Much research in software reliability focuses on using software defect reports, mostly number of defects, to construct reliability models to assess software quality. While this traditional approach is useful, it does not fully utilize information in the reports nor provides time-oriented predictions, which are important for resource scheduling and management of software testing. This paper proposes a novel approach that utilizes software defect reports to predict an estimated time required for fixing the defects found during software testing. The proposed approach applies four data mining algorithms that exploit historical qualitative and quantitative defect data for constructing predictive models. We validate the proposed approach in an empirical study using a dataset of defect reports obtained from testing of a release of a large medical system. The paper describes detailed results of our experiments.