A rapidly personalized in-hospital bloodstream infection prediction model: a multicenter retrospective study

Abstract

Background

Hospital-acquired bloodstream infections (BSIs) are associated with high morbidity and mortality. Early identification of high mortality risk patients is critical for improving outcomes. Existing prediction models, often based on single-center static indicators, lack generalizability and fail to capture dynamic physiological changes. This study aimed to develop a rapid and personalized prognostic model for mortality in patients with suspected BSIs using dynamic laboratory indicators from multicenter data.

Methods

A multicenter retrospective study was conducted using data from the medical information mart for intensive care IV (MIMIC-IV) database and the second hospital of Anhui medical (AMU2). Adults with suspected hospital-acquired BSIs were included. Demographic data, the Charlson comorbidity index, and serial laboratory values were extracted. The maximum variation rate (MVR) was used to quantify dynamic changes in laboratory parameters. Propensity score matching (PSM) and inverse probability weighting (IPW) were applied to reduce confounding. An extreme gradient boosting (XGBoost) model was developed to predict 28-day mortality, interpreted using shapely additive explanations (SHAP) values, and validated via time-dependent receiver operating characteristic (ROC) curves.

Results

Six MVR indicators were selected from among 24,606 patients: bilirubin, creatinine, C-reactive protein (CRP), hemoglobin, international normalized ratio and platelet count. The model achieved an area under the curve (AUC) of 0.74 in the internal validation cohort AMU2 and 0.71 in the external validation cohort MIMIC-IV. SHAP analysis revealed that elevated creatinine and CRP levels, in addition to reduced hemoglobin, platelet and white blood cell counts (WBC), were associated with an increased mortality risk.

Conclusions

A dynamic model based on routinely available laboratory trends effectively predicts mortality risk in patients with suspected hospital-acquired BSIs. Monitoring temporal changes in laboratory values offers superior predictive value over static measurements, supporting early clinical intervention and prognostic assessment.

Clinical trial number

Not applicable.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12879-026-13227-4.

Introduction

Bloodstream infection is a serious infectious disease that can affect hospitalized patients with significant symptoms and a poor prognosis. The pathogenesis of this condition is characterized by the invasion and multiplication of pathogenic microorganisms into the blood circulation. This often results in a rapid systemic inflammatory response that can progress to sepsis, infectious shock and multiorgan failure. This condition is associated with high morbidity and mortality rates, as well as a significant healthcare economic burden due to prolonged hospital stays, increased need for intensive care and the use of high-level antimicrobial drugs [1–4]. Among these, the situation of nosocomial-acquired bloodstream infections (hospital-acquired BSIs) are of particular critical concern, as causative organisms (e.g., methicillin-resistant Staphylococcus aureus, carbapenem-resistant Enterobacteriaceae, etc.) [5, 6] tend to exhibit higher drug resistance, resulting in limited empirical and targeted treatment options, poor efficacy, and worse clinical outcomes [7, 8]. Consequently, in clinical practice, the ability to promptly and accurately identify such high-risk patients and implement timely and effective interventions has become a pivotal component and a core challenge in enhancing patient prognosis and reducing overall morbidity and mortality.

At present, the clinical diagnosis of bloodstream infections relies heavily on the traditional “gold standard” of growth in blood cultures [9]. However, this method has notable limitations. First, this process is time-consuming and usually takes 24 to 72 h or more from sample delivery to final identification and return of drug sensitivity results. Second, its sensitivity easily interferes with a variety of factors, such as the timing of blood collection, the volume of blood collected, and the patient’s existing use of antimicrobials, which may lead to false-negative results [10–12]. This “diagnostic vacuum” forces clinicians to use empirical drugs in the absence of pathogenic evidence, which can easily lead to inappropriate selection of antimicrobials (over- or under selected) or delayed initiation of therapy, thus missing a critical therapeutic window and ultimately affecting the patient’s prognosis. Consequently, the development of a predictive tool capable of accurately identifying high-risk patients at the early stage of infection, even before the onset of bacteremia, is urgently needed. The purpose of such a tool is to achieve early warning and targeted intervention and to improve clinical outcomes.

