Model Learning from Published Aggregated Data

Abstract

In many application domains, particularly in healthcare, an access for individual datapoints is limited, while data aggregated in form of means and standard deviations are widely available. This limitation is a result of many factors, including privacy laws that prevent clinicians and scientists from freely sharing individual patient data, inability to share proprietary business data, and inadequate data collection methods. Consequently, it prevents the use of the traditional machine learning methods for model construction. The problem is especially important if a study involves comparisons of multiple datasets, where each is derived from different open-access publications where data are represented in an aggregated form. This chapter describes the problem of machine learning of models from aggregated data as compared to traditional learning from individual examples. It presents a method of rule induction from such data as well as an application of this method to constructing of the predictive models for diagnosing liver complications of the metabolic syndrome – one of the most common chronic diseases in humans. Other possible applications of the method are also discussed.