Implicit versus explicit Bayesian priors for epistemic uncertainty estimation in clinical decision support.

Area of Science:

• Artificial Intelligence
• Biomedical Informatics
• Machine Learning

Background:

• Deep learning models show promise for personalized medicine.
• Reliability issues arise with out-of-distribution data and overconfident predictions.
• Quantifying epistemic uncertainty is crucial for trustworthy clinical decision support.

Purpose of the Study:

• Compare approximate Bayesian deep learning methods for uncertainty quantification.
• Evaluate model performance in predicting prostate cancer mortality.
• Identify methods for reliable clinical decision-support tools.

Main Methods:

• Applied three approximate Bayesian deep learning methods to prostate cancer mortality data (PLCO trial).
• Compared implicit functional-prior methods (NN ensembles, VBNNs) with explicit distance-aware priors (SNGP).
• Assessed discriminative performance (AUROC) and calibration of epistemic uncertainty estimates.

Main Results:

• All methods achieved strong performance (AUROC = 0.86) with well-calibrated in-distribution probabilities.
• Implicit functional-prior methods showed reduced fidelity and biased epistemic uncertainty estimates.
• Explicitly distance-aware SNGP models provided more accurate posterior approximations and reliable uncertainty quantification.

Conclusions:

• Explicitly distance-aware Bayesian deep learning architectures offer superior uncertainty quantification.
• These methods are promising for developing trustworthy clinical decision-support systems.
• Accurate uncertainty estimation is key for reliable AI in healthcare.