CHEER: Rich Model Helps Poor Model via Knowledge Infusion.

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Machine Learning

Background:

• Deep learning (DL) models excel in healthcare with abundant multi-channel data (rich-data environments).
• Predictive model performance degrades in data-scarce settings (poor-data environments) due to limited feature channels.
• Leveraging knowledge from well-trained models in related rich-data environments is crucial for improving poor-data models.

Purpose of the Study:

• To develop a framework for transferring knowledge from DL models trained on rich data to improve models in poor-data environments.
• To address the challenge of poor predictive performance in data-limited healthcare scenarios.

Main Methods:

• Introduced the CHEER (Cross-domain Healthcare knowledge transfer Enhanced by Enrichment and Representation) framework.
• CHEER summarizes knowledge from rich-data models into transferable representations.
• Incorporated these representations into poor-data models to enhance performance.

Main Results:

• The CHEER framework demonstrated significant performance improvements over baseline methods.
• Achieved macro-F1 score increases ranging from 5.60% to 46.80% on multiple physiological datasets.
• Theoretical analysis and empirical evaluations validated the effectiveness of knowledge infusion.

Conclusions:

• The CHEER framework effectively boosts the performance of DL models in poor-data healthcare environments.
• Knowledge transfer from rich-data models via CHEER is a viable strategy for improving healthcare AI.
• This approach holds promise for applications in remote or resource-limited healthcare settings.