Machine learning-based prognostic model for bloodstream infections in hematological malignancies using Th1/Th2 cytokines.

Introduction

Bloodstream infection (BSI) is a common and severe clinical condition caused by the invasion and proliferation of various pathogens in the bloodstream. These microorganisms release toxins and metabolic products, inducing cytokine release and triggering systemic inflammatory responses that can progress to multi-organ dysfunction syndrome or death in severe cases [ 1 ]. Advances in chemotherapy, targeted therapies, hematopoietic stem cell transplantation, and supportive care have significantly improved the prognosis of patients with hematologic malignancies(HMs). However, due to the underlying characteristics of HMs, chemotherapy-induced neutropenia, and immunosuppression from treatments, the immune system is often compromised, leading to an increased susceptibility to BSIs, with an incidence reported at 20.8-24.1% [ 2 – 8 ]. In addition, patients with HMs are particularly prone to polymicrobial or recurrent BSIs. BSIs in these patients can exacerbate their condition, disrupt or delay chemotherapy, and significantly increase mortality [ 5 , 9 ].

BSI is a major challenge in the treatment of patients with HMs, particularly in the context of increasing antibiotic resistance. Early identification and effective control of BSIs are crucial for improving the prognosis and survival of patients. In addition to studying pathogen distribution patterns and antibiotic resistance to guide clinicians in choosing appropriate antibiotics, identifying risk factors that affect the prognosis of BSI patients and developing predictive models can aid in early intervention and optimize patient management, ultimately reducing mortality and improving outcomes. Current research has identified several risk factors for poor BSI prognosis in HM patients, including age > 60 years, relapsed or refractory malignancies, nosocomial infections, prolonged or severe neutropenia, inappropriate empirical antibiotic therapy, albumin < 30 g/L, septic shock, multidrug-resistant (MDR) bacteria, central venous catheter placement, urinary catheterization, and thrombocytopenia [ 4 , 7 , 10 – 12 ]. However, predictive models based on these clinical features often yield suboptimal results and require further validation.

Cytokines are powerful weapons used by immune cells to combat infections. Th1 cytokines, including interleukin-2 (IL-2), interferon-gamma (IFN-γ), and tumor necrosis factor-alpha (TNF-α), and Th2 cytokines, including interleukin-4 (IL-4), interleukin-6 (IL-6), and interleukin-10 (IL-10), are involved in mediating cellular immunity and humoral immune responses, respectively.Th1/Th2 cytokines are typically balanced, but this equilibrium can be disrupted by tumors, inflammation, or infections. Studies have shown that increased Th1/Th2 cytokines can provide diagnostic value for infections [ 13 – 20 ]. However, research on the relationship between Th1/Th2 cytokine levels during infection and BSI prognosis in patients with HMs is limited. This gap highlights the need for innovative approaches that combine Th1/Th2 data with clinical features to enhance mortality risk prediction models.

Machine learning (ML), as an emerging technology in medicine, can efficiently handle complex biomedical data and has shown significant advantages in building predictive models. In recent years, ML has made remarkable strides in the field of medicine, revolutionizing disease diagnosis, prognosis prediction, and clinical decision-making. Its applications span various medical domains, including oncology, cardiology, infectious diseases, and personalized medicine. By leveraging large-scale patient data, ML algorithms can enhance diagnostic accuracy, identify disease patterns, predict patient outcomes, and optimize treatment strategies. These innovations not only improve healthcare efficiency but also contribute to more precise and individualized patient care [ 21 – 25 ], helping clinicians identify high-risk patients early, enabling timely intervention and treatment [ 26 , 27 ], ultimately leading to better clinical outcomes. This study focuses on identifying risk factors for 30-day mortality in BSI patients with HMs, constructing predictive models using ML classification methods, and evaluating Th1/Th2 cytokines as potential biomarkers. By integrating these risk factors, we aim to develop an improved 30-day mortality prediction model for BSI through ML, promoting personalized interventions and enhancing patient management precision.

Materials and methods

Source of data

This was a real-word single-center cohort study conducted at the Department of Hematology, Affiliated Hospital of Southwest Medical University. The study was approved by the Ethics Committee of the Affiliated Hospital of Southwest Medical University. The study was retrospectively designed, and the requirement for informed consent was therefore waived. Data of 249 patients with positive blood cultures (bacteria or fungi) in HMs hospitalized from 1 January 2019 to 31 December 2023 were collected from the electronic medical record.

Study patients

Inclusion criteria were as follows: (1) patient diagnosed with HM; (2) positive blood culture; (3) complete clinical data. Exclusion criteria were: (1) patient diagnosed with non-HM; (2) incomplete clinical data; (3) patient with multiple positive blood cultures with the same pathogen during the same hospital stay; they were counted once; (4) patients with Hemophagocytic Syndrome (HLH).

Definitions

Patients with positive blood culture (excluding pollution) who had a fever (> 38 °C) and at least one of the following symptoms: chills, low oxygen saturation, hypotension, cold moist limbs, or altered consciousness, were identified as true BSI [ 28 ]. Polymicrobial bloodstream infection (Polymicrobial BSI) were defined as cases in which two or more distinct pathogens were simultaneously cultured from the venous blood samples based on the above-mentioned symptoms. MDR bacteria were defined as resistant to three or more antibacterial agents [ 29 ]. For common skin contaminants such as Diphtheroids , Bacillus spp , Propionibacterium spp , coagulase-negative staphylococci (CoNS), Viridans Streptococci , Aerococcus spp , and Micrococcus spp [ 30 , 31 ], detection in 2 or more separate blood cultures was required for a definite BSI diagnosis. Repeated sampling from the same patient at the same site was also excluded. Pulmonary infection was diagnosed in patients with acute respiratory symptoms and new onset of pulmonary infiltration on chest computed tomography. Gastrointestinal symptoms included abdominal pain, diarrhea, or lower gastrointestinal bleeding. Perianal infections were defined by perianal pain, abscess, fistula, gangrene or scrotal infection. Absolute neutrophil counts < 0.5 × 10 9 /l were defined as neutropenia.

