Time-Series Bert for Sepsis Detection: Uncovering Patient Trajectories Through Vital Sign Embeddings

Area of Science:

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

Background:

• Sepsis detection remains a challenge, requiring timely and accurate analysis of complex patient data.
• Traditional methods may not fully capture the dynamic physiological changes indicative of sepsis.

Purpose of the Study:

• To adapt the BERT model for analyzing vital sign time-series data for sepsis detection.
• To evaluate the model's ability to identify distinct patient clusters and discriminate between septic and non-septic cases.
• To explore the utility of transformer architectures in extracting meaningful patterns from electronic health records for clinical applications.

Main Methods:

• Utilized the BERT (Bidirectional Encoder Representations from Transformers) model for time-series analysis of vital signs.
• Employed unsupervised learning on model embeddings to identify patient clusters.
• Assessed classification performance using Precision-Recall Area Under Curve (PR AUC) and Receiver Operating Characteristic Area Under Curve (ROC AUC).
• Leveraged the MIMIC-III database for model training and validation.

Main Results:

• BERT embeddings successfully clustered patients, differentiating septic from non-septic cases.
• The model captured physiological complexity, evidenced by diagnosis count distributions.
• Achieved robust classification performance, indicated by high PR AUC and ROC AUC scores.
• Unsupervised analysis revealed patient subgroups with distinct physiological profiles.

Conclusions:

• Transformer-based models, like BERT, are effective for vital sign time-series analysis in sepsis detection.
• The approach enhances the ability to monitor sepsis by extracting complex physiological patterns.
• Identified patient subgroups offer potential for personalized monitoring and treatment strategies.