality prediction, another model based on 11 features selected lymphocyte count as the most influential variable. The AdaBoost model showed prominent efficacy, surpassing other model (Multilayer Perceptron, BernoulliNB, SGD) and achieving AUC values of 0.986 in training and 0.821 in validation.

 Conclusion

We developed two ML based models for predicting AB-sBSI risk and 30-day mortality. As a preliminary exploration, both models have been converted into accessible web tools. These tools are designed to assist clinicians in making informed decisions and promptly adjusting treatment strategies for critically ill patients.

 Acinetobacter baumannii diagnosis machine learning prognosis secondary bloodstream infections The author(s) declared financial support was received for this work and/or its publication. This work was supported by National Natural Science Foundation of China (82302599), and Guangdong Natural Science Foundation (2023A1515011252). section-at-acceptance Clinical Infectious Diseases Background

Acinetobacter baumannii (AB) poses a significant threat in the ICU setting, particularly due to its association with severe infections and high mortality rates among critically ill patients (Harding et al., 2018). The mortality rate among ICU patients with AB-associated bloodstream infection(AB-BSI) ranges from 34% to 43.4% (Li et al., 2024; Yehya et al., 2025). With the increasing severity of multi-drug resistance and the dwindling options for effective medications, treating AB-BSI has become highly challenging (Doi, 2019). The early diagnosis and treatment of bloodstream infection (BSI) patients are essential to their prognosis, as timely and effective management of the infection can significantly improve outcomes (Timsit et al., 2020).

In critically ill patients, AB-BSI is frequently secondary to the isolation of the same bacterial strain from other body sites, such as pulmonary, intravenous catheter, and abdominal infections, a consequence of the patients’ compromised immune systems (Wisplinghoff et al., 2000; Rodríguez-Baño et al., 2004; Trottier et al., 2007). Consequently, effective diagnostic tools for predicting the risk of AB-associated secondary bloodstream infection (AB-sBSI) and for assessing its prognosis in ICU patients remain lacking. As the use of a single marker offers limited value in the diagnosis of BSI, according to some studies, the simple addition of new biomarkers on the basis of commonly used clinical diagnostic indicators can improve the diagnostic performance to some extent, however this approach does not fully resolve the challenges encountered in clinical practice (Taneja et al., 2017).

Leveraging the power of big data, machine learning (ML) has seen significant advancements (Peiffer-Smadja et al., 2020). These models have demonstrated particular potential in infectious diseases by aiding in the diagnosis of BSI. Models based on ML show high accuracy (AUROCs 0.874-0.94) in predicting BSI, outperforming traditional methods (Liang et al., 2024). Applications include forecasting carbapenem-resistant gram-negative bacteria and early detection in critically ill children (Han et al., 2024).

Beyond diagnosis, ML is increasingly applied in ICU settings to predict critical outcomes such as mortality, readmission, length of stay, and the risk of sepsis or acute respiratory distress syndrome (Sun et al., 2023). While these advances are promising, most existing BSI prediction studies are not specifically focused on secondary BSI caused by AB in ICU, thus leaving a critical gap. The development of dedicated models for AB-sBSI and their subsequent validation through multicenter prospective studies are therefore urgently needed (Ni et al., 2022; Xu et al., 2022).

To fill this gap, the aim of this study is to construct, based on machine learning algorithms, a risk prediction model for AB-sBSI in clinical ICU patients, as well as a model for assessing the prognosis of AB-sBSI.

 Materials and methods Study population and design

A multicenter, retrospective study was conducted between 2014 and 2021 across several institutions in Guangdong Province, China. The participating hospitals included the First Affiliated Hospital of Sun Yat-sen University, Ganzhou Hospital affiliated with Guangdong Provincial People’s Hospital, Heyuan People’s Hospital, Gaoming District People’s Hospital in Foshan City, and Huizhou Central People’s Hospital. Ethical approval for this research was obtained from the Ethics Committee of the First Affiliated Hospital of Sun Yat-sen University. Patients from the Nansha Branch of The First Affiliated Hospital of Sun Yat-sen University, admitted between July 2023 and June 2025, were enrolled to form an external validation cohort, including 18 non-AB-sBSI and 8 AB-sBSI cases.

AS shown in Figure 1, a total of 4,267 intensive care unit (ICU) patients with Acinetobacter baumannii (AB) isolated from non-blood sites (e.g., respiratory tract, urinary tract, skin, and soft tissues) were initially screened. Of these, 337 ICU patients subsequently developed bloodstream infection (BSI). The final study cohort consisted of 70 patients with AB-associated secondary BSI (AB-sBSI) and 76 matched controls with non-AB-sBSI, matched based on age and sex (Mu et al., 2022).

Data-cohort-workflow.

Flowchart depicting patient selection for a study. From 4,267 patients with a positive culture of AB, 337 had bloodstream infections. Exclusions included primary bloodstream infection, coinfection of other pathogens, and incomplete records, leaving 70 for AB-sBSI and 76 for non-AB-sBSI. Both groups were divided into training and validation datasets, with AB-sBSI further split by survival (39 non-survival, 31 survival).

The exclusion criteria were as follows: (a) incomplete or missing patient data; (b) coinfection with other pathogenic microorganisms; (c) patients whose first AB-positive culture was obtained from bloodstream; (d) for patients with AB detected in multiple sample cultures, only the first episode was included to ensure data independence. Furthermore, for prognostic analysis, the AB-sBSI group was stratified into survival and non-survival subgroups according to 30-day mortality.

The term “patients with coinfections by other pathogens” in our exclusion criteria refers specifically to patients with concurrent, active, culture-proven infections at a site other than the bloodstream, which were diagnosed by the treating physicians based on a combination of clinical symptoms, radiological evidence, and microbiological culture results.

 Definition of bloodstream infection and secondary BSI

The diagnosis of bloodstream infection (BSI) required the presence of at least one positive blood culture, accompanied by two or more systemic signs of infection (including fever [>38°C], hypothermia [<36°C], tachycardia [heart rate >90/min], tachypnea [respiratory rate >20/min], leukocytosis [>12×109/L], or leukopenia [<4×109/L]) (Centers for Disease Control and Prevention, 2024).

A secondary BSI caused by Acinetobacter baumannii (AB-sBSI) was strictly defined as a BSI episode in which:

The same strain of AB was isolated from both a non-sterile site (e.g., respiratory tract, urinary tract) and a subsequent blood culture.

The non-blood site was clinically confirmed as the primary source of infection.

Primary BSI (i.e., without an identified non-blood source) and co-infections with other pathogens were excluded.

All suspected AB-sBSI cases were independently adjudicated by three senior intensive care physicians. To ensure diagnostic consistency, the physicians based their judgment on a pre-specified checklist incorporating the clinical, microbiological, and radiological criteria outlined above. The Prediction Starting Point (t=0) is definitively set as the time of first AB isolation from other site. The date of BSI onset was defined as the collection date of the first positive blood culture. The principle of the Secondary BSI Attribution Period was applied. All-cause mortality within 30 days following the first positive blood culture was defined as the primary outcome for prognostic analysis (Kontula et al., 2021).

