An Interpretable Early Dynamic Sequential Predictor for Sepsis-Induced Coagulopathy Progression in the Real-World...

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Hematology

Background:

• Sepsis-associated coagulation dysfunction significantly elevates mortality rates.
• Managing sepsis-induced coagulopathy (SIC) and disseminated intravascular coagulation (DIC) is critical for patient survival.
• Irregular clinical time-series data poses challenges for AI-driven medical applications.

Purpose of the Study:

• To develop an interpretable, real-time sequential warning model for early detection of SIC and DIC.
• To address the challenge of irregular data in AI medical applications for sepsis.
• To enable earlier clinical management of sepsis-associated coagulation disorders.

Main Methods:

• Developed and evaluated eight machine learning models, including novel algorithms, for predicting SIC and DIC onset.
• Utilized Xi'an Jiaotong University Medical College (XJTUMC) data for model development and Beth Israel Deaconess Medical Center (BIDMC) for verification.
• Annotated 12,154 SIC and 7,878 ISTH overt-DIC labels and employed AUROC for model evaluation.

Main Results:

• The eXtreme Gradient Boosting (XGBoost) model achieved high AUROCs (0.929 for SIC, 0.910 for DIC) predicting events up to 48 hours in advance.
• XGBoost demonstrated peak performance with AUROCs of 0.973 (SIC) and 0.955 (DIC) at 8 hours prior to onset.
• The novel ODE-RNN model provided continuous predictions with AUROCs of 0.962 (SIC) and 0.936 (DIC) at 8 hours earlier.

Conclusions:

• The developed AI model accurately predicts the onset of sepsis-associated SIC and DIC.
• Early prediction, up to 48 hours in advance, significantly expands the window for timely physician intervention.
• This advancement supports improved management strategies and potentially reduces mortality associated with sepsis-related coagulation disorders.