Consensus modeling: Safer transfer learning for small health systems.

Area of Science:

• Clinical Informatics
• Machine Learning in Healthcare
• Health Systems Research

Background:

• Predictive models are crucial for clinical decision support.
• Smaller health systems face challenges in developing accurate and safe models due to limited data.
• Existing methods like generic models, research networks, and transfer learning have limitations.

Purpose of the Study:

• To introduce and evaluate the consensus modeling paradigm for improving predictive models in health systems with small sample sizes.
• To compare consensus modeling against generic models, research networks, and transfer learning.

Main Methods:

• The consensus modeling paradigm was developed, utilizing a large source site to aid a smaller target site.
• The approach was evaluated on predicting postoperative complications in two health systems.
• A simulation study assessed performance relative to other methods based on target site sample size.

Main Results:

• Consensus modeling demonstrated the least over-specificity at both source and target sites.
• The approach achieved the highest combined predictive performance.
• Performance was evaluated across varying training sample sizes at the target site.

Conclusions:

• Consensus modeling offers a safer and more effective alternative for developing predictive models in resource-limited health systems.
• This paradigm addresses the limitations of existing approaches, leading to improved clinical decision support tools.