A multimodal embedding model for sepsis data representation.

Area of Science:

• Biomedical Informatics
• Clinical Data Science
• Artificial Intelligence in Medicine

Background:

• Sepsis research faces challenges due to limited labeled data and models focusing solely on tabular inputs.
• Existing models often neglect the rich information present in clinical text, hindering comprehensive patient understanding.

Purpose of the Study:

• To introduce the Sepsis Data Representation Model (SepsisDRM), an innovative embedding model designed for sepsis research.
• To develop a model capable of jointly processing both tabular and textual patient data for enhanced representation.
• To overcome limitations of existing sepsis models by integrating diverse data sources.

Main Methods:

• Developed SepsisDRM, an embedding model trained on a large dataset of 19,526 sepsis patients.
• The model jointly processes tabular data and clinical text to create comprehensive patient representations.
• Evaluated SepsisDRM's generalization across various sepsis-related tasks without task-specific fine-tuning.

Main Results:

• SepsisDRM demonstrated strong generalization capabilities across diverse sepsis-related tasks.
• The model successfully stratified patients into four clinically interpretable phenotypes.
• Achieved high AUC scores for 28-day outcome prediction: 0.92 (retrospective), 0.94 (prospective), and 0.78 (external datasets).

Conclusions:

• SepsisDRM is the first embedding model specifically developed for sepsis research.
• Establishes a novel paradigm for sepsis data analysis by integrating tabular and textual information.
• Offers a promising approach for sepsis studies and potentially other research fields requiring multimodal data integration.