 Data collection

Clinical parameters of patients were obtained from electronic medical records. We recorded demographic data, APACHE II score, underlying diseases, and previous treatments. Additionally, microbiological data, ICU and hospital stay lengths, laboratory test results (hematological, coagulation indices, and inflammatory markers), and 30-day mortality were collected. Day of ICU hospitalization was indeed defined and collected as the time from ICU admission up until the time of first AB isolation from other site.

 Isolation and identification of <italic>Acinetobacter baumannii</italic> strains

Bacterial culture was conducted according to the process of inoculation of clinical microbial specimens after specimen collection. Among them, blood culture was performed by BacT/Alert 3D system (bioMérieux, Marcyl Etoile, France). Species identification was completed using matrix-assisted laser desorption ionization–time-of-flight mass spectrometry (MALDI-TOF MS) (bioMé rieux, Marcy Etoile, France).

 Antimicrobial susceptibility testing

Antimicrobial susceptibilities for isolates were determined with Vitek system (bioMérieux, Marcy l’Etoile, France). Susceptibility testing results were interpreted under the criteria recommended by the Clinical and Laboratory Standards Institute (CLSI, 2024). The quality control strain for susceptibility testing was Escherichia coli ATCC 25922 and K. pneumoniae ATCC 700603. Susceptibility testing results have been added to Supplementary Tables S1 and S2.

 Statistical analysis

Data analysis was performed using SPSS 25.0 software (IBM Corp, Armonk, NY, USA) and Deepwise platform. Data are presented as frequencies (percentages) for categorical variables and mean ± SD or median (IQR) for continuous variables. Group comparisons utilized Student’s t-test, the Wilcoxon test, or the Pearson χ² test. Cox regression was applied to compare survivors and non-survivors among patients with AB-sBSI. The Delong test was employed to compare the AUC values of different models. P-value < 0.05 was defined as the threshold for statistical significant.

 Feature screening

The initial feature set was refined through a structured screening process prior to modeling. First, data quality filtering was applied: variables with more than 40% missing data were excluded (Gong et al., 2023; Rockenschaub et al., 2025). C -reactive protein (CRP) and C -reactive protein and albumin ratio (CAR) were removed in model for predicting the risk of AB-sBSI in ICU patients. Subsequently, univariate analysis was performed using Student’s t-test, the Wilcoxon test (for continuous variables), or the Pearson χ² test (for categorical variables), retaining variables with a p-value < 0.1. To address multicollinearity, a correlation analysis was conducted, removing one feature from any pair with a correlation coefficient > 0.9. L1-based feature selection method was then applied to further refine the predictor set. Finally, selections were reviewed for clinical applicability and cost-effectiveness. RDW was retained as previous studies have shown is an independent risk factor for predicting BSI.

 Model training and validation

We used Deepwise & Beckman Coulter DxAI platform(https://dxonline.deepwise.com) was used as statistical analysis tool for this study. It used well-established python pyradiomics (version 3.0.1) and scikit-learn (version 0.22) package. In this retrospective cohort study, model training employed 5-fold cross-validation. To ensure a rigorous performance assessment, all feature selection was conducted independently within each training fold, and the corresponding test sets were used solely for evaluation to prevent data leakage. Repeat the loop n times and report the highest metric and its hyperparameter from the validation set. Finally, 11 features were used in the construction of the AB-sBSI mortality risk prediction model and the AB-sBSI 30-days mortality risk prediction model. The area under the curveReceiver operating characteristic (AUROC) curve analysis was done to evaluate the performance of each model. The AdaBoost model was finally selected based on its superior performance. Hyperparameter of the top four performing models were showed in Supplementary Tables S3 and S4. The performance of each model was elucidated and compared using metrics such as the area under the curve (AUC), sensitivity, specificity, positive predictive value, and negative predictive value (Chen et al., 2024).

 Features included in the modelling phase

The diagnostic model for predicting AB-sBSI risk in ICU patients comprised multiple clinically relevant variables, including red cell distribution width (RDW), day of ICU hospitalization, APACHE II score, presence of tracheal intubation or tracheostomy, use of carbapenems and antifungal medications, monocyte count (MONO), mean platelet volume (MPV), MPV to platelet count ratio (MPV/PC), D-dimer (D-D), and the type of inpatient ward.

The prognostic model for assessing 30-day mortality risk incorporated the following variables: total and ICU length of stay (LOS), APACHE II score, inflammatory markers (neutrophil-to-lymphocyte ratio (NLR), lymphocyte count (LY), monocyte count (MONO), procalcitonin (PCT), platelet parameters (MPV, MPV/PC, platelet count (PLT)), and most critically, the time interval between the initial isolation of AB from non-blood sites and subsequent BSI.

 Results AB-sBSI risk prediction diagnostic model Demographic and clinical characteristics

As shown in Table 1, patients who developed AB-sBSI were characterized by longer ICU hospitalization (19.000 vs. 13.000, p=0.010), higher APACHE Ⅱ score (19.000 vs. 15.000, p=0.003). Furthermore, they were recorded more frequently tracheal intubation/tracheotomy (94.3% vs 63.2%, p=0.000) and the use of carbapenems (72.9% vs 53.9%, p=0.018) and antifungal (65.7% vs 30.3%, p=0.000). In addition, the results of laboratory examination showed that the MONO count (0.551 ± 0.379 vs 0.685 ± 0.347, p=0.045), PLT count (162.005 ± 109.138 vs 229.073 ± 138.142, p=0.002), MPV (11.900 vs 10.630, p=0.001), MPV/PC (0.080 vs 0.050, p=0.002), D-dimer (D-D) (7.890 vs 4.060, p=0.002), CRP (101.246 ± 66.237 vs 68.841 ± 66.237, p=0.016) and CAR (3.509 ± 1.933 vs 2.385 ± 1.811, p=0.039) were significant differences between two groups. Notably, it was shown that the mortality rate was also higher (55.7% vs 27.6%, p=0.001).

Baseline characteristics.