Study Predictors

Based on previous literature and the expertise of hematology physicians regarding 30-day mortality risk factors post-BSI, relevant factors were collected. While predictive factors with data loss exceeding 30% were excluded (e.g., D dimer and FDP). Clinical variables recorded for risk factor analysis and model development included age, gender, history of hypertension, history of diabetes, pathogens of BSIs, results of drug sensitivity test, state of disease, nosocomial infection, concurrent infection, white blood cell count, neutrophil count, platelet count, albumin level, procalcitonin (PCT), Creatinine (Crea), Alanine aminotransferase (ALT), Aspartate aminotransferase (AST), serum albumin (ALB), Total bilirubin (TB), Direct bilirubin (DB), Indirect bilirubin (IB), Th1/Th2 cytokines and outcomes at 30-day after BSI.

Detection of Th1/Th2 cytokines

Two milliliters of blood were collected in serum separator tubes. After clotting at room temperature or centrifugation at 2000–4000 rpm for 20 min, approximately 0.5 mL of serum was separated and sent for analysis. All samples were processed within 4 h of collection. Cytokine levels, including IL-2, IL-4, IL-6, IL-10, IFN-γ, and TNF-α (pg/mL), were measured using a CytoFLEX flow cytometer.

Model development and evaluation

In order to improve the accuracy and statistical power of the estimators, we conducted univariate and multivariate logistic regression analysis on all patients with complete clinical indicators to screen for 30-day mortality risk factors. Variables with p < 0.05 in the univariate logistic regression analysis were included in the multivariate logistic regression analysis. Final independent clinical risk factors were determined through logistic multivariate regression with P < 0.05. The Least Absolute Shrinkage and Selection Operator (LASSO) method was employed to select parameters with non-zero coefficients.

After selecting characteristic factors from all independent variables, we divided HM patients into training set and test set. Subsequently, seven ML models, including Extreme Gradient Boosting (XGBoost), Logistic Regression (LR), Light Gradient Boosting Machine (LightGBM), RandomForest (RF), Adaptive Boosting (AdaBoost), Gradient Boosting Decision Tree (GBDT) and Gaussian Naïve Bayes (GNB) were developed and refined by applying a 10-fold cross-validation methodology. The performance of these models was assessed using various metrics, including the area under the receiver operating characteristic curve (ROC), calibration, decision and learning curves and confusion matrix. Finally, the SHAP (Shapley Additive Explanations) method was applied to interpret the predicted outcomes and clarify the influence of each feature on the model’s predictions. This approach provides valuable insights, offering clinicians a practical reference for informed decision-making.

Statistical analysis

The continuous variables were expressed as mean ± standard deviation or median (IQR) and compared using t-test or Mann-Whitney U test, while categorical variables were expressed in number and percentage and compared using Chi-square test and Fisher’s exact test. Odds ratios (OR) and 95% confidence intervals (CI) were calculated for 30-day mortality risk factors analysis using the logistic regression. Variables with p < 0.05 in the univariate logistic regression analysis were included in the multivariate logistic regression analysis. The cut-off value for Th1/Th2 cytokines was determined through restricted cubic spline (RCS) regression. Non-linear continuous variables were categorized based on the cut-off values during model construction. All Statistical analyses were performed using R version 4.2.3 and python version 3.11.4.

Results

Patient characteristics

A total of 249 patients with positive blood cultures in HMs were screened, and 231 were finally included in the study based on the exclusion criteria. The overall workflow of the study is shown in Fig. 1 . The cohort consisted of 88 patients with acute myeloid leukemia (AML), 53 with acute lymphocytic leukemia (ALL), 25 with lymphoma, 20 with Myelodysplastic syndromes (MDS), 12 with Multiple Myeloma (MM) and 33 with other HMs (Table 1 ). Table 1 presents the baseline clinical characteristics of HM patients with BSI, divided into survival and death groups. The 30-day mortality in patients with HMs after BSIs was 25.54%. The median age of the patients was 50 years (range 39–61), with 123 (53%) being male, and 108 (47%) being female. The state of 49 patients with HMs was relapsed/refractory. A majority (78%) had concurrent pulmonary infections. Neutropenia existed in 73% of patients. BSIs with MDR bacteria occurred in 105 patients.

Microbiology

Two hundred fifty-one pathogens were isolated from blood culture, including gram-negative bacteria (63.64%), gram-positive bacteria (22.51%), fungi (5.63%) and polymicrobial BSI (8.23%) (Fig. 2 ). The most common gram-negative bacterium was Escherichia coli , accounting for 29.08%, followed by Klebsiella pneumoniae (21.12%) and Pseudomonas aeruginosa (7.57%). 20 strains of carbapenem resisant enterbacteriaceae (CRE) were detected in Klebsiella pneumoniae and Escherichia coli . The common gram-positive bacteria were CoNS (7.17%), Staphylococcus aureus (7.17%), and Enterococcus faecium (5.18%). Six strains of methicillin-resistant Staphylococcus aureus (MRSA) and sixteen strains of methicillin-resistant coagulase-negative staphylococcus (MRCNS) were detected in all of Staphylococcus aureus and CoNS. The most common fungus was Candida tropicalis , accounting for about 5.6%. Among the 19 cases of Polymicrobial BSI, 18 involved two pathogens and one involved three, with 11 cases were Escherichia coli combined with gram-negative bacteria, gram-positive bacteria, or fungi (Fig. 2 ).

