Faculty of Science, Technology and Medicine
Faculty of Law, Economics and Finance
Faculty of Humanities, Education and Social Sciences
Interdisciplinary Centre for Security, Reliability and Trust
Luxembourg Centre for Systems Biomedicine
Luxembourg Centre for Contemporary and Digital History
Luxembourg Centre for European Law
Luxembourg Centre for Socio-Environmental Systems
This is a hybrid event. Tune in via Webex here.
The Bayesian Z-score: toward more reliable assessment of aortic diameters
The Z-score is a conceptually simple and widely adopted standard for assessing aortic dilatation from echocardiographic measurements. It is routinely used to monitor patient progression and schedule follow-up checks. However, several criticisms have been raised due to the intrinsic limitations of the typically homoscedastic and linear predictive models. In this paper, we reinterpret the Z-score as a quantitative measure of the aleatoric uncertainty affecting aortic diameters after indexing by a limited number of predictive variables. This view reveals an additional, previously overlooked limitation: the presence of epistemic uncertainty, arising from limited or biased reference datasets. When epistemic uncertainty is high, the Z- score becomes unreliable, yet current tools fail to indicate this. We therefore propose a Bayesian reformulation based
on heteroscedastic Gaussian process regression, where diameters and their aleatoric uncertainties are modeled as random variables. In this framework, the Z-score itself is random, and clinicians receive both an expected value and a high density interval quantifying epistemic uncertainty. Trained on a merged dataset of 1,947 healthy subjects, our Bayesian Z-score detects more dilatations in at-risk patients, identifies uncertain cases, and offers a more reliable basis for clinical decision-making.
About the speaker
Paolo Frasconi is a Professor at the University of Florence, where he leads the Artificial Intelligence research group. His research focuses on machine learning for structured data, including graph neural networks, dynamical systems, and computational biology. He is a key contributor to the theory of Recurrent Neural Networks, co-authoring foundational work on the difficulty of learning long-term dependencies and Input-Output HMMs. Prior to his current role, he completed his Ph.D. at the University of Florence and was a Visiting Scholar at MIT.
The Causal Analysis of Biomedical Data Lecture Series is supported by the Luxembourg National Research Fund (FNR) RESCOM Program.
