A Combined Human-Computer Method for Building Effective Decision Support System

Computational tools for building clinical decision support systems have been developed for decades. The effects of interuser variability, i.e. the effects of different interpretations or observations by different human experts, have been typically ignored when these systems were built. This problem particularly affects machine learning methods that rely on data to build their models. In this work we present an alternative approach to build clinical decision support systems dealing with the inter-user variability problem. The method is tailored on single physician and combines its output with him/her to improve overall performance. In this paper, inter-user variability is estimated on a melanoma dataset and performances of the method are presented. The decision support system is here supposed to enhance early diagnosis of the disease. Comparison of our method with other approaches is discussed