The results of the drug sensitivity test for the common pathogens are exhibited in Fig. 3 . Escherichia coli had the lowest rate of antibiotic resistance to tigecyoline (2.74%), followed by amikacin (4.11%), meropenem (10.96%) and imipenem (10.96%). The rates of antibiotic resistance of Klebsiella pneumoniae to tigecyoline, amikacin, imipenem, and meropenem were 3.77%, 5.66%, 24.53%, and 24.53%, respectively. The frequencies for ESBL-producing Escherichia coli and Klebsiella pneumoniae strains were 43.84% and 22.64%, respectively. Pseudomonas aeruginosa had the lowest rates of antibiotic resistance to amikacin, levofloxacin and cefoperazone sulbactam, which were 0%, 0%, 5.26%, respectively. MRSA had no antibiotic resistance to linezolid and vancomycin. MRCNS was completely sensitive to linezolid and vancomycin. Enterococcus faecium had no resistance to vancomycin and linezolid. Fungal resistance was relatively high to voricozole (53.85%) and flucozole (46.15%). Meanwhile, no fungus resistant to amphotericin B, caspofungin and micafungin was found in this study.

Th1/Th2 cytokines

We performed a correlation analysis on the six Th1/Th2 cytokines (IL-2, IL-4, IL-6, IL-10, IFN-γ, and TNF-α) based on the normality of the data distribution. The results revealed a positive correlation between IL-6 and IL-10 ( r = 0.74), IL-6 and TNF-α ( r = 0.55), and IL-2 and IL-4 ( r = 0.5), as shown in Supplementary Fig. 1. In addition, in 231 patients, we utilized RCS to ascertain the cut-off values for IL-4, IL-6, IL-10, IFN-γ and TNF-α levels. The distribution of IL-4, IL-10, TNF-α and IFN-γ was found to be non-linear. The risk of 30-day mortality increased when IL-4 levels exceeded 2.492 pg/mL, IFN-γ levels surpassed 7.29 pg/mL, TNF-α exceeded 4.59 pg/mL, and IL-10 levels were higher than 107.46 pg/mL. Conversely, the distribution of IL-6 exhibited linearity, with a heightened risk of mortality when levels exceeded 875 pg/mL (Supplementary Fig. 1).

Clinical risk factors of 30-day mortality

To enhance the accuracy and statistical power of the estimators, univariate and multivariate logistic regression analyses were performed on all patients with complete clinical data to screen the risk factors for 30-day mortality. Univariate logistic regression analysis revealed that factors such as age > 60 years, relapsed/refractory HM, disease status, polymicrobial BSI, IL-4 > 2.492 pg/mL, IL-10 > 107.46 pg/mL, TNF-α > 4.59 pg/mL, IL-6 levels, ALB < 30 g/L, PLT, ALT, AST, TB, and DB were significantly associated with 30-day mortality following BSI (p < 0.05 for all, Table 2 ). Further multivariate logistic regression analysis identified the following independent prognostic factors for 30-day mortality after BSI in patients with HMs: age > 60 years ( p = 0.023, OR = 2.79, 95% CI: 1.15–6.77), disease status ( p = 0.004, OR = 0.17, 95% CI: 0.05–0.58), polymicrobial BSI ( p < 0.001, OR = 14.02, 95% CI: 3.55–55.48), IL-4 > 2.492 pg/mL ( p = 0.008, OR = 3.21, 95% CI: 1.36–7.57), IL-6 ( p = 0.021, OR = 1.00, 95% CI: 1.00–1.00), and AST ( p = 0.003, OR = 1.01, 95% CI: 1.00-1.02) (Table 2 ).

Predictors selection

A total of 231 patients with positive blood cultures in HMs were divided into a training group (161 patients) and a test group (70 patients) following a 7:3 ratio, after addressing data imbalance. Statistical analysis revealed no significant differences between the two groups except for the gender ( p = 0.005) and TB ( p = 0.007) (Table 3 ). LASSO regression is a regularization technique that performs variable selection and adjusts model complexity by incorporating penalty terms into an optimization objective. In this study, LASSO coefficient curves were generated for five non-zero variables (age, polymicrobial BSI, IL-4 levels, IL-6 levels, and AST) at the Lambda.1se value of 0.057, using 10-fold cross-validation (Fig. 4 ).

Comprehensive analysis of classified multi-model

XGBoost, LR, LightGBM, RandomForest, AdaBoost, GBDT, and GNB were trained and evaluated through 10 repetitions. Model performance was assessed using the area under the curve (AUC). The results showed that XGBoost, RandomForest, GBDT, LightGBM, AdaBoost, and LR performed the best in the training set, while Logistic Regression achieved the highest AUC in the validation set (Fig. 5 A-B, Supplementary Tables 1-–2). To further assess the clinical applicability and compare the performance of the seven models, decision curve analysis (DCA) and calibration curves were analyzed. The DCA demonstrated that the LR model provided better clinical utility, suggesting its superior suitability for clinical decision-making (Fig. 5 C). The calibration curves showed that both the Logistic and RF models exhibited higher prediction accuracy, confirming their reliability in forecasting 30-day mortality risk in HM patients with BSI (Fig. 5 D). Ultimately, the LR model was selected as the optimal model due to its robustness and lower risk of overfitting compared to other models.

