Predictive rule discovery from electronic health records

Abstract

Automated procedures are described for discovering predictive rules from electronic health records. These patient records are structured, but are not collected relative to any targeted labels or study objectives. The learning methods cycle through all features, simulating labels and converting the problem from unlabeled learning to supervised classification and regression. Each feature in turn is processed as a simulated label, and a prediction is made from the remaining features. Using a decision-rule representation for knowledge extraction, machine learning techniques are applied to a large collection of electronic health records. Many rules are readily induced with significant predictive performance. By formulating the rules as queries to a web search engine, and then counting hit frequencies, we show how medical researchers can assess and rank potential for new insight among a collection of empirically strong associations. © 2010 ACM.