Uncertainty-Aware Pre-Trained Foundation Models for Patient Risk Prediction via Gaussian Process

Area of Science:

• Artificial Intelligence in Healthcare
• Machine Learning for Clinical Decision Support
• Probabilistic Modeling in Medicine

Background:

• Patient risk prediction models are vital for proactive healthcare.
• Foundation models excel at analyzing multimodal patient data for risk prediction.
• Existing foundation models lack uncertainty quantification, limiting clinical trust.

Purpose of the Study:

• To introduce Gaussian Process-based foundation models for uncertainty-aware risk prediction.
• To enable healthcare professionals to make more informed and cautious decisions.
• To develop an architecture-agnostic approach for uncertainty quantification in foundation models.

Main Methods:

• Integration of Gaussian Processes with pre-trained foundation models.
• Development of instance-level uncertainty quantification techniques.
• Evaluation using classical classification metrics and uncertainty assessment.

Main Results:

• Proposed models achieve competitive performance on standard classification tasks.
• Prediction accuracy is significantly higher for low-uncertainty predictions.
• The method successfully quantifies uncertainty at the instance level, validating its awareness.

Conclusions:

• Gaussian Process-based foundation models enhance clinical decision-making through uncertainty quantification.
• Healthcare providers can leverage uncertainty estimates to prioritize investigations and improve patient care.
• This approach offers a principled and flexible way to build more trustworthy AI in medicine.