Cross-document Analysis for Automatic Understanding of Electronic Medical Records

Abstract

Electronic medical records contain both structured entities such as diagnosis codes and results, and unstructured entities such as textual notes typed by health care providers to record patient information during encounters. They play an integral role in assisting health care providers to manage the diagnoses and treatment plans of individual patients. % Automated understanding aims at extracting entities and relational entities from the notes, which can act as diagnosis indicators.

Existing research proposed to annotate and extract information with the lowest possible error, and surveys have been published to discuss existing work done in this field. Commercial HIS systems also exist. EPIC is an industrial system that supports automated understanding. Health Information Technology for Economic and Clinical Health is also commercial and has less support for automated understanding. The methods are often expensive and always lack support to records from developing countries.

In this thesis, we aim at improving automated understanding of electronic medical records for differential diagnosis analysis.

Differential diagnosis analysis considers several diagnostic algorithms from clinical diagnostic medical textbooks and relates them to the case at hand. For our computational model, we constructed several Bayesian networks that reflect the structure of the diagnostic algorithms. Then we devised a cross-document analysis method to learn the probabilistic parameters of these networks.

The cross-document analysis method performed cross-reference equivalence of the terms in the notes with (i) more rigorous English text with a similar distribution collected from United States Medical Licensing Examination questions, (ii) an expert based map of abbreviations, and (iii) a corpora of medical publications extracted from Pubmed. For that we devised cross-reference equivalence metrics that augmented each term with equivalent terms. Then we estimated the truth of a Bayesian network node by estimating the existence of its terms from the diagnostic algorithms in the electronic medical record in question.

We applied our method to a corpora of 151,930 clinical notes of which 3,616 are annotated. The corpora is collected from American University of Beirut Medical Center and Rafic Hariri University Hospital. The annotations are organized in a tree of diagnoses and clinical differential analysis terms.

After computing the Bayesian network parameters, we queried each Bayesian network with its diagnostic decision node as an {\em explanation} and with the electronic medical note as {\em evidence}. We considered the networks with highest scores candidates for differential analysis.

Our method successfully identified the correct diagnosis among the top two diagnostic algorithms with an average recall of 93\%, and a precision of 99\%. When considering prediction of correct diagnosis, the average precision is 64\%. The analysis often included prevailing diagnoses such as fatigue, headache, and joint sprain which health care providers refer to after eliminating more serious diagnoses.

The thesis work also presents a tool we developed for semi-automatic annotation of electronic medical records with diagnostic graphs, and use of other techniques such as Neural Networks and Hidden Markov Models that showed lower performance.