Uncertainty Quantification for Clinical Outcome Predictions with (Large) Language Models

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Informatics
• Machine Learning for Healthcare

Background:

• Language models (LMs) show promise for clinical prediction using electronic health records (EHRs).
• High-stakes healthcare applications demand reliable AI predictions, necessitating robust uncertainty quantification.
• Current AI models often lack transparency, posing risks to patient safety and ethical standards.

Purpose of the Study:

• To develop and validate a framework for uncertainty quantification of LMs in EHR tasks.
• To address uncertainty in both white-box (accessible parameters) and black-box (proprietary LMs like GPT-4) settings.
• To enhance the reliability and transparency of AI-driven clinical predictions.

Main Methods:

• Quantified uncertainty in white-box LMs using multi-tasking and ensemble techniques.
• Extended uncertainty quantification to black-box models, including proprietary LMs.
• Validated the framework on longitudinal clinical data from over 6,000 patients across ten prediction tasks.

Main Results:

• Proposed multi-tasking and ensemble methods effectively reduced model uncertainty in EHR tasks.
• Ensembling and multi-task prediction prompts demonstrated uncertainty reduction across various clinical prediction scenarios.
• The framework successfully increased model transparency in both white-box and black-box settings.

Conclusions:

• Uncertainty quantification using ensembling and multi-tasking improves the reliability of LMs for EHRs.
• The developed framework enhances AI transparency and trustworthiness in clinical decision support.
• This work advances the safe and ethical integration of AI in healthcare delivery.