Variables Non AB-sBSI (n = 76) AB-sBSI (n = 70) p-value
Basic features
Male 52 (68.4%) 54 (77.1%) 0.238
Age (year) 59.158 ± 19.835 53.986 ± 19.491 0.115
Day of ICU hospitalization (day) 13.000(7.000-21.750) 19.000(9.500-42.250) 0.010 **
APACHE Ⅱ score 15.000(11.000-20.500) 19.000(16.000-22.250) 0.003 **
The inpatient ward
EICU 6 (7.9%) 1 (1.4%) 0.150
MICU 10 (13.2%) 25 (35.7%) 0.001 **
NICU 9 (11.8%) 4 (5.7%) 0.194
RICU 1 (1.3%) 3 (4.3%) 0.555
SICU 26 (34.2%) 26 (37.1%) 0.712
Tracheal intubation/tracheotomy 48 (63.2%) 66 (94.3%) <0.001**
Usage of Antibiotics
Carbapenems 41 (53.9%) 51 (72.9%) 0.018*
Antifungals 23 (30.3%) 46 (65.7%) 0.000**
Laboratory findings
WBC (x109/L) 11.837 ± 5.747 12.548 ± 8.364 0.549
NEU (x109/L) 8.110(6.055-12.262) 8.800(6.497-13.095) 0.465
LY (x109/L) 1.049 ± 0.583 0.931 ± 0.555 0.224
MONO (x109/L) 0.685 ± 0.347 0.551 ± 0.379 0.045 *
RDW (%) 0.170(0.150-0.190) 0.170(0.152-0.190) 0.626
PLT (x109/L) 229.073 ± 138.142 162.005 ± 109.138 0.002 **
MPV (fl) 10.630(9.713-11.665) 11.900(10.500-12.900) 0.001 **
MPV/PC (fl/109/L) 0.050(0.030-0.080) 0.080(0.048-0.128) 0.002 **
PT (s) 14.680(13.230-16.800) 14.275(12.700-16.000) 0.321
INR 1.230(1.107-1.460) 1.233(1.093-1.365) 0.821
APTT (s) 39.659 ± 12.217 39.952 ± 11.103 0.883
TT (s) 16.965(15.575-18.385) 17.280(16.000-19.155) 0.153
FIB (g/L) 3.500(2.610-4.250) 2.990(2.127-4.240) 0.335
D-D (mg/L FEU) 4.060(1.870-8.470) 7.890(4.285-18.027) 0.002 **
CRP (mg/L) 68.841 ± 54.389 101.246 ± 66.237 0.016 *
ALB (g/L) 34.425(30.170-37.550) 33.815(29.402-36.250) 0.243
CAR (x10-3) 2.385 ± 1.811 3.509 ± 1.933 0.039 *
PCT (ng/mL) 7.449 ± 13.681 9.070 ± 17.008 0.553
30-days survival 55 (72.4%) 31 (44.3%) 0.001 **
Other time node
Days from Tracheal intubation to the first isolation of AB in non-blood site (day) 6.000(1.000-11.000) 5.000(1.250-9.000) 0.552
carbapenem resistant AB 54(71.1%) 48(68.6%) 0.857

*P<0.05;**P<0.01.

APACHE II, The Acute Physiology and Chronic Health Evaluation II; WBC, leucocyte count; NEU, Neutrophil count; LY, Lymphocyte count; MONO, Monocyte count; RDW, Erythrocyte hemoglobin distribution width; PLT, platelet count; MPV, mean platelet volume; MPV/PC, mean platelet volume/platelet count; PT, prothrombin time; INR, international normalized ratio; APTT, activated partial thromboplastin time; TT, thrombin time; FIB, fibrinogen; D-D, D-dimer; CRP, C reactive protein; ALB, albumin; CAR,C reactive protein/albumin; PCT, procalcitonin; AB, Acinetobacter baumannii.

 Model building and evaluation

Eleven features were chosen as the input variables in our ML model to predict AB-sBSI risk. We compared 14 distinct algorithms to ascertain which one exhibited the most satisfying performance for our datasets. Table 2 and Figure 2a illustrated the predictive capabilities of the top four models for AB-sBSI. Considering the AdaBoost model’s superior AUC in the training set and its statistically significant advantage over the other four models (p < 0.01, Table 3), we decided to adopt the AdaBoost model as our predictive algorithm. The model yielded AUROC results of 0.937 on the training dataset and 0.786 on the validation dataset, along with satisfactory specificity and sensitivity in the training dataset (0.908 and 0.814, respectively). Confusion matrix of AdaBoost model was in the Supplementary Table S5. Decision Analysis and Calibration plots of the AB-sBSI Risk Prediction Diagnostic Model was in the Supplementary Materials 3. The web application of the model can be access in https://dxonline.deepwise.com/prediction/index.html?baseUrl=%2Fapi%2F&id=49464&topicName=undefined&from=share&platformType=wisdom. A screenshot of the webpage was shown in Supplementary Image 1. Upon inputting the required parameters, the system can determine the likelihood of a patient having either AB-sBSI or Non AB-sBSI.

Prediction performance of AB-sBSI using different ML models.

Dataset ML models AUROC Sensitivity Specificity Negative predictive value Positive predictive value
Train datasetN=146 AdaBoost 0.937 (0.900-0.974) 0.814 (0.708-0.888) 0.908 (0.822-0.955) 0.842(0.752-0.904) 0.891 (0.802-0.944)
LinearDiscriminantAnalysis 0.882(0.816-0.934) 0.814 (0.712-0.888) 0.776 (0.668-0.857) 0.820 (0.722-0.890) 0.770 (0.664-0.852)
LogisticRegression 0.873 (0.817-0.929) 0.814 (0.712-0.888) 0.803 (0.697-0.881) 0.824 (0.726-0.894) 0.792 (0.685-0.872)
LinearSVC 0.884 (0.832-0.937) 0.857 (0.761-0.921) 0.776 (0.668-0.857) 0.855 (0.760-0.918) 0.779 (0.674-0.859)
Validation datasetN=146 LogisticRegression 0.787(0.712-0.862) 0.757 (0.648-0.842) 0.724 (0.617-0.810) 0.764 (0.656-0.847) 0.716 (0.608-0.804)
AdaBoost 0.786(0.711-0.860) 0.686 (0.571-0.782) 0.750 (0.643-0.835) 0.722 (0.619-0.806) 0.716 (0.604-0.808)
LinearDiscriminantAnalysis 0.786(0.710-0.863) 0.771 (0.664-0.852) 0.737 (0.630-0.823) 0.778 (0.671-0.859) 0.730 (0.623-0.816)
LinearSVC 0.782(0.706-0.859) 0.743(0.630-0.833) 0.711 (0.604-0.799) 0.750 (0.643-0.835) 0.703(0.587-0.798)

Delong test between four models in diagnostic performance.

Delong test between four models Training delong Validation delong
Models z p z p
LinearDiscriminantAnalysis-AdaBoost -2.866 0.004 ** 0.023 0.982
AdaBoost-LogisticRegression 3.111 0.002 ** -0.073 0.942
LinearSVC-AdaBoost -2.642 0.008 ** -0.135 0.893
LinearSVC-LinearDiscriminantAnalysis 1.576 0.115 -0.365 0.715
LinearSVC-LogisticRegression 1.734 0.083 -0.419 0.675
LinearDiscriminantAnalysis-LogisticRegression 0.306 0.76 -0.117 0.907

**p<0.01.

APACHE Ⅱ, The Acute Physiology and Chronic Health Evaluation II; RDW, Erythrocyte hemoglobin distribution width; MPV, mean platelet volume; MONO, Monocyte count. (A) The AUROC of Prediction performance in the train and validation datasets. (B) Feature Importance in AdaBoost model.

