Early prediction of sepsis-induced coagulopathy in the ICU using interpretable machine learning: a multi-center retrospective cohort study

TL;DR

This study developed an interpretable machine learning model to predict sepsis-induced coagulopathy in ICU patients within seven days of admission.

Contribution

The novel contribution is an interpretable machine learning model for early SIC prediction with validated performance across multiple datasets.

Findings

A LightGBM model with 13 variables achieved an AUROC of 0.885 in internal validation and 0.831 in external validation.

Key predictors included INR, platelet count, SOFA score, lactate, and comorbidities like heart failure and IHD.

The model was deployed as a web-based tool for clinical use.

Abstract

Sepsis-induced coagulopathy (SIC) is a fatal complication in ICU patients, yet early risk prediction remains challenging. This study aimed to develop an interpretable machine learning model for predicting SIC within seven days of ICU admission.

Clinical data for model development were retrieved from the Medical Information Mart for Intensive Care-IV (MIMIC-IV) database. Feature selection was performed using three distinct algorithms: least absolute shrinkage and selection operator (LASSO) regression, random forest recursive feature elimination (RF-RFE), and the Boruta method. Ten machine learning models underwent training employing 5-fold cross-validation on the training subset, with subsequent evaluation on the validation subset encompassing discrimination, calibration, and clinical utility metrics. The optimal model underwent further interpretability analysis through SHapley Additive exPlanations (SHAP) to elucidate variable contributions and their directional effects. External validation was then conducted using the electronic Intensive Care Unit Collaborative Research Database (eICU-CRD). Finally, the best-performing model was implemented as a web-based Shiny application featuring an interactive interface.

Among 10,740 patients in MIMIC-IV, 2,232 (20.78%) developed SIC within 7 days post-ICU admission. A LightGBM model with thirteen variables demonstrated optimal performance, achieving an area under the receiver operating characteristic curve (AUROC) of 0.885 (95% confidence interval (CI): 0.874–0.897) in the internal validation set and 0.831 (95% CI: 0.819–0.843) in the external eICU-CRD cohort. Key predictive variables included Prothrombin Time-International Normalization Ratio (INR), platelet count, Sequential Organ Failure Assessment (SOFA), lactate, systolic blood pressure (SBP), red cell distribution width (RDW), bicarbonate, phosphate, hemoglobin, age, the presence of heart failure (HF), ischemic heart disease (IHD) and the use of continuous renal replacement therapy (CRRT). The model was deployed as a clinician-oriented web application providing an accessible interface (https://shatao.shinyapps.io/Sepsis_Induced_Coagulopathy/).

This model demonstrated strong predictive ability and clinical interpretability, enabling early SIC identification and targeted intervention.

Figures