A Querying Method over RDF-ized Health Level Seven v2.5 Messages Using Life Science Knowledge Resources

Abstract

Background: Health level seven version 2.5 (HL7 v2.5) is a widespread messaging standard for information exchange between clinical information systems. By applying Semantic Web technologies for handling HL7 v2.5 messages, it is possible to integrate large-scale clinical data with life science knowledge resources.

Objective: Showing feasibility of a querying method over large-scale resource description framework (RDF)-ized HL7 v2.5 messages using publicly available drug databases.

Methods: We developed a method to convert HL7 v2.5 messages into the RDF. We also converted five kinds of drug databases into RDF and provided explicit links between the corresponding items among them. With those linked drug data, we then developed a method for query expansion to search the clinical data using semantic information on drug classes along with four types of temporal patterns. For evaluation purpose, medication orders and laboratory test results for a 3-year period at the University of Tokyo Hospital were used, and the query execution times were measured.

Results: Approximately 650 million RDF triples for medication orders and 790 million RDF triples for laboratory test results were converted. Taking three types of query in use cases for detecting adverse events of drugs as an example, we confirmed these queries were represented in SPARQL Protocol and RDF Query Language (SPARQL) using our methods and comparison with conventional query expressions were performed. The measurement results confirm that the query time is feasible and increases logarithmically or linearly with the amount of data and without diverging.

Conclusions: The proposed methods enabled query expressions that separate knowledge resources and clinical data, thereby suggesting the feasibility for improving the usability of clinical data by enhancing the knowledge resources. We also demonstrate that when HL7 v2.5 messages are automatically converted into RDF, searches are still possible through SPARQL without modifying the structure. As such, the proposed method benefits not only our hospitals, but also numerous hospitals that handle HL7 v2.5 messages. Our approach highlights a potential of large-scale data federation techniques to retrieve clinical information, which could be applied as applications of clinical intelligence to improve clinical practices, such as adverse drug event monitoring and cohort selection for a clinical study as well as discovering new knowledge from clinical information.

Keywords: Semantic Web; electronic health records; health level seven; information storage and retrieval; linked open data.

Figure 1

Examples of an SS-MIX2 storage structure and an HL7 message. This example HL7 message (RDE^O11) contains information on a medication order for a patient identified by 0123456789 administered on May 28, 2013. The message contains the following segments: message header (MSH), patient identification (PID), order-related information (ORC), pharmacy encoded (RXE), and timing and quantity (TQ1).

Figure 2

A medication order in the HL7 standard format, XML-encoded format, and after conversion to RDF.

Figure 3

Serialized RDF representation of a medication order in turtle format.

Figure 4

Relationships between USP, KEGG, and MEDIS DRUG used in search for atypical antipsychotic drugs that have an inhibitory effect on the 5HT2C receptor or the H1 receptor.

Figure 5

SPARQL expression of Query 1. This query searches all medication orders for drugs classified as renin angiotensin inhibitors.

Figure 6

SPARQL expression of Query 2. This query searches all cases for which a leukocyte count of 3000 or less was observed during the medication period of drug types having leukopenia or neutropenia as adverse events.

Figure 7

SPARQL expression of Query 3. This query searches all cases satisfying the criteria for impaired glucose tolerance during a period between the initial and final medications of atypical antipsychotic drugs that have a 5HT2C or H1 receptor inhibitory effect. The clinical cases that satisfy the above criteria within 60 days of the initial medication are excluded. In this query, two subqueries are used. In subquery 1, the cases having the period of initial and final medications of the atypical antipsychotic are identified. In subquery 2, the cases satisfying the criteria for impaired glucose tolerance during the period are identified.

Figure 8

The average measured execution times of Queries 1, 2, and 3 obtained through the experiments are shown in a), b), c), respectively. In each subfigure, bar graphs represent the average measured execution times of the two types of query expression with standard errors, and solid or dashed line represent the approximate average execution times.