Currently, the diagnosis of SIC predominantly depends on scoring systems such as the new Japanese Association for Acute Medicine (JAAM) score and the ISTH score for disseminated intravascular coagulation (DIC) [11]. While they present several limitations: use limited indicators, overlook complex interactions, and lack early warning capabilities. The use of machine learning algorithms in the medical field is becoming more common due to the fast progress of artificial intelligence technologies. Compared to traditional medical scoring systems, machine learning algorithms offer enhanced efficiency, accuracy, and self-optimization capabilities [12]. This study used RF feature importance scores to select features and developed eight machine learning models. The LightGBM model showed the best performance and stability.

Feature selection is critical for model optimization. This study employed RF importance scores, which assess feature contributions through multiple decision trees, improving score stability and reliability [13]. This approach evaluates each feature independently [14] and demonstrates robustness to noise and outliers [15], mitigating data quality issues. RF importance scoring identifies the most influential features, enhancing interpretability and guiding subsequent feature engineering and model optimization.

LightGBM is an efficient Gradient Boosting Framework algorithm that optimizes decision tree construction, improving training speed and memory efficiency for large-scale datasets [16]. Unlike conventional Level-wise strategies, LightGBM employs Leaf-wise growth for faster convergence and enhanced pattern recognition in complex data [17]. The algorithm demonstrates exceptional efficiency, accuracy, and robustness, with multi-language interface support facilitating model development and deployment. In medical research, LightGBM has been widely applied in target gene screening, disease diagnosis, and complication prediction [16, 18, 19]. However, increased model complexity challenges interpretability. This study addresses this by employing SHAP theory, which quantifies feature importance and visually represents feature impacts on predictions, elucidating underlying determinants and enhancing model interpretability [20].

This study assessed LightGBM robustness through calibration, DCA, and external validation. Calibration evaluates alignment between predicted and observed probabilities [21], revealing reliable predictions particularly at high probability levels. DCA facilitates assessment of clinical usefulness [22], showing the model differentiates patient risk levels to support clinical decision-making. The optimal threshold was 0.575, balancing accuracy and positive case identification. Given SIC imbalance in the test set (69.2% prevalence), we incorporated Brier score (0.142) for probabilistic calibration and reported PR-AUC (0.977) in supplementary materials, providing more informative discrimination than ROC-AUC in imbalanced settings. While delta changes may provide additional prognostic value, 24-hour means were selected to reflect cumulative metabolic and coagulatory burden during the critical early window, ensuring applicability to routine clinical practice where biomarkers are often measured once daily.

To assess model dependence on SIC diagnostic components, we reconstructed LightGBM after excluding INR, PLT, and SOFA. The excluded model achieved ROC-AUC 0.754 and PR-AUC 0.862, with marked SHAP reordering (LAC rising from 3rd to 1 st; BUN, RDW, and HCT entering top ten). This confirms that original model performance reflected genuine pathophysiological associations rather than input-output overlap. We designate the full-parameter version as “SIC-component-informed model” for maximal data utilization, and the excluded version as “SIC-component-excluded model” for studies requiring strict information independence.

The training data spanned 2008–2019, with external validation covering 2022–2024. Despite evolving sepsis treatment strategies, the model maintained AUC of 0.938, indicating that core predictors (INR, PLT, LAC) reflect common pathophysiological mechanisms of SIC rather than immediate therapeutic responses. This pathophysiology-based approach confers temporal generalizability. However, emergence of novel therapies fundamentally altering SIC trajectory may affect performance, requiring continuous monitoring and updates.

The pathophysiology of SIC is intricate, involving numerous biological pathways and molecular mechanisms. Some studies have identified several biomarkers closely associated with the development of SIC, including Neutrophil Extracellular Traps (NETs) [23], Glycosaminoglycans (GAGs) [2], Heparin-Binding Protein (HBP) [24], Extracellular Vesicles (EVs) [25], and inflammasomes [26], among others. However, due to the complexity of detection techniques and higher costs, the widespread application of these biomarkers in clinical practice faces challenges, limiting their potential in early diagnosis and treatment of SIC.

The LightGBM model using routine clinical indicators shows applicability. INR, PLT, and LAC are the top three biomarkers for predicting SIC risk. INR, derived from standardized PT, evaluates coagulability [27] and is a standalone risk factor for sepsis [28]. Sepsis patients with INR > 1.5 have higher 30-day mortality and moderate diagnostic accuracy for septic shock [29]. Elevated INR during sepsis indicates extrinsic pathway involvement [30] and worsened coagulation dysfunction, crucial for SIC diagnosis [3]. PLT plays a pivotal role in SIC; pathogens and inflammatory mediators activate PLT [31], triggering coagulopathy via the coagulation cascade [32]. PLT reduction occurs in up to 59.5% of SIC cases, with 54% experiencing severe decline [33]. The magnitude of PLT reduction indicates coagulopathy severity [34] and is bidirectionally linked to inflammation [35, 36], with reduced PLT elevating mortality risk [33]. During SIC progression, microthrombi formation causes organ hypoxia [37], elevating LAC levels. While LAC’s direct mechanism in aggravating coagulopathy remains unclear, it serves as a marker of tissue hypoperfusion and organ failure, reflecting systemic oxygen imbalance and metabolic stress. This explains its high predictive weight despite unclear direct coagulation links. In severe sepsis and septic shock, increased LAC clearance correlates with improved coagulation outcomes [38], and fibrinolytic suppression is independently linked to hyperlactatemia in sepsis with DIC [39]. Histone lactylation discovery in 2019 offers a novel perspective on LAC-coagulopathy interplay [40]; LAC directly impacts lactylation extent [41], potentially aggravating sepsis and increasing mortality [42]. This study underscores the importance of INR, PLT, and LAC in SIC pathophysiology, though precise mechanisms warrant further investigation.

This study presents several advantages: (1) it uses routine clinical data for easy implementation; (2) LightGBM demonstrated the optimal performance among multiple machine learning algorithms, providing clear individualized feature interpretation through SHAP analysis to assist clinical decision-making; and (3) external validation confirmed the model’s applicability. Limitations include: (1) the dataset spans a relatively long period (2008–2024) during which sepsis treatment protocols evolved; (2) novel infection and coagulation markers were not incorporated, potentially limiting predictive accuracy; (3) the single-center sample with a small external validation cohort requires further validation through larger multi-center prospective studies; (4) although the performance improvement of LightGBM over simpler models such as logistic regression was modest (AUC 0.937 vs. 0.918), this study effectively addressed the interpretability challenges of complex models through SHAP analysis, providing individualized clinical decision support while maintaining high predictive accuracy, and therefore LightGBM remains recommended as the preferred model; (5) the model is not applicable to patients meeting exclusion criteria (e.g., cirrhosis, trauma), and its performance in these populations remains unknown; and (6) mean imputation may underestimate variance but was selected for clinical deployability, with validation in complete data settings warranted in future studies. Additionally, while SMOTE was necessary to maintain high sensitivity while improving specificity, predicted probabilities should be calibrated to local SIC prevalence, and the tool is intended for screening rather than definitive diagnosis. Future research directions include: (1) conducting multi-center prospective studies; (2) incorporating novel biomarkers and exploring real-time risk assessment; (3) evaluating the model’s impact on patient outcomes using clinical endpoints; and (4) developing standardized clinical decision support systems, including visualization interfaces, cost-effectiveness assessments, and model monitoring mechanisms.