Dr. Zoran Obradovic is giving an invited lecture at IEEE International Conference on Bioinformatics and Biomedicine 2011. His lecture “Analysis and Integration of Inconsistent and Unreliable Biomedical Prediction Models” is based on joint papers with his PhD. students Mohamed Ghalwash (in press at Molecular Biosystems journal) and Ping Zhang (ECML PKDD 2011). Following is the abstract of this study.
In biomedical applications, multiple predictors are often developed for the same problem using multiple training datasets of various qualities. Selecting a single model from such a collection based on accuracy evaluation on a small biased set of annotated data is not necessarily the best strategy when the objective is large scale application of the model. In this talk we will discuss how to address this problem by uncertainty analysis in the reference models and in data. In addition, we will present an iterative algorithm for integrating predictions of multiple models without relying on any annotated data. The proposed solutions will be illustrated on the problem of predicting intrinsic disorder in proteins that lack a stable tertiary structure but still have important biological functions.