In recent years, the widespread popularity of electronic medical records systems and significant advances in machine learning technology have led to a surge of interest in the use of clinical big data for constructing early warning models for diseases[13, 14]. In the domain of bloodstream infections, numerous studies have been conducted that have led to the development of various risk prediction models. These models are based on algorithms such as logistic regression, random forest, or gradient boosting machines, and they integrate patients’ vital signs, laboratory indicators, and basic characteristics [15, 16]. Nevertheless, extant studies continue to present limitations. First, the majority of the models were constructed and validated on the basis of data from a single medical center; thus, extensive external validation is lacking. This has resulted in doubts being cast regarding their generalization ability and extrapolation performance in different patient groups and healthcare environments [17]. Second, the majority of models are static, performing risk assessment at a single point in time [18]. This failure to incorporate and utilize dynamic time series information on the evolution of a patient’s condition during hospitalization results in difficulties in achieving continuous risk monitoring and early warning.

The objective of this study was to develop and validate a simple, rapid, and personalized prognostic model for predicting 28-day mortality among hospitalized patients with suspected nosocomial bloodstream infection, based on dynamic changes in routine laboratory indicators. This model was based on real-world data from multiple patients, incorporating common clinical laboratory indicators and their dynamic changes. Furthermore, the interactions among relevant indicators were investigated.

Methods

Description of the data

We conducted a retrospective exploratory study of patients with suspected nosocomial bloodstream infections and utilized data from two centers to increase the generalizability of the findings. The two-center data were sourced from the Medical Information Mart for Intensive Care-IV (MIMIC-IV, version 3.1; 2008–2022) and the Second Affiliated Hospital of Anhui Medical University (AMU2; 2009–2024) [19]. All eligible patients during these periods were included without sampling, following STROBE guidelines for retrospective cohort studies [20]. The study included patients with suspected bloodstream infections during hospitalization. The inclusion criteria for suspected nosocomial bloodstream infections were as follows: (1) Blood cultures were obtained during hospitalization, with blood samples collected ≥ 48 h after admission. (2) Age > 18 years. To ensure consistency in baseline data, we excluded patients with multiple hospitalizations. The time of the first blood culture order was defined as the index time point, representing the onset of clinical suspicion. This anchor was chosen because clinical management must often begin at suspicion rather than confirmation—consistent with the low yield of blood cultures and with sepsis guidelines emphasizing early intervention upon suspicion [21]. In cases where blood cultures were performed multiple times during hospitalization, we used the earliest order date to mark the clinical suspicion of infection for a given patient.

Variables and outcomes

To better evaluate the effectiveness, we included multiple clinical feature variables—including demographics, comorbidity indices, and laboratory indicators, and adopted different extraction strategies based on the characteristics of each data type. Demographic variables (age and sex) were extracted from the initial admission record. For the comorbidity index, we used the Charlson comorbidity index to represent each patient’s overall health status. All laboratory parameters collected during hospitalization were extracted, and we considered the laboratory parameters obtained within 24 h of the first blood culture as the patient’s baseline physiological status. The primary outcome variable in this study was 28-day mortality, which was treated as a binary variable. Additional file 1 describes the extraction direction and a detailed description of all the variables.

Data preprocessing

We assessed both datasets for missing data and determined that the Second Hospital of Anhui Medical (AMU2) dataset had a higher level of completeness in terms of modeling, derivation and characterization. Variables with more than 50% missing data were excluded, and for variables with less than 50% missing data, multiple interpolations were performed to interpolate the adjusted variables to mitigate the effects of missing data. Analyses of survival/death distributions in the databases revealed significant disparities. To ensure that the death group dataset was not disproportionately influenced by the survival group dataset and to enhance the comparability of clinical outcomes, we performed propensity score matching (PSM) on the baseline information of the original dataset to control for confounding bias and used inverse probability weighting (IPW) to mitigate selection bias. The matching ratio was set at 1:2, and separate cohorts of patients in the death and survival groups were constructed to ensure balanced baseline characteristics at admission. This approach was also employed for baseline correction in the validation set.

Statistical analysis

In this study, numerical variables are presented as medians and interquartile ranges (IQR), and comparisons were performed via the Mann‒Whitney U test. Binary variables are described as frequencies and percentages, and comparisons were performed via the chi-square test or Fisher’s exact test. Through two-dimensional distribution and contour map analysis, we found that mortality rates exhibit temporal characteristics. To further describe and analyze this, we defined the maximum variation rate (MVR) as the amplitude of fluctuation between the baseline value (obtained at blood culture sampling) and the peak absolute change observed in subsequent tests. Spearman’s rank correlation coefficient was used to confirm that the maximum variation rate exhibited a low correlation with baseline values, ensuring the suitability of the included variables as independent variables for modeling. Through a grid search and cross-validation, the optimal number of variables was determined. Subsequently, an extreme gradient boosting (XGBoost) model was used to establish associations between the variables and 28-day mortality. Shapley additive explanations (SHAP) values were employed to describe the dose‒response relationships between each variable and 28-day mortality, with both internal and external validation conducted. This study employed time-dependent ROC curves to evaluate the predictive discriminatory ability of the model in both internal and external validation cohorts. Furthermore, to comprehensively assess the improvement in model performance, the traditional risk score (SOFA) [21] was used as the benchmark model in both the training and validation sets, and comparisons were made using the Net Reclassification Improvement (NRI) metric. A simple flowchart, as shown in Fig. 1, was constructed to illustrate the research process.

