Malte Blattmann

Innovation Center Computer Assisted Surgery (ICCAS), University of Leipzig

Research Associate

PLCO (Learn more about this study)

PLCO-1797

Jan 27, 2025

Towards reliable clinical decision support: Why popular stochastic deep learning approaches fail at epistemic uncertainty estimation

This project aims to explore the potential of stochastic deep learning methods for estimating epistemic uncertainty in critical clinical decision-making tasks, with a specific focus on predicting prostate cancer-related mortality. Personalized medicine increasingly relies on deep learning models to provide automated, data-driven support for complex clinical risk stratification. However, ensuring the reliability of these models when encountering unseen data remains a significant challenge.

The research will evaluate two popular stochastic deep learning approaches — neural network ensembles and mean-field Gaussian weight Bayesian neural networks — to assess their ability to capture and communicate local epistemic uncertainty. These methods have been chosen for their theoretical capacity to estimate uncertainty within the model function itself. Using data from the PLCO cancer screening trial, the project seeks to analyze their performance in producing calibrated probabilities and trustworthy uncertainty estimates.

Furthermore, the project will investigate potential limitations of these approaches in approximating the functional posterior distribution, which can lead to biased or unreliable uncertainty quantification. Based on these findings, the study aims to provide recommendations for future methodologies, including the exploration of emerging stochastic deep learning techniques that incorporate functional prior knowledge to enhance the reliability of uncertainty estimation. This research holds the potential to advance trustworthy AI systems in clinical decision support, fostering better treatment planning and outcomes in personalized medicine.

Aim 1): Evaluate the predictive performance and uncertainty estimation of stochastic deep learning methods in clinical tasks

1.1) Assess the ability of neural network ensembles and mean-field Gaussian weight Bayesian neural networks to predict prostate cancer (PCa)-related mortality with high accuracy and well-calibrated probabilities, while quantifying their epistemic uncertainty.

1.2) Investigate how these methods approximate the functional posterior distribution and identify biases in their uncertainty quantification.

Aim 2): Identify limitations of existing methods and explore potential alternatives for uncertainty quantification

2.1) Analyze why neural network ensembles and Bayesian neural networks struggle to provide reliable measures of epistemic uncertainty, focusing on factors such as model assumptions and optimization challenges.

2.2) Provide an outlook on emerging stochastic deep learning methods, particularly those incorporating functional prior knowledge, and discuss their potential to address identified limitations.