A relative similarity based method for interactive patient risk prediction

Abstract

This paper investigates the patient risk prediction problem in the context of active learning with relative similarities. Active learning has been extensively studied and successfully applied to solve real problems. The typical setting of active learning methods is to query absolute questions. In a medical application where the goal is to predict the risk of patients on certain disease using Electronic Health Records (EHR), the absolute questions take the form of “Will this patient suffer from Alzheimer’s later in his/her life?”, or “Are these two patients similar or not?”. Due to the excessive requirements of domain knowledge, such absolute questions are usually difficult to answer, even for experienced medical experts. In addition, the performance of absolute question focused active learning methods is less stable, since incorrect answers often occur which can be detrimental to the risk prediction model. In this paper, alternatively, we focus on designing relative questions that can be easily answered by domain experts. The proposed relative queries take the form of “Is patient A or patient B more similar to patient C?”, which can be answered by medical experts with more confidence. These questions poll relative information as opposed to absolute information, and even can be answered by non-experts in some cases. In this paper we propose an interactive patient risk prediction method, which actively queries medical experts with the relative similarity of patients. We explore our method on both benchmark and real clinic datasets, and make several interesting discoveries including that querying relative similarities is effective in patient risk prediction, and sometimes can even yield better prediction accuracy than asking for absolute questions.