Fig. 1.

Research process of this study

All data analysis and graphing were performed via R (version 4.5.1) and Python (version 3.13.5).

Results

Characteristics of the cohort

This retrospective study included a total of 24,605 participants: 12,948 (52.62%) were obtained from the MIMIC-IV database, and 11,657 (47.38%) were obtained from the Second Affiliated Hospital of Anhui Medical University database. Within the Second Affiliated Hospital of Anhui Medical University database, the median patient age was 58 years (IQR: 45, 70), 55.9% of patients were male, and the 28-day mortality rate was 6.3%. The baseline clinical information of the subjects included in the two databases is shown below (Table 1), with no imbalance in age or sex distribution.

Table 1.

Baseline data of AMU2 database

Variables	Overall	Survival	Death	P value
N	11,657	10,918	739
Age (years old)	58 (45, 70)	57 (45, 69)	66 (55, 76)	< 0.001
Male (%)	6518 (55.9)	6037 (55.3)	481 (65.1)	< 0.001
Charlson Comorbidity Score	2 (1, 4)	2 (1, 4)	4 (3, 6)	< 0.001
WBC (×109/L)	9.7 (6.5, 13.8)	9.6 (6.4, 13.5)	11.9 (7.4, 17.3)	< 0.001
Neutrophil count (×109/L)	8.2 (5.2, 11.9)	8.0 (5.1, 11.7)	10.1 (6.2, 15.0)	< 0.001
Lymphocyte count (×109/L)	0.9 (0.6, 1.3)	0.9 (0.6, 1.3)	0.8 (0.5, 1.3)	0.053
Monocyte count (×109/L)	0.6 (0.4, 0.9)	0.6 (0.4, 0.9)	0.7 (0.4, 1.1)	< 0.001
Hemoglobin (g/dL)	10.0 (8.0, 11.8)	10.0 (8.1, 11.8)	9.4 (7.2, 11.8)	< 0.001
Platelet (×109/L)	164 (104, 228)	8.0 (5.1, 11.7)	10.1 (6.2, 15.0)	< 0.001
Total bilirubin (mg/dL)	0.9 (0.6, 1.5)	0.9 (0.6, 1.5)	1.2 (0.7, 2.2)	< 0.001
Alanine aminotransferase (U/L)	32.0 (19.0, 63.0)	31.0 (18.0, 61.0)	41.0 (23.0, 100.0)	< 0.001
Aspartate aminotransferase (U/L)	36.0 (22.0, 71.0)	35.0 (22.0, 66.0)	58.0 (31.0, 147.0)	< 0.001
Serum Creatinine (mg/dL)	0.9 (0.7, 1.3)	0.8 (0.6, 1.2)	1.3 (0.8, 2.3)	< 0.001
Sodium (mmol/L)	139 (136, 143)	139 (136, 142)	141 (136, 147)	< 0.001
Potassium (mmol/L)	3.8 (3.5, 4.2)	3.8 (3.5, 4.2)	3.8 (3.4, 4.2)	0.786
Calcium (mg/dL)	8.0 (7.6, 8.4)	8.0 (7.6, 8.4)	8.0 (7.6, 8.4)	0.836
Prothrombin Time (s)	12.9 (11.8, 14.5)	12.8 (11.8, 14.3)	14.1 (12.5, 17.5)	< 0.001
International Normalized Ratio	1.1 (1.0, 1.2)	1.1 (1.0, 1.2)	1.2 (1.1, 1.5)	< 0.001
C-reactive protein (mg/L)	71.2 (25.0, 142.2)	68.4 (23.6, 136.6)	109.4 (46.6, 196.5)	< 0.001
Procalcitonin (ng/mL)	0.3 (0.1, 1.3)	0.3 (0.1, 1.1)	1.1 (0.3, 5.3)	< 0.001
pH	7.430(7.390,7.468)	7.434(7.400,7.469)	7.410(7.312,7.460)	0.020
Base Excess (mmol/L‌)	2.0 (0.8, 4.9)	1.9 (0.8, 4.8)	3.3 (1.0, 7.7)	0.230
PaO2 (mmHg)	76.8 (62.6, 99.4)	78.7 (63.1, 99.4)	71.4 (61.5, 98.0)	0.198
PaCO2 (mmHg)	34.7 (28.9, 40.4)	34.1 (28.9, 40.0)	35.5 (28.6, 44.0)	0.309
Lactate (mmol/L)	2.0 (1.5, 2.9)	2.0 (1.5, 2.7)	2.3 (1.6, 4.3)	0.002

