ExpertNet: A Deep Learning Approach to Combined Risk Modeling and Subtyping in Intensive Care Units.

Area of Science:

• Computational biology and bioinformatics
• Artificial intelligence in healthcare
• Clinical informatics

Background:

• Risk models are vital for disease prevention in intensive care units (ICUs).
• Diseases often present with heterogeneous subpopulations (subtypes) requiring tailored risk assessment.
• Existing subtype-aware risk models struggle with degenerate clusters and data scarcity for classifiers.

Purpose of the Study:

• To develop a deep learning model, ExpertNet, for simultaneous clustering and classification.
• To address limitations of existing models, including degenerate clusters and insufficient training data.
• To improve the accuracy and personalization of disease risk prediction in ICUs.

Main Methods:

• Developed ExpertNet, a novel deep learning architecture for joint clustering and classification.
• Incorporated specialized loss terms and network training strategies to overcome training challenges.
• Evaluated ExpertNet on large electronic medical records datasets for predicting sepsis and acute respiratory distress syndrome (ARDS) risk.

Main Results:

• ExpertNet demonstrated superior accuracy in predicting ARDS and sepsis risk compared to state-of-the-art methods.
• Achieved comparable clustering performance to existing baseline models.
• Identified clinically meaningful subtypes with distinct risk factors, validated through knowledge distillation.

Conclusions:

• ExpertNet effectively addresses technical challenges in simultaneous clustering and classification for risk modeling.
• The model provides a foundation for developing advanced subtype-aware risk prediction tools.
• ExpertNet enhances personalized risk assessment and prevention strategies in critical care settings.