Meta Analytic Data Integration for Phenotype Prediction: Application to Chronic Fatigue Syndrome

Abstract

Predictive modeling plays key role in providing accurate prognosis and enables us to take a step closer to personalized treatment. We identified two potential sources of human induced biases that can lead to disparate conclusions. We illustrate through a complex phenotype that robust results can still be drawn after accounting for such biases. Often predictive models build based in high dimensional data suffers from the drawback of lack of interpretability. To achieve interpretability in the form of description of the organism level phenomena in term of molecular or cellular level activities, functional and pathway information is often augmented. Functional information can greatly facilitate the interpretation of the results of the predictive model. However an important aspect of (vertical) data augmentation is routinely ignored, that is there could be several stages of analysis where such information could be meaningfully integrated. There is no know criteria to enable us to assess the effect of such augmentation. A novel aspect of the proposed work is in exploring possibilities of stages of analysis where functional information may be incorporated and in assessing the extent to which the ultimate conclusions would differ depending on level of amalgamation. To boost our confidence on the key findings a first level of meta-analysis is done by exploring different levels of data augmentation. This is followed by comparison of predictive models across different definitions of the same phenotype developed by different groups, which is also an extended form of meta-analysis. We have used real life data on a complex phenotype to illustrate the above. The data pertains to Chronic Fatigue Syndrome (CFS) and another novel aspect of the current work is in modeling the underlying continuous symptom measurements for CFS, which is the first for this disease to our knowledge.