In order to extract the most powerful predictive models from data collected within the PRONIA study, diverse information sources must be combined. For each subject, neurocognitive, neuroimaging and clinically observed data has been collected that is intended to provided the basis for the development of predictive models for use in individualised diagnosis and prediction. While a number of approaches may be considered in the combination of data from a diverse range of sources, here, we investigate a two-stage learning approach. An initial step produces a single (probabilistic) outcome for each modality and a second step combines these outcomes to generate a final estimate of the target class. Neurocognitive and neuroimaging data, collected as part of PRONIA, were considered as features for prediction of clinically observed global function, measured at the same time-point. Each neurocognitive test was considered as an independent modality, as were each of a range of MRI-based neuroimaging measures. Using the same target-class (a global assessment of function score less than 65), different approaches to model generation were conducted for each modality using repeated, nested, cross-validation in both stages in order ensure robust estimates of generalisation. The framework of the two-stage learning process is described, and initial results are presented for each approach to classifier-learning considered for both the first and second layer of learning outcomes. An exploration of the contribution to the final prediction from each of the input data streams is discussed and the extension of this approach to structured data-fusion and prediction is considered.