WBC, white blood cell count; PaO₂, arterial oxygen partial pressure; PaCO₂, arterial carbon dioxide partial pressure

Handling of missing values

A higher degree of completeness was achieved in modeling, derivation, and characterization (see Additional file 2) while retaining laboratory metrics with a missing data rate of less than 0.5 — including total bilirubin, alanine aminotransferase, creatinine, glutamine aminotransferase, C-reactive protein (CRP), serum sodium, serum potassium, leukocytes, hemoglobin, and platelets. Given the varying degrees of missing data across the two databases, alignment with the foundational principles of preceding studies [22–24] and the objectives of efficiency, simplicity and personalization in this study, we selected the following laboratory indices: white blood cell count (WBC), hemoglobin, platelet count, CRP, total bilirubin, and creatinine. With respect to demographic data, only sex and the Charlson comorbidity score of patients were collected since the Charlson comorbidity score already incorporates the age score.

Analyses of the survival/death distributions across the databases revealed significant differences. We therefore performed PSM on the baseline characteristics of the original dataset to control for confounding bias, and inverse IPW to mitigate selection bias, using a matching ratio of 1:2 (death: survival), which ultimately yielded 1,377 matched cases in the AMU2 database and 1401 matched samples in the MIMIC-IV database. Comparisons of the standardized mean differences (SMD) before and after matching across both databases showed a significant reduction in SMD post-matching (see Additional file 3). The same approach was also applied for baseline correction in the validation cohort.

Analysis of the relationships among the variables

On the basis of the two-dimensional distribution plot (see Additional file 4) of a single variable, we found that the variable exhibited certain temporal characteristics. To more clearly illustrate the relationship between the two, we defined the amplitude of fluctuation between the baseline value at the time of blood culture sampling and the peak of absolute change in subsequent tests as the maximum variation rate. By evaluating the correlation between the variables, we calculate the Spearman correlation coefficient (ρ) values, as shown in Additional file 5, which are ≤ 0.4 and prove that the correlation between the variables is weak. An investigation was conducted to explore the two-dimensional relationships between the baseline values of the variables and their MVRs. The results of this investigation revealed that the risk of death was greater when baseline platelet values were higher and declined faster, when baseline leukocyte values were higher and increased faster, and when baseline CRP values were high and increased faster.

Model establishment and validation

The independent variables were screened by repeating the tenfold cross-validation process five times. The analysis revealed that although the while model performance improved with an increasing number of variables, the rate of improvement slowed significantly once the number of variables reached six (see Additional file 6). Given the study’s objective of developing a forecasting tool that is both rapid and precise, we optimized the model by selecting the first six variables—a strategy aimed at striking an optimal balance between performance and simplicity (Fig. 2). We utilized SHAP values [25]to analyze the predictive contribution of each variable, identifying the top six variables in order of importance: creatinine MVR, bilirubin MVR, platelet MVR, CRP MVR, leukocyte MVR, and creatinine baseline value. The final model was based on these six variables. In this study, the XGBoost algorithm [26] was employed to construct a prediction model between the inclusion variables and patients’ 28-day mortality. To assess the discriminative performance of the model, ROC curves were plotted, and the area under the curve (AUC) was calculated. The model exhibited satisfactory discrimination in the matched training set of AMU2 database (AUC = 0.81, 95% CI: 0.78–0.83). In the subsequent validation process, the model demonstrated a high degree of predictive ability in both the internal validation set of AMU2 database (AUC = 0.74, 95% CI: 0.68–0.81) and the external validation set of the MIMIC-IV database (AUC = 0.71, 95% CI: 0.68–0.73). These findings indicate that the model possesses both adequate internal validity and generalizability to independent external datasets, as illustrated in Fig. 3. To further evaluate model performance, this study employed the SOFA score as a benchmark and conducted comparative analysis using the Net Receiver Operating Characteristic (NRI) metric. Results demonstrated that in both the training and validation sets, all NRI values were significantly greater than zero (p < 0.001 for all), indicating that the model significantly improved mortality risk prediction compared to using the SOFA score alone (Additional file 7).