The logistic regression model

Logistic regression analysis and 10-fold cross validation were performed on the training set. The results showed that the average AUC for the training set was 0.802 (0.716–0.888), the average AUC for the validation set was 0.792 (0.549–0.989), and the AUC of the test set was 0.822 (0.679–0.965) (Fig. 6 A–C). Calibration plots indicated that the 30-day mortality predictions made by the LR model closely aligned with the actual mortality observed (Fig. 6 D). On the test set, the model similarly demonstrates good net returns, especially in the range of threshold probabilities from 0.2 to 1.0, where it maintained a high level of net returns (Fig. 6 E). The confusion matrix results revealed the model’s performance across different datasets. In the training set (Supplementary Fig. 2A), the model correctly identified 92 true negatives and 31 true positives, while misidentifying 26 false positives and 12 false negatives, yielding a true positive rate (sensitivity) of 72.09% and a true negative rate (specificity) of 77.96%. In the test set (Supplementary Fig. 2B), the model correctly identified 47 true negatives and 12 true positives, misidentifying 7 false positives and 4 false negatives, resulting in a true positive rate of 75% and a true negative rate of 87.04%. Given that the performance on the validation set under the AUC index was lower than that of the test set and the ratio was below 10%, the model fitting could be considered successful. The learning curve further indicated strong fitting and high stability between the training and validation sets (Fig. 6 F). These results suggest that the LR model is suitable for the classification task on the dataset.

Model interpretation based on SHAP

To assess the relative significance of the selected factors, we employed SHAP values to illustrate how these factors predicted the 30-day mortality risk following BSI in patients with HMs. Figure 7 A represents the five most important features in the model. For each feature, the attributions of all patients are plotted, with different colored dots representing varying risk values—red dots indicating high risk and blue dots indicating low risk. The results indicate that higher levels of IL-6, AST, and age > 60 years, along with IL-4 > 2.492 pg/ml and polymicrobial BSI, were associated with an increased risk of 30-day mortality following BSI in HM patients. Figure 7 B further illustrates the impact of the top five features, ranked by importance, on the prediction outcomes. The vertical axis represents the individual features in descending order of importance, while the horizontal axis shows the average SHAP values, indicating how each feature influences the model’s predictions. To enhance the model’s interpretability, we conducted a detailed analysis of two representative samples, as shown in Fig. 7 C and D. By visualizing the SHAP values for these specific samples, we were able to discern the impact of each feature on the model’s predictions for these cases, providing a clearer understanding of the model’s decision-making process for individual patients.

Discussion

BSI is a common and significant prognostic factor in patients with HMs, and identifying key prognostic factors is essential for implementing appropriate treatments. This study included 231 patients with BSIs and analyzed pathogen distribution, antibiotic resistance patterns, and factors predictive of BSI outcomes. A logistic regression model was developed to effectively predict 30-day mortality in BSI patients with HMs. Through careful variable selection and comparison of various ML models, this study underscores the importance of key predictive factors, providing valuable tools for clinical decision-making.

The results revealed that AML and ALL were the most common hematologic malignancies associated with BSIs, emphasizing the need for more proactive infection control measures in patients with acute leukemia. Additionally, patients in the newly diagnosed and relapse/refractory stages exhibited higher rates of BSIs and mortality, whereas those in complete or partial remission experienced lower infection rates and mortality. These findings highlight the critical role of controlling the underlying disease in managing BSIs and improving patient prognosis.

After the occurrence of BSI, prompt and appropriate antimicrobial treatment is necessary, with antibiotic selection tailored to the specific pathogen species and its distribution. In this study, gram-negative bacteria remained the predominant pathogens (63.64%), with Escherichia coli (29.25%) being the most common, followed by Klebsiella pneumoniae (21.23%) and Pseudomonas aeruginosa (8.02%), which is consistent with reports from other centers in China [ 8 ]. A retrospective study from 2007 to 2017 reported that Gram-negative bacteremia accounted for 65% of cases, with Escherichia coli as the most frequent pathogen [ 28 ]. Similarly, a study in China found that Gram-negative bacteria accounted for 64.7% of BSIs in patients with HMs, with Klebsiella pneumoniae being the most common, while Gram-positive bacteria accounted for 27.7%, primarily CoNS and fungi made up 7.7% [ 5 ].In our study, Gram-positive bacteria caused 22.51% of BSIs, with CoNS and Staphylococcus aureus being the most common pathogens, consistent with other studies [ 8 , 28 , 32 , 33 ]. Fungal infections accounted for 5.63%, with Candida tropicalis being the predominant pathogen, similar to other reports [ 8 , 32 ].

Polymicrobial BSI were observed in 8.23% of cases, with 94.74% (18/19) involving two pathogens and one case with three pathogens. Among these, 57.89% (11/19) were combinations of Escherichia coli with other Gram-negative bacteria, Gram-positive bacteria, or fungi. Chen et al. reported a 2.3% incidence of polymicrobial BSI in HM patients [ 5 ], while Li et al. found a 2.76% incidence from 2012–2019 [ 34 ]. A prospective multicenter study from Italy reported that 14.4% of bacterial BSIs were polymicrobial [ 3 ].These findings indicate that the prevalence of gram-positive and gram-negative bacteria is consistent across studies, while the incidence of polymicrobial BSI may vary due to differences in hematologic malignancy types, sample sizes, and regional pathogen distribution. Clinicians should tailor antimicrobial agents based on local pathogen trends to effectively manage BSIs.

