Early Sepsis Detection Using Heterogeneous Structured ICU Data with Explainable Deep Learning

Area of Science:

• Critical Care Medicine
• Biomedical Informatics
• Artificial Intelligence in Healthcare

Background:

• Sepsis is a life-threatening condition requiring early detection in intensive care units (ICUs) for improved patient outcomes.
• Electronic health records (EHRs) contain valuable data for developing predictive models.
• Deep learning offers potential for analyzing complex EHR data for sepsis prediction.

Purpose of the Study:

• To retrospectively evaluate various deep learning architectures for early sepsis prediction.
• To predict sepsis onset up to 6 hours before clinical recognition using structured EHR data.
• To compare the performance of different deep learning models in identifying sepsis.

Main Methods:

• Utilized hourly structured EHR variables (vitals, labs, demographics) for model training.
• Evaluated architectures: CNN, LSTM, GRU, Bi-LSTM, TCN, Transformer, and CNN-ViT.
• Applied median imputation for missing values and class-weighted loss for imbalance; used SHAP and attention for interpretability.

Main Results:

• The hybrid Convolutional Neural Network-Vision Transformer (CNN-ViT) model demonstrated the strongest performance for the minority class.
• CNN-ViT achieved 88.25% accuracy, 0.7480 recall, 0.454 F1-score, and 0.48 AUPRC.
• Performance remained relatively stable under internal distribution shifts, indicating robustness.

Conclusions:

• Combining local feature extraction with temporal and attention-based modeling enhances early sepsis prediction from structured ICU data.
• The CNN-ViT architecture shows significant potential for improving early sepsis detection.
• Further prospective clinical validation is needed to assess real-world deployment effectiveness.