Integration Of Declarative and Procedural Knowledge for The Management of Chronic Patients


Published on

Yuval Shahar, M.D., Ph.D.
Medical Informatics Research Center
Department of Information Systems Engineering
Ben-Gurion University
Beer Sheva, Israel
(16/10/08, Plenary session 3)

1 Like
  • Be the first to comment

No Downloads
Total views
On SlideShare
From Embeds
Number of Embeds
Embeds 0
No embeds

No notes for slide
  • Integration Of Declarative and Procedural Knowledge for The Management of Chronic Patients

    1. 1. Medical Informatics Research Center Department of Information Systems Engineering Ben-Gurion University Beer Sheva, Israel Yuval Shahar, M.D., Ph.D. Integration Of Declarative and Procedural Knowledge for The Management of Chronic Patients
    2. 2. The BGU Medical Informatics Research Center <ul><li>The main academic center in this area in Israel </li></ul><ul><li>Integrates researchers from the engineering, medical, natural science, and management faculties of Ben Gurion University </li></ul><ul><ul><li>Laboratory based in the Information Systems Engineering Department </li></ul></ul><ul><ul><li>20 M.Sc./Ph.D students from multiple faculties and universities </li></ul></ul><ul><li>Focuses on medical decision-support systems, especially in chronic-patient management </li></ul><ul><ul><li>Intelligent monitoring of chronic patients over time </li></ul></ul><ul><ul><li>Automated Intelligent interpretation of large numbers of longitudinal clinical data </li></ul></ul><ul><ul><li>Automated support to guideline-based care </li></ul></ul><ul><ul><li>Quality assurance and quality assessment of medical care </li></ul></ul><ul><ul><li>Data mining of large clinical databases, in particular over time </li></ul></ul><ul><ul><li>Individual and policy-oriented medical decision making </li></ul></ul>
    3. 3. <ul><li>Procedural Knowledge: </li></ul><ul><li>Clinical Practice Guidelines </li></ul><ul><li>Clinical Practice Guidelines ( CPGs ) are a set of schematic plans for management of patients with a particular clinical condition (e.g., diabetes) </li></ul><ul><li>CPGs have a potential to improve medical care and to reduce medical costs </li></ul><ul><li>CPGs are powerful if and only if they are: </li></ul><ul><ul><li>Accessible at the point of care </li></ul></ul><ul><ul><li>Actually applied and followed </li></ul></ul><ul><li>Computer-based techniques are required to support automation of guideline-based care </li></ul>
    4. 4. <ul><li>Guideline-based Care Tasks </li></ul><ul><li>Specification & Representation </li></ul><ul><ul><li>Necessitates a complete methodology </li></ul></ul><ul><li>Validation & Verification </li></ul><ul><li>Search & Retrieval </li></ul><ul><li>Display & Browsing </li></ul><ul><li>Applicability & Eligibility determination </li></ul><ul><li>Runtime Application </li></ul><ul><li>Retrospective Quality Assessment </li></ul>
    5. 5. <ul><li>Runtime Guideline Application: The Overall Picture </li></ul>Care provider (monitoring, therapy) Clinical-guideline library Clinical data repository Medical expert Knowledge engineer Decision-support system Health-care manager (QA, planning, analysis) (Storage, search, retrieval, display) (Knowledge specification & maintenance)
    6. 6. <ul><li>The Guideline Conversion Problem </li></ul><ul><li>Knowledge Acquisition aspects: </li></ul><ul><ul><li>Expert physicians are not programmers </li></ul></ul><ul><ul><li>Knowledge engineers do not understand the clinical semantics of the guidelines nor its tacit knowledge </li></ul></ul><ul><li>Functional aspects: </li></ul><ul><ul><li>Free-text representations are useful for additional tasks such as for guideline search, retrieval, and display </li></ul></ul><ul><ul><li>Formal representations are essential for automation </li></ul></ul>How will the large mass of free-text guidelines be converted to machine-comprehensive, executable representations, while preserving the benefits of the textual format?
