Semantic Approach for Text Understanding of Chief Complaints Data

Abstract

We propose a semantic approach to processing free form text information such as chief complaints using formal knowledge representation and Description Logic (DL) reasoning. Our method extracts concepts and as much contextual information as is available in the text. Output is a computationally interpretable representation of this information using the Resource Definition Framework (RDF) and UMLS Metathesaurus [1].

Introduction

Chief complaints are often represented textually and as a mixture of complex and context-dependant lexical symbols with little formal sentence structure [2]. Human experts usually comprehend text in its right context intuitively and effortlessly, but use of chief complaint data by computers is a challenge. Semantic approaches to text understanding are concerned with the meanings of terms and their relationships, driven by an explicit model, rather than their syntactic forms. Explicit representation of domain concepts along with computer reasoning enables a knowledgeable computer agent to identify domain concepts in a text and pinpoint relevant relationships if they make sense according to an existing formal model available to the agent [3].

Method

Our methodology uses RDF [4] and the Web Ontology Language (OWL) for information and knowledge representation [5]. DL inferences are used for classification and case matching. Since there is no guarantee that text will have a formal sentence structure, our methodology considers the entire chief complaint as a single term. After a text preparation process that includes spell-checking and expanding known abbreviations and patterns, a syntactic term parser computes an index of all permutations of plausible subterms extractable from a term based on word location, order, part of speech, and word count. From all plausible subterms, only those under five words long are processed further, assuming that the relevant context for a given concept might be found within 1–4 degrees of separation from the word(s) representing that concept. The MMTx linguistic analysis tool [6] from NLM is used to map eligible subterms to the UMLS Metathesaurus.

Outputs from MMTx include UMLS semantic types [1, 7] for each mapped concept and a mapping score. Only semantic types with a perfect mapping score of 1000 are processed further. An indexer then creates an RDF representation of the original term, its subterms mapped to UMLS, indexed by the order in which they appear in the term and indexed by their semantic types. A subterm may have multiple UMLS maps and one UMLS map may occur more than once in the term or have more than one semantic type.

We have developed an OWL model to represent clinical evidence (observations such as diseases and syndromes, signs and symptoms, trauma or poisoning, healthcare procedures, etc.) as a temporal event having spatial aspects, quantitative and qualitative modifiers, and contextual aspects such as age, presenter, causation, or negation. The model is an extension of the UMLS Semantic Net represented in OWL-DL. A computer agent uses this model, a set of rules, and DL reasoning to interpret the relationship between subterms and their semantic types according to the model.

Results

Our method extracts important clinical observations in nearly all runs and the relevant contextual information in most cases, if they exist. Failures are frequently related to semantically ambiguous or irregular iterations such as ‘referred by doc to check lab’ or ‘patient does not eat/drink/diarrhea’.