A time series driven model for early sepsis prediction based on transformer module.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Critical Care Medicine

Background:

• Sepsis is a leading cause of mortality in intensive care units (ICUs).
• Early detection and intervention are crucial for improving patient survival rates.
• Traditional methods often lack the predictive power for timely sepsis diagnosis.

Purpose of the Study:

• To develop and evaluate novel predictive models for early sepsis detection.
• To leverage time-series patient data for enhanced prediction accuracy.
• To compare the efficacy of CNN-Transformer and LSTM-Transformer architectures for sepsis prediction.

Main Methods:

• Collected time-series patient data at 4, 8, and 12 hours prior to sepsis diagnosis.
• Utilized Convolutional Neural Network-Transformer (CNN-Transformer) and Long Short-Term Memory-Transformer (LSTM-Transformer) architectures.
• Employed the SHAP algorithm for model interpretability and feature weight visualization.

Main Results:

• The proposed Transformer-based models demonstrated a significant improvement (approx. 20%) over traditional recurrent neural networks.
• Achieved high average performance metrics: accuracy (0.964), precision (0.956), recall (0.967), and F1 score (0.959).
• Transformer models showed exceptional predictive capability, especially within a 12-hour window before sepsis onset.

Conclusions:

• Transformer-based deep learning models offer a promising approach for early sepsis prediction.
• These models enhance clinical decision-making by providing interpretable and accurate predictions.
• Early prediction facilitates timely intervention, potentially reducing sepsis-related mortality in ICUs.