Developing an Ontology-based Metadata Management System for Heterogeneous Clinical Databases By Quddus Chong Winter 2002

Outline  Towards a clinical data warehouse  Integrating heterogeneous data sources  Clinical abstractions as Ontologies  Managing database metadata  The data mediator approach  Using Protégé-2000

Towards a Clinical Data Warehouse  Clinical Data Warehousing is the application of Data Warehousing concepts to allow clinical data about a large patient population to be analyzed to perform clinical quality management and medical research.  In a data warehouse environment, data has the following properties: Data is organized by subject, or domain-level concepts, rather than by function. Data from various operational systems is integrated, by definition or by content. Data is archived in non-volatile storage to allow temporal analysis. Data is recorded with a temporal dimension (e.g. timestamp) Data is optimized for decision making (DSS) or analysis (OLAP).

Integrating Heterogeneous Data Sources  The main challenge in integrating data from heterogeneous sources is in resolving schema and data conflicts.  Approaches to this problem include using a federated database architecture, or providing a multi-database interface. These approaches are geared more towards providing query access to the data sources than towards supporting analysis.  Types of data integration: Physical integration – convert records from heterogeneous data sources into a common format (e.g. ‘.xml’). Logical integration – relate all data to a common process model (e.g. a medical service like ‘diagnose patient’ or ‘analyze outcomes’). Semantic integration – allow cross-reference and possibly inferencing of data with regards to a common metadata standard or ontology (e.g. HL7 RIM, OIL+DAML).

Clinical abstractions as Ontologies  An ontology is a explicit specification of the conceptualization of a domain. Information models (such as the HL7 RIM) and standardized vocabularies (such as UMLS) can be part of an ontology. An ontology provides a core component in a Knowledge-Based System.  In the clinical research field, ontologies have been used in computerized guideline modeling. This allows the development of applications to provide recommendations (e.g. to make indications for the use of surgical procedures), to identify deviations in practices, and screening services (e.g. evaluate patient eligibility).  Benefits of using ontologies include: Facilitating sharing between systems and reuse of knowledge Aiding new knowledge acquisition Improving the verification and validation of knowledge-based systems.

Managing database metadata  Metadata is the detailed description of the instance data; the format and characteristics of the populated instance data; instances and values dependent on the requirements/role of the metadata recipient.  Metadata is used in locating information, interpreting information, and integrating/transforming data.  Being able to maintain a well-organized and up-to-date collection of the organization’s metadata is a great step towards improving overall data quality and usage. However this task is complicated by the different quality and formats of metadata available (or not) from the heterogeneous data sources, and the consistency in updating existing metadata.  A common metadata architecture is essential to keeping data manageable.

The Data Mediator approach  In this project, we will attempt to develop an extensible and adaptable architecture to perform integration of heterogeneous data sources into a data warehouse environment using a ontology-based data mediator approach.  The components of this architecture include: Knowledge base – stores the ontology; consists of:  The abstraction model – domain-level concepts  The database description model – metadata record of data sources  The mappings model – how data elements relate to attributes in the abstraction model  The transformations model – metadata of available methods to transform data elements from one data source to another Data mediators – provides each data source an interface to the warehouse and resolving data conflicts between any different representations; necessary classes generated from the ontology. Data warehouse – provides access to integrated data for analysis and decision-making.

Patient model (adapted from SMI Dharma model)  The patient-data information model defines the classes and attributes of patient data for an Electronic Patient Record (EPR).  The patient-data model consists of: a Patient class whose instances hold demographic information about specific patients a Note_Entry class that describes qualitative observations about patients a Numeric_Entry class that represent results of quantitative measurements an Adverse_Event class that models adverse reactions to specific substances a Condition class that represent medical conditions that persist over time, and two intervention classes Medication and Procedure, that model drugs and other medical procedures that have been recommended, authorized, or used.  The defining characteristic of entities in the patient-data model is that they are assertions about demographic and clinical conditions of specific patients.

Database metadata model (adapted from Critchlow et. al.)  The metadata model here contains the information needed for the data integration process.  The database description model contains language independent class definitions that closely mirror the physical layout of a source database. In our prototype model, the database description is simply a class containing a set of database entries. A model is provided for two distinct entry- types: field-entries (from flat-file data sources) and column-entries (from relational data sources). Entries are essentially instances of the attribute class.  Modeling the database metadata as an ontology provides flexibility when trying to describe heterogeneous data sources. For instance, the model can be easily extended to describe Native XML databases.  How the models are used in data integration: The source database attributes are mapped to the appropriate abstraction characteristic through mappings. When an abstraction defines multiple representations for the same characteristic attribute, transformation functions are defined to convert between them.

A prototype architecture Source db 1 Target db Abstractions Data Descriptions Data Mappings Transformation Descriptions Ontology Server (Relational DBMS, e.g. MySQL) Source db 2 (Object-Relational DBMS, e.g. Postgresql) (Data Warehouse environment, e.g. SQL Server) Mediator Interface 1 Mediator Interface 2 Warehouse Mediator *possible use of JDBC metadata to obtain db descriptions *possible use of XSLT to perform data transformations *alternatively, a common metadata exchange standard such as XMI could be used *abstraction model in the ontology is extensible to any domain *XML data binding could be used to generate APIs for data validation or transformation *key goal: develop the ontology server as a component, use EJB or.NET *ontologies can be created and modified via Protégé-2000 tool; underlying format is RDF

Using Protégé-2000  Protégé-2000 is a experimental knowledge-acquisition tool, written in Java, that allows users to import, export and create their own ontologies.  The tool itself is extensible; a programming developer kit is available for instructions on creating plug-ins: ‘tabs’ - user interface between a ontology model in Protégé and another knowledge-based application. ‘slot-widget’ – user interface for viewing and acquiring slot values for new instances. backend plug-ins – specify the mechanism that Protégé-2000 will use to store the ontology.

Screenshot: Creating the classes and slots of an ontology

Screenshot: Viewing the newly created ontology model

References  Pedersen T. B., Jensen C. S., “Research Issues in Clinical Data Warehousing” In Proceedings of the 10 th International Conference on Scientific and Statistical Database Management, pg , July 1998 (available online: )  Critchlow T., Ganesh M., Musick R., “Meta-Data Based Mediator Generation” In Proceedings of the 3 rd IFCIS Conference on Cooperative Information Systems, August 1998 (available online: )  Tu S. et. al. “A Flexible Approach to Guideline Modeling” AMIA Annual Symposium, 1999 (available online: web.stanford.edu/pubs/SMI_Abstracts/SMI html ) web.stanford.edu/pubs/SMI_Abstracts/SMI html