Graphs showing machine learning model performance and feature importance. The first graph (a) displays ROC curves for four models: LinearDiscriminant, LinearSVC, LogisticRegression, and AdaBoost, with separate plots for train and validation datasets. The second graph (b) highlights feature importance in the AdaBoost model, with RDW as the most significant feature, followed by MPV, APACHEIIscore, tracheal intubation or tracheotomy, day of ICU hospitalization, and monocyte count. Clinical features importance

Eleven features served as predictors in the AdaBoost model. Figure 2b highlights the top six laboratory tests with the highest significance features within the AdaBoost model. Each of these six tests had an importance value exceeding 0.05, with RDW holding the most substantial weight. This was followed by MPV, APACHE II score, the presence of tracheal intubation/tracheotomy, and MONO, ranked by their significance.

 Comparison of the AdaBoost model with CRP

We conducted a comparative analysis of the predictive performance between the AdaBoost model and the use of CRP alone, which is one of the most prevalent biomarkers for BSI in clinical practice (as shown in Figure 3a) (Lai et al., 2025). The findings revealed that the AdaBoost model outperformed CRP (AUROC 0.904 vs. AUROC 0.660).

CRP, C reactive protein; APACHE Ⅱ, The Acute Physiology and Chronic Health Evaluation II; NLR, neutrophil-to-lymphocyte ratio. (A) The area under the receiver operating characteristic curves for the comparison of the AdaBoost prediction model with CRP. (B) The area under the receiver operating characteristic curves for the comparison of the AdaBoost prognosis model with APACHE Ⅱ and NLR.

Two receiver operating characteristic (ROC) curves comparing AdaBoost to other models. Chart a compares AdaBoost with C-reactive protein (CRP), showing higher sensitivity and specificity for AdaBoost. Chart b compares AdaBoost with the APACHE II score and neutrophil-to-lymphocyte ratio (NLR), showing similar performance distinctions. Interlabortary external validation

After data preprocessing, patients from the Nansha Branch of The First Affiliated Hospital of Sun Yat-sen University, admitted between July 2023 and June 2025, were enrolled to form an external validation cohort, including 18 non-AB-sBSI and 8 AB-sBSI cases. The performance of the AdaBoost prediction model on this external validation set is provided in Supplementary Tables S6–S8.

 AB-sBSI prognostic prediction model Demographic and clinical characteristics

As reported in Table 4, the baseline characteristics of non-survival and survival groups in patients with AB-sBSI showed that non-survival group based on 30-days mortality was more associated with shorter hospitalization (18.000 vs 59.000, p=0.000), shorter ICU hospitalization (14.000 vs 37.000 days, p=0.009) and higher APACHE Ⅱ score (20.758 ± 4.555 vs 16.815 ± 5.664 points, p=0.004). Routine blood test indices (LY, MONO, PLT, MPV, MPV/PC) and inflammatory markers (PCT) were significantly different between groups (p < 0.05). Additionally, it was found that non survival group had less time to develop AB-sBSI (4.000 vs 13.000, p<0.001).

Comparisons of baseline characteristics of AB-sBSI patients between Non survival and survival.

Variables Non survival (n = 39) Survival (n = 31) p-value
Male 31 (79.5%) 23 (74.2%) 0.761
Age (year) 54.615 ± 18.613 53.194 ± 20.827 0.764
Day of hospitalization (day) 18.000(12.750-29.250) 59.000(38.500-89.500) <0.001 **
Day of ICU hospitalization (day) 14.000(8.000-23.500) 37.000(17.000-74.500) 0.009 **
Temperature (°C) 36.992 ± 0.853 37.129 ± 1.039 0.547
APACHE Ⅱ score 20.758 ± 4.555 16.815 ± 5.664 0.004 **
Laboratory findings
WBC (x109/L) 10.810(5.335-15.082) 11.220(9.440-13.920) 0.605
LY (x109/L) 0.713 ± 0.462 1.185 ± 0.553 <0.001 **
MONO (x109/L) 0.462 ± 0.407 0.654 ± 0.321 0.037 *
RDW (%) 1.537 ± 4.843 2.672 ± 5.874 0.379
PLT (x109/L) 126.905 ± 102.120 206.164 ± 102.808 0.002 **
MPV (fl) 12.175(11.485-13.805) 10.690(10.070-11.970) 0.004 **
MPV/PC (fl/109/L) 0.110(0.065-0.215) 0.060(0.040-0.080) 0.004 **
CRP (mg/L) 110.953 ± 73.786 91.029 ± 57.449 0.355
ALB (g/L) 32.033 ± 4.842 34.450 ± 4.937 0.094
PCT (ng/mL) 3.530(1.710-18.690) 1.115(0.532-5.008) 0.008 **
ProBNP (pg/mL) 3205.00(1156.00-8610.00) 1457.77(738.70-3508.50) 0.085
Days from the first isolation of AB in non-blood site to blood site 4.000(3.000-9.000) 13.000(6.000-25.500) <0.001 **

*P<0.05;**P<0.01.

APACHE II, The Acute Physiology and Chronic Health Evaluation II; WBC, leucocyte count; NEU, Neutrophil count; LY, Lymphocyte count; MONO, Monocyte count; RDW, Erythrocyte hemoglobin distribution width; PLT, platelet count; MPV, mean platelet volume; MPV/PC, mean platelet volume/platelet count; CRP, C reactive protein; ALB, albumin; PCT, procalcitonin; ProBNP, Pterior-brain natriuretic peptide.

Cox regression analysis indicated that the number of hospitalization days, NLR, MPV/PC, and the time interval from the initial isolation of AB bacteria in a non-blood site to the occurrence of AB-sBSI were significantly associated with the 30-day mortality rate among patients diagnosed with AB-sBSI (p < 0.05) (Table 5).

Cox regression analysis of between non-survival and survival in patients with AB-sBSI.

Variables Univariable analysis Multivariable analysis
HR HR (95%CI) p-value RR RR (95%CI) p-value
Day of hospitalization (day) 0.919 0.885-0.954 0.000 ** 0.937 0.894-0.983 0.008 **
Day of ICU hospitalization (day) 1.000 0.996-1.003 0.806
APACHE Ⅱ score 1.071 1.006-1.140 0.031 *
LY 0.114 0.043-0.301 0.000 **
MONO 0.211 0.065-0.680 0.009 **
PLT 0.993 0.989-0.997 0.001 **
NLR 1.035 1.021-1.048 0.000 ** 1.022 1.005-1.039 0.011 *
MPV 1.007 0.996-1.019 0.209
MPV/PC 465.327 37.371-5794.118 0.000 ** 1608.562 11.782-219609.852 0.003 **
PCT 1.022 1.007-1.037 0.004 **
Days from non-blood to blood site 0.887 0.838-0.940 0.000 ** 0.863 0.770-0.966 0.010 *

APACHE II, The Acute Physiology and Chronic Health Evaluation II; LY, Lymphocyte count; MONO, Monocyte count; PLT, platelet count; NLR, neutrophil-to-lymphocyte ratio; MPV, mean platelet volume; MPV/PC, mean platelet volume/platelet count; PCT, procalcitonin.

