BOBM: an adaptive deep learning framework for extended-window sepsis prediction with cross-institutional...

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Informatics
• Deep Learning for Healthcare

Background:

• Sepsis is a life-threatening condition requiring early detection, but current systems face challenges due to its complexity and dynamics.
• Diagnostic and therapeutic delays are common in sepsis management, highlighting the need for improved early warning systems.

Purpose of the Study:

• To introduce the Bayes-Optimized Boosting-Mamba Tab Model (BOBM), a novel deep learning algorithm for enhanced sepsis risk prediction.
• To improve the timeliness and accuracy of sepsis detection and intervention through advanced computational methods.

Main Methods:

• Developed a deep learning algorithm (BOBM) with bidirectional optimization and an ensemble framework.
• Incorporated dynamic agent models for computational efficiency and adaptability.
• Optimized model architecture for different temporal phases and used SHapley Additive exPlanations for interpretability.

Main Results:

• Achieved high AUC (0.86-0.95) on the MIMIC-IV dataset across various pre-onset temporal windows.
• Demonstrated superior performance (AUC: 0.978-0.982) on two independent external datasets.
• Validated robust generalizability and adaptability in sepsis risk prediction.

Conclusions:

• The BOBM model significantly extends the potential intervention timeframe for sepsis compared to traditional methods.
• The model offers dynamic monitoring capabilities from 7 to 28 days before clinical suspicion.
• This approach lays the groundwork for developing regional sepsis surveillance systems.