Retrieval of Similar Electronic Health Records Using UMLS Concept Graphsby:Laura plaza & ALBERTO DIAZ Presented by: PriyaWadhwa
Outline • Introduction • Related work • EHRs • Concept Annotation • Retrieval of information • Negation Detection • Evaluation Methodology • Conclusion & future work
Introduction • Physicians often use information from previous clinical cases. • Large no. of records make decision making exhaustive and unfeasible. • Previous knowledge-most significant factor in decision making.
Contd… • Medical records’ information stored as free text. • Method of retrieving similar clinical cases by mapping the text onto UMLS concepts. • Information is difficult to analyze. • Negation detection plays an important role in understanding. • May be considered a case of Information retrieval stated as “Given a reference record, retrieve other records from the clinical database that are similar to the referenece one”.
Patient records represented as semantic graphs. • Vertices –UMLS concepts. • Edges- is-a relationship b/w them.
Related Work • Term based indexing used. • Its simple & powerful. • Concept based indexing can improve performance. • SAPHIRE-the pioneer • Development of methods for indexing medical documents from bibliographic DB. • CLEF- development of NLP of Medical records.
Electronic health records • Unique attributes • Structure and content vary with user needs. • Mix of highly structured information and idiosyncratic narrative text. • Images may be included. • Size vary • Negation Detection important. • Negation in natural language can be subtle.
Contd.. • Terminology & writing practices makes concept detection ambitious task. • Synonyms & homonyms. • Neologism • Elisions & abbreviations.
Concept Annotation • UMLS – most popular biomedical terminologies in NLP application. • 3 main components: Specialist lexicon Metathesaurus Semantic Network • Meta Map Transfer tool
Automatic Retrieval of Similar EHRs • Concept graph based • Consists of 4 steps • Lets take an example: • Clinical history : Eleven year old with ALL, bone marrow transplant on Jan2,now with 3 day history of cough. • Impression: No focal pneumonia. Likely chronic changes at the left lung base. Mild anterior wedging of the thoracic vertebral bodies.
Extraction of UMLS concepts • Text in EHR mapped onto UMLS concepts using MetaMap • Physician consulted • Two clinical records can be similar if: the same symptom or sign present patients received same diagnosis same test or procedure reported same medicament has been administered
Negation Detection • Detect negated concepts in EHRs • Absent symptoms or diseases not relevant • Simple lexical scanner used
Semantic graph representation • Creating graph based representation for each EHR • Concepts identified in previous step retrieved from UMLS metathesaurus • All concept hierarchies are merged • Each concept is assigned a weight • Weight(A,B)=|a&b|/|a or b| a = set of all parents of concept A including A b= set of all parents of concept B including B
Computing similarity b/w EHRs • Compute similarity b/w the graph based representation of two patients records • Given two graphs, A & B • Each concept of A which is present in B assigns a score equal to weight of that concept in graph A and 0 otherwise • Sum of the scores for all concepts in A is computed • Result is normalized in interval [0,maximum similarity]
Evaluation Methodology • Collection of radiology reports • 50 reports obtained from corpus • Ten categories used • Two hospital physicians queried • Comparison b/w EHRs retrieved by the system and the ones retrieved by the experts
Conclusion and Future work • Novel approach to automatic retrieval of similar EHRs • Method gets a richer representation than traditional models • High precision • Future work will test the method on different evaluation collection. • User oriented evaluation. • Take into consideration spelling errors.