Brian Tom (Theme Lead), Jessica Barrett, Oscar Rueda and William Astle
The overall mission of this theme is to better characterize clinical and biological heterogeneity in order to improve understanding, prognosis, prediction, tailoring of treatment and healthcare decisions. We address the analytical challenges around precision medicine and precision health arising from important clinical and healthcare questions. We collaborate in many areas of healthcare, including arthritis, cancer, cardiovascular disease, blood donation and transfusion, multi-morbidity, and neurodegeneration.
We develop and apply statistical and machine learning methods to multi-modal and longitudinal endotypic and phenotypic individual-level data, in addition to environmental and lifestyle risk factors, for subgroup and latent structure identification. Specific areas of focus include clustering, machine-learning techniques for high-dimensional biomarkers, integration of different data types, Bayesian dynamic latent variable processes, modelling within-individual variability in longitudinal biomarkers and mechanistic models of blood cell biology.
Personalized risk and prognostic modelling
We develop personalized multi-modal approaches/tools for timely identification of risk, prognosis and trajectories. Statistical methodology based on multi-state models, dynamic prediction, longitudinal and joint modelling, machine learning (e.g. deep learning) and risk stratification are pursued. Specific areas of interest are quantification of risk uncertainty, scalability, informative observation in electronic health records, multi-state modelling, transportability, interface of statistical/machine learning approaches and Bayesian formulation of landmark approaches.
Decision-making and actionable insight
Methods are being developed for optimal decision rules in complex static and dynamic settings, for optimizing the allocation of donated blood and for building decision-support tools for managing blood stocks. We address causal inference questions around continuous-time interventions, causal estimands, modelling treatment pathways and causal attribution in multi-state models. Other areas we are exploring include use of Value of Information approaches, identification of drug-sensitive patients and matching of patients to avatars based on genomic and transcriptomic profiles.