This study received ethical approval from the Ethics Committee of the First Affiliated Hospital of Shandong First Medical University (Approval No. YXLL-KY-2022-024). The requirement for informed consent was granted a waiver due to the retrospective nature and minimal risk of this study.

This retrospective investigation focused on patients discharged between 1 January 2018, and 31 December 2021, who received treatment with cefazolin sodium, cefoperazone/sulbactam sodium, cefminol sodium, amoxicillin/sulbactam sodium, or piperacillin/tazobactam sodium. We utilized clinical data extracted from electronic medical records at the First Affiliated Hospital of Shandong First Medical University. The data for this study was collected from Shandong Provincial Qianfoshan Hospital Healthcare Big Data Platform (SPQHHBDP). The SPQHHBDP integrates multi-source data from hospital information system (HIS), electronic medical records (EMR), laboratory information management system (LIS), picture archiving and communication system (PACS), nursing information system. The encrypted personal identification number was used as a unique identifier to interlink each person’s data information in the above-mentioned database. The patient’s identification number, name, address, and telephone number were encrypted and not accessible to the investigators. Therefore, this study fully adhered to ethical principles and protected patients' privacy.

Data were retrieved from patient medical records via the hospital management systems and healthcare big data platforms. This comprehensive dataset included patient demographics, surgical details, laboratory test results, and diagnostic information, as shown in Supplementary Table S1. For missing data, we employed two strategies: (1) If the proportion of subjects with missing values was less than 5% of the total, those records were removed. (2) For variables with a missing rate below 10%, the missing values were filled using the patient’s information from other hospitalizations within our institution. The dataset was obtained from the hospital’s EHR system, capturing patients from various departments and covering a diverse patient population in terms of age, gender, and clinical diagnoses. To minimize selection bias, random sampling was employed, and bias analysis was conducted to ensure the generalizability of the results.

Inclusion criteria comprised patients with a hospital stay of ≥48 h, aged ≥18 years, and two or more tests for prothrombin time (PT), activated partial thrombin time (APTT), thrombin time (TT), and platelet (PLT) during hospitalization. Exclusion criteria included patients with an unclear route of administration, non-intravenous administration, those receiving heparin, low molecular weight heparin, warfarin, or other anticoagulants, patients lacking baseline parameters or presenting baseline abnormalities, and those with underlying diseases predisposing to coagulation dysfunction.

Baseline values for drug-related coagulation dysfunction (DRCD) definition were determined from the last test before the initial medication use. As shown in Table 1, the baseline abnormalities were defined as TT or PT exceeding the normal upper limit by 3 s or APTT surpassing the normal upper limit by 10 s. Coagulation parameter abnormalities were identified as a 25% increase in APTT, PT, or TT post-medication compared to baseline.

Statistical analyses were performed using R software (version 3.6.3) on the healthcare big data platform at the First Affiliated Hospital of Shandong First Medical University. Data with normal distribution were presented as mean ± standard deviation, analyzed by t-test or analysis of variance. Non-normally distributed data were expressed as median and interquartile range, assessed by the Wilcoxon rank-sum test. Categorical data were represented as counts or percentages and compared using the Chi-square test. Multivariate logistic regression analysis reported odds ratios (ORs) and 95% confidence intervals (95% CIs). A two-sided p-value <0.05 was considered statistically significant.

Five machine learning methods including logistic regression (LR), random forest (RF), gradient boosting machine (GBM), extreme gradient boosting (XGBoost), and light gradient boosting machine (LightGBM), were employed for predictive model development. The advantages of Logistic Regression are strong explanatory power and high computational efficiency; Random Forest has strong anti-overfitting ability and can perform feature importance evaluation; XGBoost and LightGBM are both improved algorithms based on GBM. XGBoost has better performance, while LightGBM has faster training speed. Therefore, we have selected these five algorithms for the construction of the prediction model and selected the optimal model from them. Patient inclusion involved random division into training and validation sets at a 4:1 ratio. Models underwent 5-fold cross-validation (5-CV) on training sets and independent validation on separate validation sets. To prevent overfitting, univariate analysis screened significantly correlated variables (p < 0.05). The features used in the machine learning models were selected based on the results of multivariate logistic regression analysis, where parameters with statistically significant associations (p < 0.05) were included. The optimal penalty coefficient (λ) was determined by ten-fold cross-validation, identifying characteristic variables with predictive value. Model performance was evaluated using area under the subject operating characteristic curve (AUC), accuracy (ACC), sensitivity (SEN), and specificity (SPE). ACC gauged overall correctness, while SEN and SPE assessed positive and negative sample prediction accuracy, respectively.

