The purpose of this work is to reduce the human experts’ workload for building systematic reviews that are used in evidence-based medicine, by applying machine learning techniques. We propose to use a committee of classifiers to rank biomedical papers based on the predicted relevance to the topic that is under review. Then, we select two subsets of papers: one that represents the top and another that represents the bottom of the ranked list. These subsets are considered zones where abstracts are labeled with high confidence as being relevant or irrelevant to the topic of the review. We explain how these prediction zones are determined and how they constitute a source for workload reduction. Since the experimental results show that prediction performance, achieved on the prediction zones by applying proposed ML techniques, is generally on the same level as the human prediction performance, using this approach can lead to significant workload reduction for the human experts involved in the systematic reviews process.

Joint work with S. Matwin, D. Inkpen, O. Frunza, A. Razavi, M. Sehatkar, and L. Seaward