    7. 7. The Digital Electronic Guideline Library (DeGeL) [Shahar, Young et al, JBI 2004] <ul><li>Implements an incremental , collaborative process for specification of guidelines by medical experts, clinical knowledge editors, and knowledge engineers, catering for their different skills </li></ul><ul><li>Uses a hybrid , multiple-representation model: </li></ul><ul><ul><ul><li>Semi-structured format (marking-up text segments, using semantic labels from a target guideline ontology) </li></ul></ul></ul><ul><ul><ul><li>Semi-formal (includes control information) </li></ul></ul></ul><ul><ul><ul><li>Formal (machine executable) </li></ul></ul></ul><ul><li>Includes tools for upload and semantic indexing of guideline source documents, semantic markup using multiple GL ontologies, concept-based (using the semantic indices) and context-sensitive (within labeled content) search & retrieval, and runtime application </li></ul><ul><li>Supports a multiple-role authorization model </li></ul>
    8. 8. The Lifecycle of a Hybrid Guideline in DeGeL <ul><li>A collaboration among medical experts (consensus formation), clinical knowledge editors (semi-structuring and semi-formal markup), and knowledge engineers (formal structuring), using multiple representation formats </li></ul>
    9. 9. The GESHER CPG Structuring and Maintenance Tool: Text Markup and Hierarchical Plan Definition
    10. 10. The Uruz Web-Based Semi-Structuring Mark-Up Tool
    11. 11. The Uruz Semi-Formal Plan-Body Wizard
    12. 12. Concept-Based Search in Vaidurya, Using The Semantic Axes
    13. 13. Context-Sensitive Search in Vaidurya, Using the Marked-Up Knowledge Roles
    14. 14. The VisiGuide Guideline-Browsing Tool <ul><li>Enables visualization, browsing, and exploration of search results </li></ul><ul><li>Exploits the original semantic classification and the internal marked-up structure </li></ul>
    15. 15. Declarative Knowledge: The Need for Intelligent Interpretation of Multiple Time-Oriented Clinical Data <ul><li>Many medical tasks, especially those involving chronic patients, require extraction of clinically meaningful concepts from multiple sources of raw, longitudinal, time-oriented data </li></ul><ul><ul><li>Example: “Modify the standard dose of the drug, if during treatment , the patient experiences a second episode of moderate anemia that has persisted for at least two weeks ” </li></ul></ul><ul><li>Examples of clinical tasks that require temporal reasoning: </li></ul><ul><ul><li>Monitoring and Diagnosis </li></ul></ul><ul><ul><ul><li>Searching for “a gradual increase of fasting blood-glucose level” </li></ul></ul></ul><ul><ul><li>Therapy </li></ul></ul><ul><ul><ul><li>Following a treatment plan based on a clinical guideline </li></ul></ul></ul><ul><ul><li>Quality assessment </li></ul></ul><ul><ul><ul><li>Comparing observed treatments with those recommended by a guideline </li></ul></ul></ul><ul><ul><li>Research </li></ul></ul><ul><ul><ul><li>Detection of hidden dependencies over time between clinical parameters </li></ul></ul></ul>
    16. 16. The Need for Intelligent Mediation: The Gap Between Raw Clinical Data and Clinically Meaningful Concepts <ul><li>Clinical databases store raw, time-stamped data </li></ul><ul><li>Care providers and decision-support applications reason about patients in terms of abstract, clinically meaningful concepts, typically over significant time periods </li></ul><ul><li>A system that automatically answers queries or detects patterns regarding either raw clinical data or concepts derivable from them over time, is crucial for effectively supporting multiple clinical tasks </li></ul>
    17. 17. The IDAN Temporal-Mediation Architecture [Boaz and Shahar, AIM 2005] <ul><ul><li>Enables access to heterogeneous time-oriented clinical data sources </li></ul></ul><ul><ul><li>Supports querying for both patient raw data and their abstractions </li></ul></ul><ul><ul><li>Integrates clinical data sources with clinical knowledge sources </li></ul></ul><ul><ul><li>Supports applications such as the KNAVE-II interactive visual-exploration tool and the Spock guideline-application tool </li></ul></ul>
