Natural Language Processing of Medical Reports

Abstract

A significant amount of information regarding the observations, assessments, and recommendations related to a patient's case is documented within free-text medical reports. The ability to structure and standardize clinical patient data has been a grand goal of medical informatics since the inception of the field - especially if this structuring can be (automatically) achieved at the patient bedside and within the modus operandi of current medical practice. A computational infrastructure that transforms the process of clinical data collection from an uncontrolled to highly controlled operation (i.e., precise, completely specified, standard representation) can facilitate medical knowledge acquisition and its application to improve healthcare. Medical natural language processing (NLP) systems attempt to interpret free-text to facilitate a clinical, research, or teaching task. An NLP system performs translates a source language (e.g., free-text) to a target surrogate, computer-understandable representation (e.g., first-order logic), which in turn can support the operations of a driving application. NLP is really then a transformation from a representational form that is not very useful from the perspective of a computer (a sequence of characters) to a form that is useful (a logic-based representation of the text meaning). In general, the accuracy and speed of translation is heavily dependent on the end application. This chapter presents work related to natural language processing of clinical reports, covering issues related to representation, computation, and evaluation. We first summarize a number of typical clinical applications. We then present a high-level formalization of the medical NLP problem in order to provide structure as to how various aspects of NLP fit and complement one another. Examples of approaches that target various forms of representations and degrees of potential accuracy are discussed. Individual NLP subtasks are subsequently discussed. We conclude this chapter with evaluation methods and a discussion of the directions expected in the processing of clinical medical reports. Throughout, we describe applications illustrating the many open issues revolving around medical natural language processing.