Deep Bayesian Gaussian processes for uncertainty estimation in electronic health records

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Medicine

Background:

• Clinical decision making relies on deep learning, but accurate confidence estimation remains a challenge.
• Current methods like deep Bayesian neural networks and sparse Gaussian processes have limitations in expressiveness and interpretability.
• Existing deep kernel learning models capture uncertainty from latent spaces but ignore raw data uncertainty, hindering interpretability.

Purpose of the Study:

• To develop a comprehensive uncertainty estimation method by integrating deep Bayesian learning and deep kernel learning.
• To improve the reliability and interpretability of deep learning models in clinical settings.
• To address the limitations of existing uncertainty quantification techniques in healthcare.

Main Methods:

• A novel framework merging deep Bayesian learning and deep kernel learning was developed.
• The method was applied to predict heart failure, diabetes, and depression using large-scale electronic medical records.
• Uncertainty estimation was evaluated based on data insufficiency, misclassification identification, and prediction confidence.

Main Results:

• The proposed method demonstrated superior uncertainty capture compared to Gaussian processes and deep Bayesian neural networks.
• It showed comparable generalization performance while effectively indicating data insufficiency and identifying misclassifications.
• The model proved less prone to overconfident predictions, particularly for minority classes in imbalanced datasets.

Conclusions:

• The integrated deep Bayesian and deep kernel learning approach offers a more comprehensive uncertainty estimation for clinical decision support.
• This method enhances model interpretability by leveraging uncertainty information for risk factor analysis.
• The findings suggest a pathway towards more trustworthy and reliable AI in healthcare.