 Evaluation of different models

Based on the highest AUC of the AdaBoost model in the training set, and its positive difference compared to the other four models (p < 0.01) in Table 6, we ultimately selected the AdaBoost model as our AB-sBSI prognostic prediction model. As shown in Table 7, the AUROC results of AdaBoost were 0.986 and 0.821 on the train and validation datasets, with acceptable specificity and sensitivity in train dataset (0.949 and 0.839, respectively). Figure 4a demonstrated all four models have effective predictive performance in AB-sBSI prognosis. Confusion matrix of AdaBoost model was in the Supplementary Table S9. The web application of the model can be access in https://dxonline.deepwise.com/prediction/index.html?baseUrl=%2Fapi%2F&id=49468&topicName=undefined&from=share&platformType=wisdom. A screenshot of the webpage was shown in Supplementary Image 1. Upon inputting the required parameters, the system can determine the likelihood of a patient having either non-survival and survival.

Delong test between four models in prognostic performance.

Delong test between four models Training Delong Validation Delong
Models z p z p
AdaBoost-SGD 2.51 0.012 * -1.072 0.284
AdaBoost-MultilayerPerceptron 2.44 0.015 * -1.194 0.232
AdaBoost-BernoulliNB 2.731 0.006 ** -1.85 0.064
SGD-MultilayerPerceptron -1.263 0.206 -0.147 0.883
SGD-BernoulliNB 0.313 0.754 -0.526 0.599
MultilayerPerceptron-BernoulliNB 0.652 0.514 -0.524 0.6

Prognosis prediction performance of the outcome of patients with AB-sBSI.

Dataset ML models AUROC Sensitivity Specificity Negative predictive value Positive predictive value
Train dataset AdaBoost 0.986 (0.968-1.000) 0.839 (0.681-0.932) 0.949 (0.826-0.990) 0.881 (0.757-0.951) 0.929 (0.788-0.984)
MultilayerPerceptron 0.931 (0.877-0.985) 0.806 (0.637-0.911) 0.872 (0.723-0.950) 0.850 (0.701-0.935) 0.833 (0.668-0.932)
BernoulliNB 0.921(0.861-0.980) 0.839 (0.681-0.932) 0.821 (0.657-0.922) 0.865 (0.705-0.949) 0.788 (0.621-0.900)
SGD 0.926 (0.868-0.983) 0.742 (0.568-0.868) 0.923 (0.792-0.978) 0.818 (0.669-0.914) 0.885 (0.724-0.961)
Validation dataset BernoulliNB 0.881(0.803-0.959) 0.806 (0.647-0.910) 0.743 (0.574-0.867) 0.829 (0.666-0.926) 0.714 (0.550-0.840)
MultilayerPerceptron 0.867(0.784-0.950) 0.806 (0.647-0.910) 0.795 (0.636-0.901) 0.838 (0.681-0.930) 0.758 (0.595-0.874)
SGD 0.865(0.779-0.951) 0.710 (0.535-0.842) 0.821 (0.664-0.919) 0.780 (0.627-0.888) 0.759 (0.582-0.885)
AdaBoost 0.821(0.721-0.920) 0.677 (0.504-0.816) 0.795 (0.636-0.901) 0.756 (0.598-0.871) 0.724 (0.543-0.858)

APACHE Ⅱ score, The Acute Physiology and Chronic Health Evaluation II; LY, Lymphocyte count; PCT, Calcitonin; Days of AB in NB to B, Days from the first isolation of AB in non-blood site to blood site. (A) The AUROC of Prognosis Prediction performance in the train and validation datasets. (B) Feature Importance in AdaBoost model of Prognosis.

a. Two ROC curves show the AUROC of prognosis prediction models comparing train and validation datasets. Curves for SGD, MultilayerPerceptron, AdaBoost, and BernoulliNB are plotted. b. A horizontal bar chart displays feature importance in the AdaBoost model of prognosis, highlighting factors like lymphocyte count, hospitalization duration, and APACHE II score. Clinical features importance

Figure 4b shows the top six significance values of the selected laboratory tests in AdaBoost model. Each of these six tests exhibited an importance greater than 0.2, with LY having the utmost weight. This was followed by Days of hospitalization, Days of ICU hospitalization, APACHE II score, PCT and Days from the first isolation of AB in non-blood site to blood site, in order of significance.

 Comparison of the AdaBoost model with APACHE II and NLR

We conducted a comparative analysis of the predictive performance between the AdaBoost model and the use of APACHE II and NLR alone, which are the important prevalent biomarkers for prognostic prediction in clinical practice (as shown in Figure 3b) (Liang and Yu, 2022). The findings revealed that the AdaBoost model outperformed APACHE II and NLR (AUROC 0.972 vs. AUROC 0.700 and 0.703).

 Discussion

In this study, we established a diagnostic model for predicting the risk of AB-sBSI in ICU patients and a prognostic model for assessing the associated 30-day mortality risk. These models have been converted into web-based tools to provide clinicians with actionable insights for rapidly adjusting treatment strategies. In line with prior evidence that incorporating overlooked variables can improve predictive accuracy beyond standard models, this study incorporated underutilized clinical variables—such as time-related metrics, invasive mechanical ventilation, and catheterization—to enhance performance (Zhou et al., 2025).

In this study, the AB-sBSI risk model outperformed competing models for both diagnosis and 30-day mortality, as well as demonstrated higher predictive accuracy than established markers such as C-reactive protein, APACHE II, and NLR. The superior performance of AdaBoost can be attributed to its intrinsic ensemble learning mechanism. Unlike linear models such as Linear Discriminant Analysis and Logistic Regression, AdaBoost iteratively refines its predictions by focusing on misclassified instances through sample weight adjustments. This adaptive nature allows it to effectively capture the complex, nonlinear interactions among heterogeneous clinical variables characteristic of the ICU setting (Christodoulou et al., 2019; Henry Basil et al., 2024). The markedly higher AUC observed with AdaBoost underscores the advantage of this approach in modeling the multifaceted pathophysiology of AB-sBSI. Furthermore, feature importance analysis visually identified key clinical predictors, reinforcing that the model’s decisions are driven by a composite of clinically relevant variables rather than relying on a single biomarker.

Given that ICU patients are often critically ill with compromised immunity, AB-sBSI portends a poor prognosis. Early identification of sBSI risk in AB-colonized patients is crucial, as it opens a critical window for initiating targeted anti-infective therapies (Gouel-Cheron et al., 2022). While recent studies have advanced the development of pragmatic diagnostic models for drug-resistant infections—such as an XDRAB risk model with Lasso regression (Shi et al., 2024) and a high-performing hybrid random forest model for ICU multi-drug resistant bacteria (He and Qu, 2023)—they have not specifically addressed the prediction of sBSI in the ICU setting (Roimi et al., 2020). This gap underscores the distinct focus and novelty of the present study.