Early and effective antimicrobial therapy is critical to reducing mortality in HM patients with BSI. However, due to the low positivity rate and delayed results of blood cultures, empirical therapy remains the primary approach in clinical practice. The widespread use of antibiotics has led to the emergence of resistant strains, which are progressively developing increased resistance, transitioning from low to high resistance and from single-drug to multidrug resistance, complicating empirical treatment. The development of resistance in gram-negative bacteria is mainly associated with the production of β-lactamases, which hydrolyze β-lactam antibiotics, modify antibiotic targets, enhance drug efflux pumps, or decrease the permeability of the outer membrane, leading to resistance [ 35 ]. In the context of increasing antibiotic resistance and limited treatment options, understanding local epidemiology and pathogens resistance patterns is essential for selecting the appropriate antibiotics. This study found that common gram-negative bacteria were most sensitive to amikacin, meropenem, imipenem, and piperacillin-tazobactam. Escherichia coli exhibited the highest resistance to ampicillin, cefotaxime, and ceftriaxone, with nearly 50% resistance to levofloxacin and 35.62% to ceftazidime. Klebsiella pneumoniae showed 24.53% resistance to meropenem and imipenem, 35.85% to ceftazidime, and 45.28% to piperacillin-tazobactam. Pseudomonas aeruginosa demonstrated high resistance (89.47%) to ceftriaxone, cefotaxime, and tigecycline, while it remained sensitive to ceftazidime. No resistance to linezolid or vancomycin was detected in gram-positive bacteria; however, MRSA and MRCNS were identified [ 36 ]. Fungal infections, particularly Candida tropicalis , were also prevalent in HM patients, consistent with other reports from China [ 8 , 32 ]. Resistance to voriconazole in fungal infections exceeded 50%, suggesting that alternatives such as amphotericin B, caspofungin, or micafungin should be considered for antifungal treatment. Xiao et al. reported a 40.5% mortality rate from fungal BSIs in HMs, which is significantly higher than that of other common fungal infections [ 37 ]. Thus, understanding fungal distribution and antifungal resistance is crucial for effective treatment.

BSI is a significant cause of mortality in patients with HMs, with a reported 30-day mortality of 20-30% [ 2 , 4 , 5 , 8 ]. In our study, the 30-day mortality was 25.5%. Early identification of prognostic factors and the development of predictive models are crucial for effective management and improving outcomes in HM patients with BSIs. This study explores the predictive factors for BSI outcomes and develops a logistic regression model using ML classification techniques to effectively predict 30-day mortality in these patients. The results identified independent risk factors for 30-day mortality in HM patients, including age > 60 years, polymicrobial BSI, IL-4 > 2.492 pg/ml, and elevated IL-6 and AST levels.

Polymicrobial BSIs, though less common than monomicrobial BSIs, are frequently encountered in clinical practice, particularly among patients with HMs. These patients often experience immune suppression due to both the malignancy and treatments such as chemotherapy, targeted therapy, immunotherapy, or bone marrow transplantation.This immunocompromised state increases susceptibility to infections caused by multiple pathogens, resulting in polymicrobial BSIs that present complex clinical challenges. While few studies have explored the impact of polymicrobial BSIs on the prognosis of HM patients, our study found that polymicrobial infections are a significant risk factor for 30-day mortality. These infections complicate clinical management because treatment decisions must account for the resistance profiles of multiple pathogens. For example, when Escherichia coli combined with Klebsiella pneumoniae , the latter’s resistance to carbapenems may limit treatment options, reducing the efficacy of carbapenem-based therapies. Additionally, combinations of Escherichia coli and gram-positive cocci (e.g., Enterococcus species) may occur, necessitating combination therapy—such as carbapenems paired with vancomycin or linezolid—to ensure adequate coverage of both pathogens. However, this can lead to potential drug interactions, which require careful monitoring. Given the complexity of polymicrobial BSIs, a comprehensive, individualized treatment approach is crucial for improving outcomes and reducing mortality in HM patients.

Th1 and Th2 cells play distinct roles in immune responses: Th1 cells secrete IL-2, IFN-γ, and TNF-α to drive cellular immunity, while Th2 cells mediate humoral immunity by stimulating B cell proliferation and releasing cytokines like IL-4, IL-6, and IL-10. Disruptions in the balance of Th1/Th2 responses can occur in infections, tumors, and autoimmune diseases. Cytokines, due to their critical role in immune responses, have emerged as important diagnostic biomarkers for infections [ 38 ].Studies have demonstrated that IL-6 are useful diagnostic markers for bacteremia in both pediatric and adult hematologic patients [ 14 , 17 , 20 ]. In particular, IL-6 have shown predictive value for severe infections in pediatric hematologic patients [ 20 ].Zhang L et al. [ 18 ] found that levels of IL-6 were elevated in newly diagnosed Non-Hodgkin Lymphoma patients with respiratory infections compared to those without infections. Specifically, IL-6 levels ≥ 102.6 pg/mL were indicative of bacterial lung infection with bacteremia. However, few studies have investigated the prognostic value of Th1/Th2 cytokines in HMs with BSI. IL-4 and IL-6 play critical roles in immune regulation and inflammatory responses, which may influence the prognosis of BSI in HMs. Our study identifies IL-4 and IL-6 as independent risk factors for 30-day mortality following BSI in HM patients. IL-4, a Th2 cytokine, is primarily involved in anti-inflammatory responses and immune modulation, potentially affecting the host’s ability to clear infections. In contrast, IL-6 is a well-established pro-inflammatory cytokine that contributes to the acute-phase response and systemic inflammation, which are closely linked to sepsis severity and poor outcomes in immunocompromised patients [ 15 ]. Furthermore, excessive IL-4 and IL-6 levels can trigger an overactive immune response, causing inflammatory damage and multi-organ dysfunction. Given these findings, routine monitoring of Th1/Th2 cytokines in HM patients with BSI is recommended. Patients with elevated IL-4 and IL-6 levels should be closely monitored for signs of organ dysfunction and receive prompt treatment to prevent complications. Proactive immune modulation may help reduce mortality and improve outcomes.