Fig. 2.

SHAP value rankings of variable contributions to 28-day mortality. SHAP value is on the x-axis: large positive value—feature contributes strongly to predict infection; large negative value—feature contributes strongly to predict control

Fig. 3.

ROC curves for the XGBoost model in training, validation, and on the MIMIC4 database

Model interpretation

Subsequent to the development of the model, SHAP values were utilized to visualize the dose‒response relationship between each variable and 28-day mortality (Fig. 4). Positive values indicate an increased risk, whereas negative values are associated with protective effects. An elevated MVR of creatinine demonstrated a significant positive correlation with mortality risk, suggesting that dynamic monitoring of renal function deterioration is closely associated with poor prognosis. Total bilirubin MVR demonstrated a distinctive U-shaped association with mortality risk, indicating that both rapid increases and precipitous decreases were tied to higher mortality risk. Furthermore, a sharp decline in platelet count emerged as a potent predictor of mortality risk. Conversely, a significant increase in CRP, a reliable indicator of the systemic inflammatory response, unmistakably signals an increased mortality risk.

Fig. 4.

Relationship between variables and 28-day mortality based on SHAP values. (A) Effect of creatine MVR on 28-day mortality. (B) Effect of total bilirubin MVR on 28-day mortality. (C) Effect of CRP MVR on 28-day mortality. (D) Effect of platelet MVR on 28-day mortality

Discussion

A novel model was developed to predict the 28-day risk of death in patients with early clinical suspicion of nosocomial bloodstream infection. In patients with early clinical suspicion of nosocomial bloodstream infection, our study revealed that while baseline values are clinically meaningful, dynamic changes in indicators—reflecting their temporal characteristics—are also relevant. We therefore defined a new variable: the maximum rate of variation (MVR). The final model comprised six variables: creatinine MVR, bilirubin MVR, platelet MVR, CRP MVR, leukocyte MVR, and creatinine baseline value. All these variables are widely used in clinical practice and easily accessible.

In this exploratory, retrospective quantitative analysis, the association between each variable and patient risk of death was explored in depth. After constructing the prediction model, we used SHAP values to visualize the dose–response relationship between each variable and 28-day mortality. Variable analyses based on SHAP values revealed a nonlinear relationship between each indicator and the 28-day risk of death, as well as the importance of specific patterns of change. Specifically, elevated creatinine MVR was significantly positively correlated with the risk of death, suggesting that dynamic monitoring of renal function deterioration is essential for prognostic assessment [27]. The pathophysiological mechanisms underlying this process are likely closely related to inadequate renal perfusion and direct injury caused by inflammatory mediators resulting from sepsis. Conversely, the total bilirubin MVR exhibited a distinctive “U-shaped” correlation, indicating that both elevated and reduced values were associated with increased risk. The latter may be attributable to elevated baseline bilirubin levels in the patients themselves, with the subsequent precipitous decline potentially indicative of intricate fluctuations in the disease rather than a straightforward improvement [28]. Furthermore, a precipitous decrease in the PLT serves as an effective predictor of mortality risk, a finding that aligns with the prevailing clinical observation that thrombocytopenia frequently signifies critical states, including severe infections and disseminated intravascular coagulation [29]. Conversely, CRP, an acute-phase reactant produced by the liver, is a reliable indicator of systemic inflammation [30]. The dramatic increase in CRP in the present study, a marker of an intense systemic inflammatory response, clearly indicates an increased risk of death.

Subsequent analysis of the two-dimensional interaction effects between the baseline level of each variable and its MVR was conducted via SHAP values, and the corresponding relationships were plotted. Additional file 8, which is based on the SHAP approach, provides a more intuitive and accurate reflection of the variables’ contributions to the model output than do the two-dimensional relationship graphs plotted prior to modeling. This provides an intuitive visualization of the association between the magnitude of change in key variables (MVR) and predicted mortality risk, transforming complex nonlinear relationships into a practical tool for rapid bedside clinical assessment. The combined effect of the baseline level and trend of each variable on the risk of death was as follows: Creatinine: the risk of death was elevated at lower baseline levels and at a faster rate of increased in creatinine. CRP and white blood cells: the risk of death was elevated at higher baseline levels and at a faster rate of increase. It has been demonstrated that the risk of death is increased at higher baseline levels and that rates of decline are faster in relation to those of platelets. This finding underscores the importance of continuous monitoring of the temporal evolution of these laboratory parameters in clinical practice, as it enables comprehensive patient condition evaluation and prognosis prediction.

