1 / 14

Kahn's TOPAZ System (1988)

Temporal Reasoning and Planning in Medicine Temporal Reasoning In Medical Information Systems (II) Yuval Shahar M.D., Ph.D. Kahn's TOPAZ System (1988). • Input: point-based , unambiguous time-stamped clinical data • An integrated , multiple-temporal–model interpretation scheme

presley
Télécharger la présentation

Kahn's TOPAZ System (1988)

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Temporal Reasoning and Planning in MedicineTemporal ReasoningIn Medical Information Systems (II)Yuval Shahar M.D., Ph.D.

  2. Kahn's TOPAZ System (1988) • Input: point-based, unambiguous time-stamped clinical data • An integrated, multiple-temporal–model interpretation scheme • A numeric model that models quantitatively underlying processes and modifies an atemporal, population-based model into a temporal, patient-specific model • A symbolic, interval-based model that aggregates clinically interesting events into an interval-based hierarchy using context-specific rules • A symbolicstate-based model that generates text paragraphs in the domain's language from the interval-based abstractions, using ATNs • The TNET management system and the temporal-query language TQuery • TNET was extended to ETNET, a knowledge-representation, temporal-reasoning and temporal-management system that used context-specific rules

  3. Summarization Steps in TOPAZ

  4. The TOPAZ System: An Analysis • Highly domain-specific model (one parameter, anatomic site. Disease, drug); not clear how reusable it would be across other clinical domains • Many domains defy complete quantitative numeric modeling - when should modeling stop? (another component might create instability) - which data should be considered? (historical versus recent data) - which parameter should be modified if necessary? (credit assignment) - how should over-fitting be avoided? (especially if a crucial parameter is missing) • The expected population predictions compared to patient-specific predictions (since these are smoother), not to the observed data - detection of change in patient parameters is more difficult - generation of explanations is more challenging, since users look at observed data

  5. Larizza’s M-HTP System(1992) • Monitors heart-transplant patients • Uses a temporal net similar to Kahn's TNET, and a relational database management system • An object-oriented visit hierarchy of visits, off which patient-specific parameters and their values are indexed • An object-oriented knowledge base of significant episodes -Parameter objects include constructs such as Hemoglobin-decrease • A temporal-pattern-matching language used for antecedents of rules: "An episode of PLT-count-decrease overlaps an episode of WBC-count-decrease for at least 3 days during last week"

  6. The Patient-Visits Hierarchy in M-HTP

  7. The Significant-Episodes Taxonomy in M-HTP

  8. Significant Episodes in M-HTP REPRESENTATION OF SIGNIFICANT EPISODES 25 Nov 90 05 Dec 90 15 Dec 90 25 Dec 90 04 Jan 91 Negative_CMV_viremia + + 56 45 25 42 59 24 CMV_antigenemia_increase 24 59 CMV_viremia_increase + + 31 28 37 56 WBC_decrease 59 21

  9. The M-HTP System: An Analysis • A domain-specific instance of more general architectures • The knowledge base encodes hard-coded instances such as WBC-decrease as opposed to instances of the general class gradients for the decreasing value - does not enable inheritance of default values from the gradient class - does not enable sharing of properties among all hematological gradients • No separation of domain-independent knowledge from domain-specific properties - abstractions are not “first class citizens” in the knowledge base, and cannot be described and manipulated using the full language

  10. Kohane’s TUP system(1986) • A temporal utilities package (TUP) system • Demonstrated by the temporal hypothesis reasoning in patient history taking (THRIPHT) medical expert system • A point-based, flexible representation based on the range relation (RREL) creates a constraint network using a restricted, computationally tractable algebra (in particular, no disjunctions such as A <before|after> B) (RREL <point 1 specification> <point 2 specification> <lower bound> <upper bound> <context>) e.g., (RREL ((event MI) (type BEGIN-INTERVAL) (event SGOT-PEAK) (type BEGIN-INTERVAL) (24 hrs) (48 hrs)) • Can reason about alternate temporal hypotheses (monitor contexts)

  11. The TrenDx System(Haimovitz and Kohane, 1993) • Built on top of Kohane's constraint-network TUP system • Encodes patterns as trend templates (TTs) that describe typical clinical patterns as a set of vertical and horizontal constraints • a TT has a set of value constraints of the form minf(D) MAX; min, MAX, are minima and maxima of the function f defined over the measureable parameters D in the temporal range of the interval. • TTs can be matched to partial patterns by maintaining an agenda of candidate patterns that might fit the data (even one point) - A goal-directed approach to pattern matching, starting with pattern • Tested on cases in the growth-chart and intensive-care domains

  12. A Normal-Growth TT in TrenDx

  13. The TrenDx System: An Analysis • Different goals from other systems: Matching of top-level temporal patterns to raw data rather than abstraction or summarization • No knowledge base or abstraction hierarchy such as in IDEFIX, etc. - TTs need to be re-constructed for new tasks: parts are not reused - no knowledge roles (e.g., significant deviation) that can be reused - does not enable use of inheritance among pattern types and instances - no capability for answering queries regarding intermediate-level abstractions - no capability for using intermediate abstractions in the input, since pattern matching must start from raw data • Acquisition of new TT involves definition of all levels of abstraction at the same time - no intent to facilitate elicitation from users • Like other systems, assumes incomplete information about the clinical domain which prevents construction of complete numerical models - associative patterns when knowledge is incomplete and data is sparse

  14. Summary • Temporal representations in medical information systems have moved from symbolic tokens and strings (e.g., MYCIN and Internist-I) to syntactic-level temporal control systems (e.g., TCS), to explicit knowledge-based systems (e.g., VM, IDEFIX), and finally to semantic-level temporal abstraction systems (e.g., TrenDx, M-HTP) • A major issue that is at the focus of current research is facilitation of acquisition, maintenance, reuse and sharing of domain-specific (medical) knowledge

More Related