Our study also highlighted the significant association between elevated AST levels and 30-day mortality following BSI in HM patients. AST, an enzyme found primarily in the liver, heart, and kidneys, is released into the bloodstream when these organs are damaged. In the context of BSI, apoptosis, metabolic disturbances, and mitochondrial dysfunction can cause multi-organ injury. Specifically, BSI-induced damage may impair cardiac function, reduce myocardial contractility, and decrease cardiac output, leading to pulmonary and systemic congestion, as well as insufficient organ perfusion [ 39 ]. Additionally, microcirculatory dysfunction in the kidneys can cause tubular cell dysfunction and apoptosis, contributing to acute kidney injury [ 40 ]. Given the critical role of AST as an indicator of organ dysfunction, routine monitoring of AST levels should be incorporated into the clinical management of HM patients with BSI. This will enable early detection of organ damage and guide appropriate therapeutic interventions to protect organ function during anti-infective therapy. By closely monitoring AST and other relevant biomarkers, clinicians can better manage the risk of multi-organ dysfunction and improve patient outcomes in this high-risk population.

Despite its strengths, our study has certain limitations. First, as a retrospective study, selection bias is an inherent limitation. Patient inclusion was based on pre-existing medical records, which may not fully represent the broader population. To minimize this bias, we applied strict inclusion and exclusion criteria, ensuring that only patients meeting well-defined diagnostic standards were included. Additionally, appropriate imputation techniques were used to handle missing values. Second, our study only included internal validation, which may limit the generalizability of the model. External validation using independent datasets, as well as prospective studies, will be necessary to further refine and confirm our findings. Third, potential heterogeneity within the patient population exists due to the diverse nature of hematologic malignancies, which may influence the model’s performance across different subgroups. Finally, as a single-center study, our results may be subject to regional variations. Multi-center studies will be required to validate our findings and enhance the generalizability of our model.

Conclusion

We identified age > 60 years, polymicrobial BSI, IL-4 > 2.492 pg/mL, IL-6 and AST levels as independent risk factors for 30-day mortality in HMs with BSI. Based on these factors, we developed and validated a logistic regression prediction model with strong predictive accuracy. The model, incorporating easily accessible clinical indicators, aids in the early identification of high-risk patients, facilitating optimized management and timely interventions to reduce disease progression and BSI-related mortality in this patient population.

Electronic Supplementary Material

Below is the link to the electronic supplementary material.

Supplementary Material 1

Acknowledgements

The authors sincerely thank the patients whose valuable data were vital to this research.

Abbreviations

Bloodstream infection

Hematologic malignancies

Xtreme Gradient Boosting

Logistic Regression

Light Gradient Boosting Machine

RandomForest

Adaptive Boosting

Gradient Boosting Decision Tree

Gaussian Naïve Bayes

Receiver operating characteristic curve

Shapley Additive Explanations

Multidrug-resistant

Interleukin-2

Interferon-gamma

Tumor necrosis factor-alpha

Interleukin-4

Interleukin-6

Interleukin-10

Machine learning

Hemophagocytic Syndrome

Coagulase-negative staphylococci

Procalcitonin

Creatinine

Alanine aminotransferase

Aspartate aminotransferase

Serum albumin

Total bilirubin

Direct bilirubin

Indirect bilirubin

Least Absolute Shrinkage and Selection Operator

Confidence intervals

Odds ratios

Restricted cubic spline

Acute myeloid leukemia

Acute lymphocytic leukemia

Myelodysplastic syndromes

Multiple Myeloma

Carbapenem resisant enterbacteriaceae

Methicillin-resistant Staphylococcus aureus

Methicillin-resistant coagulase-negative staphylococcus

Area under the curve

Decision curve analysis

Author contributions

Qin Li and Nan Lin collected cases and wrote the main manuscript. Zuheng Wang and Yuexi Chen analyzed the data. Yuli Xie drew the picture. Xuemei Wang and Jirui Tang collected the data. Yuling Xu and Min Xu interpreted the data. Na Lu, Yiqian Huang, and Jiamin Luo completed the tables. Li Jing and Zhenfang Liu designed the study and critically revised the manuscript. All authors have read and approved the final version of the manuscript.

Funding

This work was supported by the National Natural Science Foundation of China (No. 82160039) and the Research Foundation of Southwest Medical University (2024ZKY027).

Data availability

The datasets generated and/or analyzed are not publicly available owing to ethical and legal causes. Nevertheless, they can be made available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

The research project was approved by the Ethics Committee of the Affiliated Hospital of Southwest Medical University (KY2024541) and conducted in accordance with the Declaration of Helsinki. The Ethics Committee of the Affiliated Hospital of Southwest Medical University approved this study protocol and waived the obligation for informed consent because of the retrospective nature of the study.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.

Clinical trial number

Not applicable.

Footnotes

Publisher’s note

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

Qin Li and Nan Lin contributed equally to this work.

Contributor Information

Zhenfang Liu, Email: liuzhenfang@gxmu.edu.cn.

Li Jing, Email: 2251270990@qq.com.

References

1. Gotts JE, Matthay MA. Sepsis: pathophysiology and clinical management. BMJ. 2016;353:i1585.

2. Åttman E, Aittoniemi J, Sinisalo M, Vuento R, Lyytikäinen O, Kärki T, Syrjänen J, Huttunen R. Etiology, clinical course and outcome of healthcare-associated bloodstream infections in patients with hematological malignancies: a retrospective study of 350 patients in a Finnish tertiary care hospital. Leuk Lymphoma. 2015;56(12):3370–7.

3. Trecarichi EM, Pagano L, Candoni A, Pastore D, Cattaneo C, Fanci R, Nosari A, Caira M, Spadea A, Busca A, et al. Current epidemiology and antimicrobial resistance data for bacterial bloodstream infections in patients with hematologic malignancies: an Italian multicentre prospective survey. Clin Microbiol Infect. 2015;21(4):337–43.

