Context identification in electronic medical records

Abstract

In order to automate data extraction from electronic medical documents, it is important to identify the correct context of the extracted information. Context in medical documents is provided by the layout of documents, which are partitioned into sections by virtue of a medical culture instilled through common practice and the training of physicians. Unfortunately, formatting and labeling is inconsistently adhered to in practice and human experts are usually required to identify sections in medical documents. A series of experiments tested the hypothesis that section identification independent of the label on sections could be achieved by using a neural network to elucidate relationships between features of sections (like size, position from start of the document) and the content characteristic of certain sections (subject-specific strings). Results showed that certain sections can be reliably identified using two different methods, and described the costs involved. The stratification of documents by document type (such as History and Physical Examination Documents or Discharge Summaries), patient diagnoses and department influenced the accuracy of identification. Future improvements suggested by the results in order to fully outline the approach were described.