The study analyzed and modeled a cohort of 9,529 participants. The incidence of coagulation dysfunction in patients receiving cefazolin sodium, cefoperazone/sulbactam sodium, cefminox sodium, amoxicillin/sulbactam sodium, and piperacillin/tazobactam sodium was found to be 2.4% (51/2,161), 5.4% (102/1,900), 1.5% (25/1,626), 5.5% (139/2,535), and 4.8% (61/1,282), respectively.

The study conducted both single-factor (as shown in Supplementary Tables S2–S6) and multi-factor analysis (as shown in Table 2–6) for each β-lactam antibiotic. In the case of cefazolin sodium (CAS), independent factors associated with DRCD included malignant tumors (OR: 4.52, 95% CI 2.000–10.997) and glucocorticoids (OR: 2.172, 95% CI 1.0118–4.716).

For cefoperazone/sulbactam sodium (CSSS), independent factors for DRCD were malignant tumors (OR: 3.932, 95% CI 2.135–7.613), anticoagulants (OR: 2.00, 95% CI 1.110–3.620), antiepileptic drugs (OR: 2.515, 95% CI 1.392–4.596), and diuretics (OR: 1.807, 95% CI 1.024–3.173).

The risk factors for DRCD associated with cefminox sodium (CMS) included daily dose (OR: 0.487, 95% CI 0.258–0.865), anticoagulants (OR: 5.267, 95% CI 1.774–15.458), antiepileptics (OR: 3.213, 95% CI 1.046–10.081), anxiolytics (OR: 3.225, 95% CI 1.089–9.070), and proton pump inhibitors (OR: 0.162, 95% CI 0.034–0.909).

Mezlocillin/sulbactam sodium (MMS) DRCD independent risk factors comprised drug use time (OR: 1.187, 95% CI 1.090–1.291), contrast agents (OR: 5.763, 95% CI 2.206–14.607), and coagulants (OR: 2.159, 95% CI 1.052–4.239).

Piperacillin/tazobactam sodium (PTS) independent risk factors for DRCD were anticoagulants (OR: 4.370, 95% CI 2.537–7.579) and antiepileptic drugs (OR: 5.567, 95% CI 2.297–14.792).

Internal and external verification results for each β-lactam antibiotic revealed optimal AUC and ACC values, as shown in Figures 1–5. For CAS, the AUC was 0.798. CSSS had an AUC of 0.768, CMS 0.919, MMS 0.783, and PTS 0.867. Validation results illustrated the robustness of the prediction models, with key factors impacting DRCD risk elucidated through ROC curve analyses.

For CAS, factors such as daily dose, non-steroidal anti-inflammatory drugs, anxiolytics, antiepileptics, antipsychotics, and antiarrhythmic drugs were identified as highly correlated with DRCD in cefazolin sodium. For CSSS, major factors affecting the DRCD risk model of cefoperazone sodium and sulbactam sodium included surgery, antiepileptic drugs, antidepressants, diuretics, and malignant tumors. Similarly, for CMS, significant factors impacting the occurrence of DRCD in cefminox sodium included non-steroidal anti-inflammatory drugs, daily dose, diuretics, anxiolytics, and hypertension. In MMS, factors such as age, surgery, contrast agent, pneumonia, procoagulants, and others were found to have a considerable impact on the production of DRCD of mezlocillin/sulbactam sodium. For PTS, the occurrence of DRCD in injectable piperacillin/tazobactam was significantly influenced by risk factors such as anxiolytics, non-steroidal anti-inflammatory drugs, antiarrhythmic drugs, pneumonia, anticoagulants, and hypertension.

The predictive models can be integrated into clinical decision support systems to identify patients at high risk of DRCD, enabling timely intervention and personalized medication adjustments, ultimately improving patient outcomes.