4. Tang Y, Cheng Q, Yang Q, Liu J, Zhang D, Cao W, Liu Q, Zhou T, Zeng H, Zhou L, et al. Prognostic factors and scoring model of hematological malignancies patients with bloodstream infections. Infection. 2018;46(4):513–21.

5. Chen S, Lin K, Li Q, Luo X, Xiao M, Chen M, Zhu H, Chen Y, Wu X, Zeng Y, et al. A practical update on the epidemiology and risk factors for the emergence and mortality of bloodstream infections from real-world data of 3014 hematological malignancy patients receiving chemotherapy. J Cancer. 2021;12(18):5494–505.

6. Okamoto A, Kanda Y, Kimura SI, Oyake T, Tamura K. Predictive and risk factor analysis for bloodstream infection in high-risk hematological patients with febrile neutropenia: post-hoc analysis from a prospective, large-scale clinical study. Int J Hematol. 2021;114(4):472–82.

7. Zhao Y, Lin Q, Zhang T, Zhen S, Wang J, Jiang E, Mi Y, Qiu L, Han M, Wang J, et al. Pseudomonas aeruginosa bloodstream infection in patients with hematological diseases: Clinical outcomes and prediction model of multidrug-resistant infections. J Infect. 2023;86(1):66–117.

8. Wang J, Wang M, Zhao A, Zhou H, Mu M, Liu X, Niu T. Microbiology and prognostic prediction model of bloodstream infection in patients with hematological malignancies. Front Cell Infect Microbiol. 2023;13:1167638.

9. Liao WC, Chung WS, Lo YC, Shih WH, Chou CH, Chen CY, Tu CY, Ho MW. Changing epidemiology and prognosis of nosocomial bloodstream infection: A single-center retrospective study in Taiwan. J Microbiol Immunol Infect. 2022;55(6 Pt 2):1293–300.

10. Tumbarello M, Trecarichi EM, Caira M, Candoni A, Pastore D, Cattaneo C, Fanci R, Nosari A, Spadea A, Busca A, et al. Derivation and validation of a scoring system to identify patients with bacteremia and hematological malignancies at higher risk for mortality. PLoS ONE. 2012;7(12):e51612.

11. Albasanz-Puig A, Gudiol C, Puerta-Alcalde P, Ayaz CM, Machado M, Herrera F, Martín-Dávila P, Laporte-Amargós J, Cardozo C, Akova M, et al. Impact of the Inclusion of an Aminoglycoside to the Initial Empirical Antibiotic Therapy for Gram-Negative Bloodstream Infections in Hematological Neutropenic Patients: a Propensity-Matched Cohort Study (AMINOLACTAM Study). Antimicrob Agents Chemother. 2021;65(8):e0004521.

12. Shi N, Kang J, Wang S, Song Y, Yin D, Li X, Guo Q, Duan J, Zhang S. Bacteriological Profile and Antimicrobial Susceptibility Patterns of Gram-Negative Bloodstream Infection and Risk Factors Associated with Mortality and Drug Resistance: A Retrospective Study from Shanxi, China. Infect Drug Resist. 2022;15:3561–78.

13. Tang Y, Liao C, Xu X, Song H, Shi S, Yang S, Zhao F, Xu W, Chen X, Mao J, et al. Evaluation of Th1/Th2 cytokines as a rapid diagnostic tool for severe infection in paediatric haematology/oncology patients by the use of cytometric bead array technology. Clin Microbiol Infect. 2011;17(11):1666–73.

14. Tang Y, Liao C, Xu X, Song H, Shi S, Yang S. Th1/Th2 cytokine profiles in G+/G- bacteremia in pediatric hematology/oncology patients. Pediatr Blood Cancer. 2012;58(1):50–4.

15. Lenski M, Tonn JC, Siller S. Interleukin-6 as inflammatory marker of surgical site infection following spinal surgery. Acta Neurochir (Wien). 2021;163(6):1583–92.

16. Surbatovic M, Popovic N, Vojvodic D, Milosevic I, Acimovic G, Stojicic M, Veljovic M, Jevdjic J, Djordjevic D, Radakovic S. Cytokine profile in severe Gram-positive and Gram-negative abdominal sepsis. Sci Rep. 2015;5:11355.

17. Xia T, Xu X, Zhao N, Luo Z, Tang Y. Comparison of the diagnostic power of cytokine patterns and procalcitonin for predicting infection among paediatric haematology/oncology patients. Clin Microbiol Infect. 2016;22(12):996–1001.

18. Zhang L, Zhang J, He H, Ling X, Li F, Yang Z, Zhao J, Li H, Yang T, Zhao S, et al. Increased Cytokine Levels Assist in the Diagnosis of Respiratory Bacterial Infections or Concurrent Bacteremia in Patients With Non-Hodgkin’s Lymphoma. Front Cell Infect Microbiol. 2022;12:860526.

19. Zhang WF, Xu Y, Gao QP, Mao HW, Xu B, Qu H. The Levels of Inflammatory Regulatory Factors in Acute Leukemia Complicated by Infection and Its Clinical Significance. Zhongguo Shi Yan Xue Ye Xue Za Zhi. 2022;30(2):435–40.

20. Lin SP, Xu XJ, Liao C, Zhao N, Chen YY, Song H, Xu WQ, Liang J, Shen DY, Zhang JY, et al. The predictive utility of cytokines, procalcitonin and C-reactive protein among febrile pediatric hematology and oncology patients w ith severe sepsis or septic shock. Pediatr Hematol Oncol. 2024;41(1):1–14.