In the diagnostic model for predicting the risk of AB-sBSI in ICU patients, by leveraging readily accessible clinical indicators, this model enables early identification of sBSI risk following AB isolation, thereby facilitating timely infection control and treatment optimization. Notably, the finding that RDW and MPV were the influential predictors in our model provides critical insights. From a pathophysiological perspective, infection-related oxidative stress is known to inhibit erythrocyte precursor maturation and promote premature release of immature cells into circulation—a mechanism that may explain the elevated RDW values observed in AB-sBSI patients (Mahmood et al., 2014). In the setting of BSI, pro inflammatory cytokines (e.g., IL-6, TNF-α) disrupt iron metabolism and erythropoiesis, leading to the release of immature, variably sized red cells and an increase in RDW (Jin et al., 2025). Previous studies have shown that an elevated RDW on admission independently predicts mortality, length of stay, and the risk of subsequent bloodstream infection in critically ill patients (Tseng et al., 2020). Accordingly, considering both statistical significance and clinical applicability, we incorporated RDW as a variable in the design of our models. The prominence of tracheal intubation/tracheotomy as a predictor underscores its role in AB-sBSI risk. This intervention compromises natural airway defenses (Zilberberg et al., 2016; de Carvalho Baptista et al., 2018), consistent with its established association with infection (Zoabi et al., 2021; Choi et al., 2024).

Furthermore, ML further identified modifiable ICU-related predictors. Although these relationships are correlational, they suggest that reducing ICU stay and optimizing airway management could help mitigate AB-sBSI risk. Future studies could adopt a target-trial emulation (TTE) framework to integrate predictive modeling with causal inference, enabling rigorous evaluation of these factors (Yang et al., 2025).

The 30-day mortality risk model leverages explainable AI to provide interpretable clinical decision support. A crucial finding is the significantly shorter transition time from non-blood AB isolation to bacteremia in non-survivors. This abbreviated interval likely reflects greater disease severity, potentially due to factors such as initial inappropriate antibiotic therapy or more frequent invasive procedures (Ju et al., 2018). Moreover, this parameter likely encapsulates the dynamic interplay between bacterial virulence and compromised host defenses, offering critical insights for timely intervention (Holmes et al., 2024).

Moreover, the significantly higher LY counts observed in the survival group suggested that LY may serves as a surrogate for adaptive immune competence. BSI induced lymphocyte apoptosis and functional exhaustion result in lymphopenia, which correlates with impaired pathogen clearance, higher rates of secondary infection, and increased organ dysfunction (Yang et al., 2015). The clinical accessibility of LY counts within routine complete blood count (CBC) parameters, combined with this mechanistic rationale, supported its inclusion in our model.

Although our findings confirm that the NLR is an independent risk factor for mortality—consistent with prior studies (Huang et al., 2024; Jin et al., 2024; Wei et al., 2024)—it was ranked below six other variables in feature importance. This discrepancy highlights the methodological distinction between Cox regression, which identifies statistically significant associations with outcome at the cohort level, and machine learning feature importance, which reflects a variable’s contribution to individual-level prediction accuracy (Qiu et al., 2020; Sharafi et al., 2024). NLR may represent a universal risk factor with limited interpatient variability, thus offering high epidemiological relevance but lower discriminative power in a multivariable predictive model.

This study was conducted in a multicenter ICU cohort with inherent clinical heterogeneity. While our model provides a general predictive framework, the importance of specific predictors may vary across patient subgroups (e.g., surgical vs. medical). Future work should therefore pursue subgroup-specific analyses to support more personalized prognosis (Yang et al., 2024).

This study has several limitations. First, despite being a multi-center dataset, the sample size was moderate, and external validation was limited, which restricts the model’s stability and generalizability. Second, the retrospective design is inherently susceptible to selection bias and likely underrepresents milder cases. Third, although internal validation methods were applied, the model’s robustness may be compromised by the suboptimal predictor-to-sample ratio and the mean imputation for handling missing data. Therefore, independent external validation is mandated prior to any clinical implementation.

 Conclusion

We have developed two readily accessible online prediction tools: an AB-sBSI Risk Prediction Diagnostic model and a 30-day mortality risk prediction model for ICU patients. Designed to support clinical decision-making, these tools provide a foundation for individualized treatment strategies. As an initial exploratory step, they hold potential to aid in improving patient outcomes and guiding infection prevention and control measures in the ICU. Further validation across diverse populations and healthcare settings will be essential to confirm their broader applicability.

 Data availability statement

The datasets used and/or analyzed during the current study available from the corresponding author on reasonable request. Requests to access the datasets should be directed to chenyli3@mail.sysu.edu.cn.

 Ethics statement

This study was approved by the Institutional Review Board and Ethics Committee of The First Affiliated Hospital of Sun Yat-sen University and conducted according to the Declaration of Helsinki. Consent was obtained from all participants. The studies were conducted in accordance with the local legislation and institutional requirements. The participants provided their written informed consent to participate in this study.

 Author contributions

HC: Writing – review & editing, Writing – original draft, Formal Analysis. WG: Writing – review & editing, Data curation, Methodology. XZ: Data curation, Formal Analysis, Writing – original draft. PL: Data curation, Software, Writing – original draft. TD: Writing – original draft. HX: Validation, Resources, Writing – original draft. PC: Methodology, Project administration, Software, Writing – review & editing, Resources. YC: Writing – original draft, Project administration, Funding acquisition, Methodology, Writing – review & editing.

 Conflict of interest

The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

 Generative AI statement

The author(s) declare that Generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

 Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

 Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fcimb.2025.1667176/full#supplementary-material.