Cephalosporin-induced DRCD can be attributed to the molecular structure containing the N-methylthiotetrazolium side (NMTT), directly impacting vitamin K participation in carboxylation reactions and heightening the risk of bleeding (Lipsky, 1983). In contrast, penicillin-related DRCD may result from penicillin binding to platelet surface proteins, rendering them antigenic and triggering an immune response, leading to decreased platelet count and DRCD (Chong et al., 2013). The stronger association between cefminox and piperacillin with coagulation dysfunction may be related to their specific molecular structures that affect clotting pathways. For instance, cefminox has been associated with reduced prothrombin activity, leading to increased risk of coagulation disorders (Wu et al., 2022).

Analyzing patient demographics, age emerges as a contributing factor to DRCD related to β-lactam antibiotics, with a higher incidence in the elderly. Age-related factors such as diminished immunity, organ aging, poor renal excretion, and drug accumulation elevate the risk of adverse reactions. Thus, caution is warranted to ensure medication safety for the elderly (Fu and Perloff, 2022). Furthermore, our findings underscore a positive correlation between total dosage and DRCD risk, emphasizing the importance of adhering to recommended antibiotic doses in clinical applications (Rochoy et al., 2022).

Gender appears to influence DRCD occurrence, particularly evident in a higher likelihood among men receiving cefoperazone injection. These gender-related differences may be linked to body fat composition and hydrophilic drug combinations, suggesting the necessity for personalized clinical guidance (Karamian et al., 2022).

Patients with underlying diseases, particularly those with malignant tumors, exhibited a significantly higher incidence of DRCD. This association might be linked to circulating tumor cells affecting the coagulation system, extending APTT, TT, or PT times, and elevating DRCD risk (Ward et al., 2021). Chronic kidney disease patients, particularly in the study of piperacillin/tazobactam sodium for injection, demonstrated a heightened susceptibility to DRCD, possibly due to alterations in coagulation system interactions under metabolic conditions in renal failure, as suggested by studies on the impact of microparticles on coagulation (Lutz et al., 2014).

Contrary to general expectations, our study revealed higher DRCD incidence in hypertensive patients treated with cefminox sodium for injection and piperacillin for injection. This unexpected finding prompts further investigation to explore potential connections between hypertension and DRCD, especially in the context of diuretic use (Antza et al., 2018; Kearsley and Stocks, 2021).

Analyzing concurrent medications, anticoagulant use was associated with an increased DRCD risk, seemingly contradictory to the expected therapeutic effects. However, this association was likely influenced by the inclusion of patients undergoing surgeries that required anticoagulants. Similarly, non-steroidal anti-inflammatory drugs exhibited a significant impact, affecting COX-1 activity and transiently influencing platelet function (Scharf, 2012).

Inconsistencies arose in the relationship between glucocorticoid use and DRCD, with our study indicating a higher incidence, contrary to previous research (Graversen et al., 2012; Coelho et al., 2015; Isidori et al., 2015). This unexpected outcome suggests the need for a more nuanced exploration of the potential liver-related effects of glucocorticoids on DRCD occurrence.

Surprisingly, the inclusion of proton pump inhibitors appeared to inhibit DRCD occurrence, suggesting a potential protective effect. This aligns with their clinical use in alleviating gastrointestinal bleeding, underscoring the importance of considering the comprehensive medication profile (Lanas et al., 2018).

Examining bacterial associations, Streptococcus and Gram-negative bacteria, including Pseudomonas aeruginosa, Acinetobacter baumannii, and Klebsiella pneumoniae, were linked to increased DRCD risk. The interaction between these bacteria and the coagulation system components was identified as a potential contributor to DRCD (Herwald et al., 2003; Bergmann and Hammerschmidt, 2007).

Surgery emerged as a significant related factor, supporting existing research on postoperative patients being more prone to DRCD (Ranucci, 2015; Bolliger and Tanaka, 2017). The local trauma and mechanical bleeding associated with surgery might contribute to DRCD, necessitating vigilant monitoring during antibiotic administration in surgical settings.

One of the limitations of this study is its retrospective design, which limits the ability to infer causal relationships between β-lactam antibiotic use and coagulation dysfunction. Additionally, since the data were collected from a single hospital, the generalizability of the findings to other populations or healthcare settings may be limited. Future studies should aim to validate the model on external datasets from different regions or healthcare systems to enhance its applicability.