21. Li Y, Wu X, Yang P, Jiang G, Luo Y. Machine Learning for Lung Cancer Diagnosis, Treatment, and Prognosis. Genom Proteom Bioinform. 2022;20(5):850–66.

22. Anh DT, Takakura H, Asai M, Ueda N, Shojaku H. Application of machine learning in the diagnosis of vestibular disease. Sci Rep. 2022;12(1):20805.

23. C PD. A systematic review on machine learning and deep learning techniques in cancer survival prediction. Prog Biophys Mol Biol. 2022;174:62–71.

24. Mohammadzadeh I, Hajikarimloo B, Niroomand B, Eini P, Ghanbarnia R, Habibi MA, Albakr A, Borghei-Razavi H. Application of artificial intelligence in forecasting survival in high-grade glioma: systematic review and meta-analysis involving 79,638 participants. Neurosurg Rev. 2025;48(1):240.

25. Mohammadzadeh I, Niroomand B, Eini P, Khaledian H, Choubineh T, Luzzi S. Leveraging machine learning algorithms to forecast delayed cerebral ischemia following subarachnoid hemorrhage: a systematic review and meta-analysis of 5,115 participants. Neurosurg Rev. 2025;48(1):26.

26. Greener JG, Kandathil SM, Moffat L, Jones DT. A guide to machine learning for biologists. Nat Rev Mol Cell Biol. 2022;23(1):40–55.

27. Kang MW, Kang Y. Utilizing deep learning-based causal inference to explore vancomycin’s impact on continuous kidney replacement therapy necessity in blood culture-positive intensive care unit patients. Microbiol Spectr 2024:e0266224.

28. Amanati A, Sajedianfard S, Khajeh S, Ghasempour S, Mehrangiz S, Nematolahi S, Shahhosein Z. Bloodstream infections in adult patients with malignancy, epidemiology, microbiology, and risk factors associated with mortality and multi-drug resistance. BMC Infect Dis. 2021;21(1):636.

29. Magiorakos AP, Srinivasan A, Carey RB, Carmeli Y, Falagas ME, Giske CG, Harbarth S, Hindler JF, Kahlmeter G, Olsson-Liljequist B, et al. Multidrug-resistant, extensively drug-resistant and pandrug-resistant bacteria: an international expert proposal for interim standard definitions for acquired resistance. Clin Microbiol Infect. 2012;18(3):268–81.

30. van Walraven C, Wong J. Independent influence of negative blood cultures and bloodstream infections on in-hospital mortality. BMC Infect Dis. 2014;14:36.

31. Apostolopoulou E, Raftopoulos V, Terzis K, Elefsiniotis I. Infection Probability Score, APACHE II and KARNOFSKY scoring systems as predictors of bloodstream infection onset in hematology-oncology patients. BMC Infect Dis. 2010;10:135.

32. Cai LJ, Wei XL, Wei YQ, Guo XT, Jiang XJ, Zhang Y, Yu GP, Dai M, Ye JY, Zhou HS, et al. A single-center study on the distribution and antibiotic resistance of pathogens causing bloodstream infection in patients with hematological malignancies. Zhonghua Xue Ye Xue Za Zhi. 2023;44(6):479–83.

33. Di Domenico EG, Marchesi F, Cavallo I, Toma L, Sivori F, Papa E, Spadea A, Cafarella G, Terrenato I, Prignano G, et al. The Impact of Bacterial Biofilms on End-Organ Disease and Mortality in Patients with Hematologic Malignancies Developing a Bloodstream Infection. Microbiol Spectr. 2021;9(1):e0055021.

34. Li M, Du M, Li H, Liu D, Liu Y. Epidemiology, resistant pathogens, and causes of early death in cases of bloodstream infection in patients with hematological malignancies from 2012–2019. Infect Med. 2022;1(1):23–30.

35. Darby EM, Trampari E, Siasat P, Gaya MS, Alav I, Webber MA, Blair JMA. Molecular mechanisms of antibiotic resistance revisited. Nat Rev Microbiol. 2023;21(5):280–95.

36. Haddad S, Jabbour JF, Hindy JR, Makki M, Sabbagh A, Nayfeh M, Boustany M, El-Zein S, Tamim H, Zakhem AE, et al. Bacterial bloodstream infections and patterns of resistance in patients with haematological malignancies at a tertiary centre in Lebanon over 10 years. J global Antimicrob Resist. 2021;27:228–35.

37. Xiao H, Tang Y, Cheng Q, Liu J, Li X. Risk Prediction and Prognosis of Invasive Fungal Disease in Hematological Malignancies Patients Complicated with Bloodstream Infections. Cancer Manage Res. 2020;12:2167–75.

38. Akdis M, Aab A, Altunbulakli C, Azkur K, Costa RA, Crameri R, Duan S, Eiwegger T, Eljaszewicz A, Ferstl R, et al. Interleukins (from IL-1 to IL-38), interferons, transforming growth factor beta, and TNF-alpha: Receptors, functions, and roles in diseases. J Allergy Clin Immunol. 2016;138(4):984–1010.

39. Cimolai MC, Alvarez S, Bode C, Bugger H. Mitochondrial Mechanisms in Septic Cardiomyopathy. Int J Mol Sci. 2015;16(8):17763–78.

40. Dellepiane S, Marengo M, Cantaluppi V. Detrimental cross-talk between sepsis and acute kidney injury: new pathogenic mechanisms, early biomarkers and targeted therapies. Crit Care. 2016;20:61.

Associated Data

Supplementary Materials

Supplementary Material 1

Data Availability Statement

The datasets generated and/or analyzed are not publicly available owing to ethical and legal causes. Nevertheless, they can be made available from the corresponding author on reasonable request.