Host: Institute of Population Health
About the event:
Individual-level medical prediction is an age-old statistical problem. However, recent advancements in measurement platforms, notably the availability of multilevel "omics" data, or the possibility of monitoring the individual via serial measurements of protein biomarkers, present new methodological challenges. The two talks in this workshop, given by experts in the field, discuss burning issues in this area. The talks will be pitched to a statistically literate audience, although also medical researchers exposed to statistical data analysis may benefit from them.
15:00 - 15:45 Bart Mertens (Department of Biostatistics and Bioinformatics, Leiden University Medical Centre, The Netherlands)
TITLE: Augmented prediction in Omics applications - old wine in new bottles?
ABSTRACT: We consider the problem of constructing prediction models in situation where several distinct sets of omics measurements are available on the same patient, typically representing different levels of human biology (multilevel omics data). A specific problem that may be of interest in this context is the question whether it is worthwhile to include a new omics source in addition to existing omics measurements for better prediction. Another, but similar, example would be that of adding a novel set of omics measures to an established set of clinical predictors. In both these situations the problem is posed of assessing the added value in a predictive sense. We discuss the issues involved and propose a general methodological approach to the problem of assessing the added value of novel omics predictors.
15:45 -16:00 Coffee break
16:00 - 16:45 Carlo Berzuini (Centre for Biostatistics, University of Manchester, United Kingdom)
TITLE: Predicting medical outcomes from serial biomarker measurements: should we model the process in reverse time?
ABSTRACT: Repeated biomarker measurements generated from each member of a cohort of healthy individuals can be monitored for an early identification of those individuals in the cohort who are about to develop a disease or clinical event, even before they show clinical symptoms. The prediction is based on an understanding of the dynamic changes (e.g., changes in slope) that the marker exhibits in the imminence of the event. Contrary to traditional screening approaches based on a population-wide "normal" range of marker values, emphasis on the dynamics of the marker trajectory personalises the screening to each individual’s own baseline, for a more sensitive monitoring. The problem here is how to analyse longitudinal marker measurements jointly with event occurrence data. A relatively novel idea in this area, (brought into the spotlight by an imminent paper by P. McCullagh) is to model marker evolution backwards in time, using event occurrence as origin, without necessarily going all the way back to time zero. This may be a good idea if the individual's hazard tends to remain constant for a long time, only to rise if and when changes in the biomarker behaviour announce an imminent event. We shall discuss the approach in relation to (i) simplicity of the modeling, (ii) extension to case-control studies and personalized treatment (iii) computation of the predictive distributions and (iv) role in biomarker discovery. We illustrate the method through an example in early cancer detection.
Please register for this event using Eventbrite