Intelligent Prediction Platform for Sepsis Risk Based on Real-Time Dynamic Temporal Features: Design Study.

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Decision Support Systems
• Sepsis Pathophysiology

Background:

• Sepsis development in ICUs is rapid, necessitating early diagnosis and intervention.
• Real-time prediction models are crucial for sepsis management but often lack timeliness and interpretability.
• Existing AI models for sepsis prediction face limitations in real-time performance and clinical transparency.

Purpose of the Study:

• To develop a real-time sepsis prediction model with high timeliness and clinical interpretability.
• To dynamically predict sepsis risk in ICU patients.
• To establish a practical, tailored sepsis prediction platform for clinical use.

Main Methods:

• A retrospective analysis framework incorporating a real-time prediction module and an interpretability module.
• Utilized 3-hour dynamic temporal features from 8 noninvasive physiological indicators (heart rate, respiratory rate, SpO2, MAP, SBP, DBP, temperature, glucose).
• Employed TreeSHAP for model interpretability, linking AI outputs to physiological significance, and integrated into a web-based platform.

Main Results:

• The sepsis prediction model achieved an accuracy of 0.7 and an AUC of 0.76 in the test cohort.
• TreeSHAP effectively visualized feature contributions, enhancing model transparency and anomaly identification.
• The web-based platform improved clinical utility with real-time risk assessment and actionable insights.

Conclusions:

• The developed platform provides real-time, dynamic sepsis risk warnings for ICU patients.
• Supports timely clinical decision-making for critically ill patients.
• Enhances sepsis management through integrated prediction and interpretability.