26 October 2026 - 26 October 2026
2:00PM - 3:00PM
Room MCS2068, Mathematical Sciences & Computer Science Building, Durham University, Upper Mountjoy Campus, Stockton Road, Durham DH1 3LE
Free
Biostatistics Unit Seminar with speaker Dr Brian Tom from the MRC Biostatistics Unit, University of Cambridge on the topic: A generalized Bayes framework for multi-task learning in the context of outcome-guided clustering.
Maths Building in the distance with grass in the foreground
Seminar Title - A generalized Bayes framework for multi-task learning in the context of outcome-guided clustering
Brian Tom in collaboration with Sida Chen, Helene Ruffieux and Sylvia Richardson
MRC Biostatistics Unit, University of Cambridge
Abstract: In precision medicine, there is growing interest in discovering disease endotypes from high-dimensional biomarkers which associate with clinically relevant outcomes or phenotypes. Motivated by an application in knee osteoarthritis (OA) using synovial fluid protein markers, we aimed to identify OA endotypes linked to disease severity (based on Kellgren-Lawrence grade). Bayesian profile regression (Molitor et al. 2010) was adopted to perform this outcome-guided clustering. However, although the cluster solution found was stable across training and validation/test sets, the prediction of disease severity from these clusters was poor in the validation/test sets; thus, limiting its practical use. The poor out-of-sample performance can partly be attributed to the weak influence of the single binary outcome relative to the high dimensional set of continuous biomarkers used for determining the clustering structure. To redress this imbalance, we propose a generalized Bayesian framework for multi-task learning with outcome-guided clustering formulated as a decision problem with separate loss components for the clustering and prediction tasks. This framework offers greater flexibility and is decision aligned to the scientific goal(s). Additionally, it admits standard Bayesian profile regression as a special case. To improve the interpretability of the weighting of the different loss components, we propose a principled standardization that places each task risk on a common scale, thereby separating the specification of relative task priorities from the learning rate used for posterior uncertainty calibration. We discuss how candidate structural solutions are identified and then selected using goal-aligned out-of-sample criteria, and how uncertainty is calibrated for specific inferential targets.
This will be a hybrid event via teams click here to join on the day.