    18. 18. The IDAN Temporal-Abstraction Mediator (Boaz and Shahar, AIME 2003, AIM 2005) Temporal- Abstraction Controller Knowledge- Acquisition tool Standard Medical Vocabularies service KNAVE-II Knowledge Service Temporal- Abstraction Service Data Access Service Medical Expert Clinical User
    19. 19. The KNAVE-II Browsing and Exploration Interface [Shahar et al., AIM 2006] Overall pattern Raw clinical data Intermediate abstractions Medical knowledge browser Concept search
    20. 20. Evaluation of KNAVE-II (Palo Alto Veterans Administration Health Care System) <ul><li>Eight clinicians with varying medical/computer use backgrounds </li></ul><ul><ul><li>A second study used 6 additional clinicians and more difficult queries </li></ul></ul><ul><li>Each user was given a brief demonstration of the interface </li></ul><ul><li>The evaluation used an online database of more than 1000 bone-marrow transplantation patients followed for 2 to 4 years </li></ul><ul><li>Each user was asked to answer 10 queries common in oncology protocols, about individual patients, at increasing difficulty levels </li></ul><ul><li>A cross-over study design compared the KNAVE-II module versus two existing methods (in the 2 nd study, users chose which): </li></ul><ul><ul><li>Paper charts </li></ul></ul><ul><ul><li>An electronic spreadsheet (ESS) </li></ul></ul><ul><li>Measures: </li></ul><ul><ul><li>Quantitative : time to answer and accuracy of responses </li></ul></ul><ul><ul><li>Qualitative : the Standard Usability Score ( SUS ) and comparative ranking </li></ul></ul>
    21. 21. The KNAVE-II Evaluation Results (Martins et al., MEDINFO 2004, AIM 2008) <ul><li>Direct Ranking comparison : KNAVE-II ranked first in preference by all users </li></ul><ul><li>Detailed Usability Scores : The Standard Usability Scale ( SUS) mean scores: KNAVE-II 69, ESS 48, Paper 46 (P=0.006) (more than 50 is user friendly) </li></ul><ul><li>Time to answer : </li></ul><ul><ul><li>Users were significantly faster using KNAVE-II as the level of difficulty increased, up to a mean of 93 seconds difference versus paper, and 27 seconds versus the ESS, for the hardest query (p = 0.0006) </li></ul></ul><ul><ul><li>The second evaluation, using more difficult queries and more advanced features of KNAVE-II, emphasized the differences even further: The comparison with the ESS showed a similar trend for moderately difficult queries (P=0.007) and for hard queries (p=0.002); on the average, study participants answered each of the two hardest queries 277 seconds faster using KNAVE than the ESS </li></ul></ul><ul><li>Correctness : </li></ul><ul><ul><li>Using KNAVE-II significantly enhanced correctness versus using paper , especially as level of difficulty increased, even in the initial study (P=0.01) (99% accuracy with K-II versus only 78% paper accuracy, 1 st study; 92% with K-II vs. 57% for ESS, 2 nd study) </li></ul></ul><ul><ul><li>The correctness scores for KNAVE-II versus ESS in the second study, which used more difficult queries, are significantly higher for all queries (p<0.0001) </li></ul></ul>
    22. 22. Exploration of Patient Populations: The VISITORS System [Klimov and Shahar, AMIA 2005, MEDINFO 2007] <ul><li>VIS ualizat I on and exploration of T ime- O riented raw data and abstracted concepts for multiple patient R ecord S </li></ul><ul><ul><li>Knowledge-based time-oriented interpretations of the raw clinical data </li></ul></ul><ul><ul><li>Visual display and interactive exploration </li></ul></ul><ul><ul><li>Multiple -record aggregation and association </li></ul></ul>
    23. 23. The VISITORS Main Interface Multiple-subjects raw data Medical knowledge browser Subject groups Distribution of derived patterns over time Concept search
    24. 24. Temporal Association Charts <ul><ul><li>The data of each subject are connected by a line </li></ul></ul>Abstractions for the same subject group are connected; support and confidence are indicated by width and hue