References Centers for Disease Control and Prevention . (2024). Bloodstream infection (BSI) events ( National Healthcare Safety Network). Available online at: https://www.cdc.gov/nhsn/pdfs/pscManual/4PSC_CLABScurrent.pdf (Accessed March 16, 2024). Chen J. Wang Y. Hong M. Wu J. Zhang Z. Li R. . (2024). Application of peripheral blood routine parameters in the diagnosis of influenza and Mycoplasma pneumoniae. Virol. J. 21, 162. doi: 10.1186/s12985-024-02429-4, PMID: 39044252 Choi M. H. Kim D. Park Y. Jeong S. H. (2024). Development and validation of artificial intelligence models to predict urinary tract infections and secondary bloodstream infections in adult patients. J. Infect. Public Health 17, 10–17. doi: 10.1016/j.jiph.2023.10.021, PMID: 37988812 Christodoulou E. Ma J. Collins G. S. Steyerberg E. W. Verbakel J. Y. Van Calster B. (2019). A systematic review shows no performance benefit of machine learning over logistic regression for clinical prediction models. J. Clin. Epidemiol. 110, 12–22. doi: 10.1016/j.jclinepi.2019.02.004, PMID: 30763612 de Carvalho Baptista I. M. Martinho F. C. Nascimento G. G. da Rocha Santos C. E. Prado R. F. D. Valera M. C. (2018). Colonization of oropharynx and lower respiratory tract in critical patients: Risk of ventilator-associated pneumonia. Arch. Oral. Biol. 85, 64–69. doi: 10.1016/j.archoralbio.2017.09.029, PMID: 29031240 Doi Y. (2019). Treatment options for carbapenem-resistant gram-negative bacterial infections. Clin. Infect. Dis. 69, S565–S575. doi: 10.1093/cid/ciz830, PMID: 31724043 Gong H. Ke X. Wang A. Li X. (2023). An interpretable machine-learning model for predicting in-hospital death risk in patients with cardiac arrest: based on the US Medical Information Mart for Intensive Care Database IV 2.0. Chin. J. Crit. Care Med. 38, 528537. doi: 10.12290/xhyxzz.2022-0733 Gouel-Cheron A. Swihart B. J. Warner S. Mathew L. Strich J. R. Mancera A. . (2022). Epidemiology of ICU-Onset Bloodstream Infection: Prevalence, Pathogens, and Risk features Among 150,948 ICU Patients at 85 U.S. Hospitals. Crit. Care Med. 50, 1725–1736. doi: 10.1097/CCM.0000000000005662, PMID: 36190259 Han H. J. Kim K. Park J. D. (2024). Early detection of bloodstream infection in critically ill children using artificial intelligence. Acute Crit. Care 39, 611–620. doi: 10.4266/acc.2024.00752, PMID: 39587863 Harding C. M. Hennon S. W. Feldman M. F. (2018). Uncovering the mechanisms of Acinetobacter baumannii virulence. Nat. Rev. Microbiol. 16, 91–102. doi: 10.1038/nrmicro.2017.148, PMID: 29249812 He J. Qu T. (2023). Progress in the risk prediction model of multiresistant bacteria infection in intensive care unit. Chin. J. Clin. Infect. Dis. 16, 384–390. doi: 10.3760/cma.j.issn.1674-2397.2023.05.008 Henry Basil J. Lim W. H. Syed Ahmad S. M. Menon Premakumar C. Mohd Tahir N. A. Mhd Ali A. . (2024). Machine learning-based risk prediction model for medication administration errors in neonatal intensive care units: A prospective direct observational study. Digit Health 10, 20552076241286434. doi: 10.1177/20552076241286434, PMID: 39430694 Holmes C. L. Albin O. R. Mobley H. L. T. Bachman M. A. (2024). Bloodstream infections: mechanisms of pathogenesis and opportunities for intervention. Nat. Rev. 23, 210–224. doi: 10.1038/s41579-024-01105-2, PMID: 39420097 Huang L. Weng B. Gu X. Wang Y. Wang M. Weng J. . (2024). Performance of various pneumonia severity models for predicting adverse outcomes in elderly inpatients with community-acquired pneumonia. Clin. Microbiol. Infect. 30, 1426–1432. doi: 10.1016/j.cmi.2024.07.008, PMID: 39002660 Jin S. Yin J. B. Li W. Zang L. L. (2024). Effect of neutrophil to lymphocyte ratio on prognosis of elderly patients with severe sepsis combined with diabetes mellitus. BMC Geriatr. 24, 211. doi: 10.1186/s12877-024-04757-0, PMID: 38424501 Jin J. Yu L. Zhou Q. Du Q. Nie X. Yin H. Y. . (2025). Development and validation of a multidimensional predictive model for 28-day mortality in ICU patients with bloodstream infections: a cohort study. Front. Cell Infect. Microbiol. 15. doi: 10.3389/fcimb.2025.1569748, PMID: 40692680 Ju M. H. Yao Y. L. Du C. L. Chen S. Song Y. L . (2018). Subsequent multidrug-resistant bacteremia is a risk factor for short-term mortality of patients with ventilator-associated pneumonia caused by acinetobacter baumannii in intensive care unit: A multicenter experience. Chin. Med. J. 131, 361–363. doi: 10.4103/0366-6999.223859, PMID: 29363656 Kontula K. S. K. Skogberg K. Ollgren J. Järvinen A. Lyytikäinen O . (2021). Population-based study of bloodstream infection incidence and mortality rates, Finland, 2004-2018. Emerging Infect. Dis. 27, 2560–2569. doi: 10.3201/eid2710.204826, PMID: 34546161 Lai S. Li X. Cai D. Mei C. Liang Z . (2025). Prognostic value of NPR and CLR-based nomogram modeling in elderly patients with Acinetobacter baumannii bloodstream infection. BMC Geriatr. 25, 234. doi: 10.1186/s12877-025-05884-y, PMID: 40205541 Li L. Chen D. Liu P. Dai L. Tang Z. Yi S. . (2024). Risk factors for development and mortality of bloodstream infections caused by carbapenem-resistant Acinetobacter baumannii. Infect. Drug Resist. 17, 5699–5706. doi: 10.2147/IDR.S484546, PMID: 39720614 Liang Q. Ding S. Chen J. Chen X. Xu Y. Xu Z. . (2024). Prediction of carbapenem-resistant gram-negative bacterial bloodstream infection in intensive care unit based on machine learning. BMC Med. Inform Decis Mak 24, 123. doi: 10.1186/s12911-024-02504-4, PMID: 38745177 Liang P. Yu F. (2022). Value of CRP, PCT, and NLR in prediction of severity and prognosis of patients with bloodstream infections and sepsis. Front. Surg. 9. doi: 10.3389/fsurg.2022.857218, PMID: 35345421 Mahmood N. A. Mathew J. Kang B. DeBari V. A. Khan M. A. (2014). Broadening of the red blood cell distribution width is associated with increased severity of illness in patients with sepsis. Int. J. Crit. Illn Inj Sci. 4, 278–282. doi: 10.4103/2229-5151.147518, PMID: 25625057 Mu S. Xiang H. Wang Y. Wei W. Long X. Han Y. . (2022). The pathogens of secondary infection in septic patients share a similar genotype to those that predominate in the gut. Crit. Care (London England) 26, 68. doi: 10.1186/s13054-022-03943-z, PMID: 35331299 Ni S. Xu P. Zhang K. Zou H. Luo H. Liu C. . (2022). A novel prognostic model for Malignant patients with Gram-negative bacteremia based on real-world research. Sci. Rep. 12, 11644. doi: 10.1038/s41598-022-15126-5, PMID: 35804024 Peiffer-Smadja N. Rawson T. M. Ahmad R. . (2020). Machine learning for clinical decision support in infectious diseases: a narrative review of current applications. Clin. Microbiol. Infect. 