Based on large-scale multicenter data, this study introduced the dynamic metric of the “maximum rate of change” to comprehensively capture disease progression trends. The model was developed for the purpose of predicting the occurrence of bloodstream infections in hospitalized patients. It has been shown to be both rapid and straightforward in its application and to be personalized to individual patients. Compared with preceding studies, the primary benefit of this model is its comprehensive utilization of readily available routine clinical indicators. Importantly, while ensuring practicality and scalability, the system in question has been engineered to achieve high-efficiency prediction with minimal indicators. This has been accomplished in such a manner as to effectively meet the demand for immediate clinical assessment. With respect to reliability, the model exhibited consistent and robust discriminatory performance across stringent internal validation and independent external validation, thus confirming its excellent robustness and generalizability. To provide further clarification regarding its clinical value, the recommendation is for subsequent studies to conduct comparative analyses with existing mainstream prediction models. In summary, the present research provides a bloodstream infection risk predictive tool that has been validated by multicenter data, offering reliable performance and ease of clinical implementation.

Limitations

The present study was retrospective in design. Despite the use of multiple models and methods to address potential biases, residual confounders and uncontrolled biases may still persist. Additionally, the model was trained and validated solely on data from two large tertiary hospitals and has not yet been validated in broader healthcare settings such as community hospitals or different healthcare systems. We attempted validation using other international cohorts, such as the Amsterdam University Medical Center dataset and the Korean INSPIRE dataset, but were unable to do so due to the requirement for complete time-series laboratory metrics. Second, owing to inherent limitations of the data from the two databases, some data were proportionately missing, and numerous laboratory indicators were not yet included in this study. The reliability of our conclusions could be enhanced through the conduct of randomized controlled clinical trials or further validation with larger sample sizes. Third, although laboratory measurements were standardized within each center, inter-center variability in assay methods or reporting practices may introduce bias and limit direct comparability. In the future, we plan to incorporate more representative cohorts from diverse sources to further validate the model’s generalizability and clinical applicability.

Conclusion

In summary, the present study constructed a simple, rapid and personalized nosocomial bloodstream infection-related mortality risk prediction model on the basis of routinely available clinical indicators and their dynamic changes. This model is applicable to the risk assessment of patients with suspected bloodstream infections in general wards. The present study revealed that focusing on the dynamic trends of laboratory or physiological indicators has greater value and significance in predicting patient death and clinical prognosis than does the independent assessment of single measurements.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

AMU2

The Second Hospital of Anhui Medical

AUC

Area Under the Curve

BSIs

Bloodstream Infections

CRP

C-Reactive Protein

IPW

Inverse Probability Weighting

IQR

Interquartile Range

MIMIC-IV

Medical Information Mart for Intensive Care-IV

MVR

Maximum Variation Rate

PaCO₂

Arterial Carbon Dioxide Partial Pressure

PaO₂

Arterial Oxygen Partial Pressure

PLT

Platelet

PSM

Propensity Score Matching

ROC

Receiver Operating Characteristic

SHAP

Shapely Additive Explanations

SMD

Standardized Mean Differences

WBC

White Blood Cell Count

XGBoost

Extreme Gradient Boosting

Author contributions

MN: Investigation, Writing – original draft, Conceptualization. WL: Visualization, Validation, Writing – original draft, Data curation. ML: Writing – review & editing, Methodology, Validation. SZ: Supervision, Writing – review & editing. XC: Supervision, Writing – review & editing. ZZ: Resources, Writing – review & editing, Conceptualization. YZ: Resources, Writing – review & editing, Conceptualization.

Funding

The author(s) declare financial support was received for the research and/or publication of this article. This work was supported by the Research Fund of the Second Affiliated Hospital of Anhui Medical University.

Data availability

MIMIC is a public database, and all patient information has been anonymized. We obtained permission to use the data center, and permission to use these data has been granted at https://www.citiprogram.org, with record ID 60227244. Deidentification measures were taken to ensure patient privacy, and we downloaded the complete file and obtained authorization for analysis. The original contributions presented in the study are included in the article/Supplementary Material. Further inquiries can be directed to the corresponding authors.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

I give my consent for information about myself/my child or ward/my relative (circle as appropriate) to be published in European Journal of Medical Research.

Authors information

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.

Competing interests

The authors declare no competing interests.