    25. 25. <ul><li>Assists the care-provider in applying the guideline to a specific patient at the point of care </li></ul><ul><li>Currently geared to the Asbru guideline ontology </li></ul><ul><li>Supports the guideline-application process to some extent regardless of: </li></ul><ul><ul><li>Guideline representation level </li></ul></ul><ul><ul><li>Access to an EPR </li></ul></ul><ul><li>Maintains an interactive dialog with the user concerning patient data and recommended actions </li></ul><ul><li>Keeps detailed records (log) of the guideline application and of the user’s decisions </li></ul>Integration of Procedural and Declarative Knowledge: The Spock Guideline-Application System [Young and Shahar, AMIA 2005; Young et al., JBI 2007]
    26. 26. Spock’s Hybrid-Application Matrix:
    27. 27. Overview of the Spock System Architecture
    28. 32. Integration of Procedural and Declarative Knowledge: Retrospective Quality Assessment and The Guideline-Matching Task <ul><li>Task : Match a set of time-oriented data patterns (tests, observations, actions) to a set of time-oriented procedures (i.e., guidelines) </li></ul><ul><li>Supports the performance of guideline-based quality and effectiveness assessment automatically </li></ul><ul><li>Suggests how to proceed in the application of a potentially identified guideline </li></ul><ul><li>Finds which guidelines are in usage , if any (e.g. for research purposes) </li></ul>
    29. 33. The Guideline Recognition Problem <ul><li>In practice </li></ul><ul><li>There is no record in the EMR as to which guideline the physician chooses to follow </li></ul><ul><li>Physicians often do not adhere to all parts of a guideline </li></ul><ul><li>There is a need to match vague and non-deterministic information in the guideline to exact data the EMR </li></ul><ul><li>There is a need to specify ahead of time the distance/similarity functions between each guideline component and the data </li></ul><ul><li>The output must include a continuous degree of fit and a justification for that degree </li></ul>0.83 ( with justification )
    30. 34. The Guideline-Matching Method [Boldo and Shahar, 2004; Boldo, 2007] <ul><li>A knowledge base of abstract features (e.g., “ diastolic blood pressure state ”) is built for each given guideline </li></ul><ul><ul><li>Each feature includes an importance measure , an abstraction function and a fuzzy membership function </li></ul></ul><ul><li>Each medical record is transformed automatically into a representation of abstract features, using the knowledge base </li></ul><ul><li>For each record and a given guideline, a score of the (fuzzy) degree of match of the record to the guideline is generated, using </li></ul><ul><ul><li>The degree of match of each feature </li></ul></ul><ul><ul><li>The importance measure of the feature </li></ul></ul>
    31. 35. <ul><li>Knowledge base : Two recent anti-hypertensive guidelines, JNC6 and JNC 7 </li></ul><ul><li>Patient data : 773 EMRs; 485 EMRs that contain initial drug therapy </li></ul><ul><ul><li>Site: 16 medical centers in Israel </li></ul></ul><ul><ul><li>Gender : 238 female patients; 247 male patients </li></ul></ul><ul><ul><li>Ages : between 21 and 90 years old </li></ul></ul><ul><li>Results : A very good match (P < 0.01) with the clinical expert’s judgments </li></ul>Evaluation of The Guideline-Matching Method
    32. 36. <ul><li>Summary </li></ul><ul><li>Medical decision-support systems need to integrate clinical procedural and declarative knowledge, with a significant potential for improvement in the quality of chronic-patient medical care and for reduction of its cost </li></ul><ul><li>Automated guideline application can be supported at the point of care – possibly at the patient’s home! </li></ul><ul><li>DeGeL : A comprehensive hybrid , multiple-ontology architecture that caters for the full life cycle of a guideline </li></ul><ul><li>IDAN : An intelligent temporal mediator to patient data </li></ul><ul><li>KNAVE-II and VISITORS support intelligent visualization & exploration of individual and multiple electronic patient records, respectively </li></ul>
    33. 37. <ul><li>Acknowledgements </li></ul><ul><li>This research was supported in part by NIH award No. LM-06806 </li></ul><ul><li>We thank our Stanford University and Ben Gurion University collaborators, Drs. Mary Goldstein, Susana Martins, Laszlo Vaszar, Lawrence Basso, Yair Liel, Reuven Sobel, Guy Bar, Avi Yarkoni, Tal Marom, Akiva Leibovitch and Eitan Lunenfeld, for their priceless assistance and contribution </li></ul><ul><li>For more information: </li></ul>