26, 584–595. Qiu X. Gao J. Yang J. Hu J. Hu W. Kong L. . (2020). A comparison study of machine learning (Random survival forest) and classic statistic (Cox proportional hazards) for predicting progression in high-grade glioma after proton and carbon ion radiotherapy. Front. Oncol. 10. doi: 10.3389/fonc.2020.551420, PMID: 33194609 Rockenschaub P. Akay E. M. Carlisle B. G. Hilbert A. Wendland J. Meyer-Eschenbach F. . (2025). External validation of AI-based scoring systems in the ICU: a systematic review and meta-analysis. BMC Med. Inform Decis Mak 25, 5. doi: 10.1186/s12911-024-02830-7, PMID: 39762808 Rodríguez-Baño J. Cisneros J. M. Fernández-Cuenca F. Ribera A. Vila J. Pascual A. . (2004). Clinical features and epidemiology of Acinetobacter baumannii colonization and infection in Spanish hospitals. Infect. Control Hosp Epidemiol. 25, 819–824. doi: 10.1086/502302, PMID: 15518022 Roimi M. Neuberger A. Shrot A. Paul M. Geffen Y. Bar-Lavie Y. (2020). Early diagnosis of bloodstream infections in the intensive care unit using machine-learning algorithms. Intensive Care Med. 46, 454–462. doi: 10.1007/s00134-019-05876-8, PMID: 31912208 Sharafi M. Mohsenpour M. A. Afrashteh S. Eftekhari M. H. Dehghan A. Farhadi A. . (2024). Factors affecting the survival of prediabetic patients: comparison of Cox proportional hazards model and random survival forest method. BMC Med. Inform Decis Mak 24, 246. doi: 10.1186/s12911-024-02648-3, PMID: 39227824 Shi J. Mao X. Cheng J. Shao L. Shan X. Zhu Y. . (2024). Risk features and predictive model for nosocomial infections by extensively drug-resistant Acinetobacter baumannii. Front. Cell Infect. Microbiol. 14. doi: 10.3389/fcimb.2024.1475428, PMID: 39403207 Sun Y. He Z. Ren J. Wu Y. (2023). Prediction model of in-hospital mortality in intensive care unit patients with cardiac arrest: a retrospective analysis of MIMIC -IV database based on machine learning. BMC Anesthesiol 23, 178. doi: 10.1186/s12871-023-02138-5, PMID: 37231340 Taneja I. Reddy B. Damhorst G. Dave Zhao S. Hassan U. Price Z. . (2017). Combining biomarkers with EMR data to identify patients in different phases of sepsis. Sci. Rep. 7, 10800. Timsit J. F. Ruppé E. Barbier F. Tabah A. Bassetti M. (2020). Bloodstream infections in critically ill patients: an expert statement. Intensive Care Med. 46, 266–284. doi: 10.1007/s00134-020-05950-6, PMID: 32047941 Trottier V. Namias N. Pust D. G. Nuwayhid Z. Manning R. Marttos A. C. . (2007). Outcomes of Acinetobacter baumannii infection in critically ill surgical patients. Surg. Infect. (Larchmt) 8, 437–443. doi: 10.1089/sur.2006.029, PMID: 17883360 Tseng P. Y. Chen Y. T. Wang C. H. Chiu K. M. Peng Y. S. Hsu S. P. . (2020). Prediction of the development of acute kidney injury following cardiac surgery by machine learning. Crit. Care 24, 478. doi: 10.1186/s13054-020-03179-9, PMID: 32736589 Wei Z. Zhao L. Yan J. Wang X. Li Q. Ji Y. . (2024). Dynamic monitoring of neutrophil/lymphocyte ratio, APACHE II score, and SOFA score predict prognosis and drug resistance in patients with Acinetobacter baumannii-calcoaceticus complex bloodstream infection: a single-center retrospective study. Front. Microbiol. 15. doi: 10.3389/fmicb.2024.1296059, PMID: 38322313 Wisplinghoff H. Edmond M. B. Pfaller M. A. Jones R. N. Wenzel R. P. Seifert H. . (2000). Nosocomial bloodstream infections caused by Acinetobacter species in United States hospitals: clinical features, molecular epidemiology, and antimicrobial susceptibility. Clin. Infect. Dis. 31, 690–697. doi: 10.1086/314040, PMID: 11017817 Xu J. Chen X. Zheng X. (2022). Acinetobacter baumannii complex-caused bloodstream infection in ICU during a 12-year period: Predicting fulminant sepsis by interpretable machine learning. Front. Microbiol. 13. doi: 10.3389/fmicb.2022.1037735, PMID: 36386651 Yang J. Zhang B. Hu C. Jiang X. Shui P. Jiajie Huang J. . (2024). Identification of clinical subphenotypes of sepsis after laparoscopic surgery. Laparosc Endosc Robotic Surg. 7, 1626. doi: 10.1016/j.lers.2024.02.001 Yang J. Wang L. Chen L. Zhou P. Yang S. Shen H. . (2025). A comprehensive step-by-step approach for the implementation of target trial emulation: Evaluating fluid resuscitation strategies in post-laparoscopic septic shock as an example. Laparoendosc. Robot. Surg. 8, 28–44. doi: 10.1016/j.lers.2025.01.001 Yang M. Li L. J. Su N. Lin J. Wang J. (2015). Dynamic monitoring of the neutrophil/lymphocyte ratio could predict the prognosis of patients with bloodstream infection. Chin. J. Clin. Infect. Dis. 27, 471476. doi: 10.3760/cma.j.issn.20954352.2015.06.011, PMID: 26049186 Yehya A. Ezzeddine Z. Chakkour M. Dhaini Z. Bou Saba M. S. Bou Saba A. S. . (2025). The intricacies of Acinetobacter baumannii: a multifaceted comprehensive review of a multidrug-resistant pathogen and its clinical significance and implications. Front. Microbiol. 16. doi: 10.3389/fmicb.2025.1565965, PMID: 40444001 Zhou Z. Liu S. Qu F. Wei Y. Song M. Guan X. (2025). Development and validation of a clinical prediction model for pneumonia - associated bloodstream infections. Front. Cell Infect. Microbiol. 15. doi: 10.3389/fcimb.2025.1531732, PMID: 40171157 Zilberberg M. D. Nathanson B. H. Sulham K. Fan W. Shorr A. F. (2016). Multidrug resistance, inappropriate empiric therapy, and hospital mortality in Acinetobacter baumannii pneumonia and sepsis. Crit. Care 20, 221. doi: 10.1186/s13054-016-1392-4, PMID: 27417949 Zoabi Y. Kehat O. Lahav D. Weiss-Meilik A. Adler A. Nea S. (2021). Predicting bloodstream infection outcome using machine learning. Sci. Rep. 11, 20101. doi: 10.1038/s41598-021-99105-2, PMID: 34635696

Edited by: Ilknur Erdem, Namik Kemal University, Türkiye

Reviewed by: Zhongheng Zhang, Sir Run Run Shaw Hospital, China

Polyxeni Karakosta, University General Hospital Attikon, Greece

AB, Acinetobacter baumannii; APACHE II, The Acute Physiology and Chronic Health Evaluation II; CAR,C -reactive protein and albumin ratio; NLR, Neutrophil-to-lymphocyte ratio; ML, Machine learning; MPV/PC, Mean platelet volume/platelet count; RDW, Erythrocyte hemoglobin distribution width; sBSI, Secondary bloodstream infections; WBC, White Blood Cell count.