Temporal Reasoning and Planning in Medicine Automated Support

Temporal Reasoning and Planning in Medicine Automated Support to Guideline-Based Care Yuval Shahar, M. D. , Ph. D.

Clinical Guidelines • • A standard of care, typically an experts’ consensus Usually specifies diagnostic and therapeutic procedures Also known as clinical protocols (e. g. , in oncology); care plans A powerful method to standardize and improve the quality of medical care [Grimshaw and Russel, 1993]. • Increasingly widespread use, to spur best practices in medical care and to incorporate evidence-based medicine • Computer-based techniques needed to automated the support of guideline-oriented medical care. • Example tasks to be supported: determining the applicability of a guideline for a given patient, monitoring the application of the guideline, assessing the effectiveness of the guideline

Characteristics of Automated Support to Guideline-Based Care • Dialog: Care provider automated support system • Both have relative strengths: – Care provider: Better access to patient data and to medical knowledge – Automated system: Better access to guidelines and to temporal patterns • The aim is synergy

Automated Support for Clinical Guidelines: Examples of Prescriptive approaches – DILEMMA, PRESTIGE, Proforma, Prodigy (UK/EU) – Oncocin, T-Helper, EON, ATHENA (Stanford) – Arden Syntax/MLMs (Columbia/LDS) – GLIF (Columbia, Harvard, Stanford) – Active. Guidelines (Epic Systems Co. , USA) – The Pavia web-based diabetes-therapy project (Italy)

Automated Support for Clinical Guidelines: Critiquing Approaches – VT-Attending (Miller, Yale) – Hyper. Critique (Van der Lei and Musen, Rotterdam) – The Asgaard project (Stanford, Vienna, London) integrates both prescriptive and critiquing approaches by representing both the prescribed default actions and the underlying process and outcome intentions

Requirements for Automated Protocol-Based Care • Ability to deal with complexity of patient data (e. g. , time dependencies, abstractions, missing data) • Ability to deal with complexity of protocol actions (e. g. , actions which are themselves protocols) • A scalable and maintainable computational architecture

The Arden Syntax (Hripcsak et al. , SCAMC 1990) • Named after the Arden Homestead in NY, in which representatives from ten universities discussed sharing of medical knowledge • Represents medical knowledge as independent units called Medical Logical Modules (MLMs) • Uses a Pascal-like programming language to encode highly specific rules, grounded in the local institution’s database schema • General medical logic (encoded in the Arden syntax) separated from institution-specific component (encoded in the local query language and terms) • An ASTM standard

The Arden Syntax: An Example • Maintenance: – title: Agranulocytosis and trimethoprim/sulfamethoxazole – author: Dr. Bonzo • Library: – keywords: granulocytopenia; agranulocytosis ; trimethoprim; sulfamethoxazole – citations: 1. Anti-infective drug use. . . Archives of Internal Medicine 1989; 149(5): 1036 -40 • Knowledge – type: data driven; – data: • anc: = read last 2 from ({query for ANC} where it occurred within the past 1 week); • pt_taking_tms : = read exist {query for TMS order}; • evoke: on storage of {ANC}; – logic: • if pt_taking_tms and last anc < 1000 and decrease of anc > 0 then conclude true else conclude false; – action: • store “Caution: The patient’s relative granulocytopenia may be exacerbated by trimethoprim/sulfamethoxazole. ”;

The Arden Syntax: Issues • Difficulty in reuse of general clinical knowledge within different contexts, even within a single system (e. g. , what is “mild anemia”) leading to difficulties in maintenance (Shwe et al. , SCAMC 1992) • Sharing problems encountered when MLMs were transported from Columbia to LDS (Utah); most difficulties due to local query and vocabulary differences as well as local practices (Pryor and Hripcsak, SCAMC 1993) • Difficulty in representation of continuous therapy plans (each MLM represents a well-defined, independent rule) • Lack of ability to represent and reuse higher, meta-level problem-solving knowledge

The EON Project (Musen et al, JAMIA 1996) • A general, client–server architecture that developers can use to build systems that support automated reasoning about guideline-directed care • Includes reusable components, such as – A therapy planner (the episodic skeletal-plan–refinement method) – A temporal mediator (Tzolkin) to the patient database, which includes • the RÉSUMÉ temporal-abstraction system • the Chronus temporal-maintenance system – An eligibility-determination module (Yenta) – A domain knowledge base server • Uses the Common Object Request Broker Architecture (CORBA) as a communication protocol

The EON Architecture: A Conceptual View • Problem-solving components that have task-specific functions (e. g. , planning, classification) • A central database system for queries of both – Primitive patient data – Temporal abstractions of patient data • A shared knowledge base of protocols and general medical concepts

EON as “Middleware” • Software components designed for – incorporation within other software systems (e. g. , hospital information systems) – reuse in different applications of protocol-based care

The EON Architecture: A Graphical View Domain knowledge base ORB ESPR ORB Yenta Other PSMs Guideline-acquisition tool Tzolkin controller RÉSUMÉ Chronus Patient database ORB CORBA BUS

The EON Protégé-Based Guideline-Acquisition Tool

The ATHENA/EON Hypertension-Management System

GLIF (Machado et al. , JAMIA 1998) • Guideline Interchange Format: A specification language • Resulted from the Inter. Med multiple-center collaboration effort (Columbia, Harvard, Stanford) • Attempts to integrate key lessons from MLMs, GEODECM, MTBA, EON • Intended to enable representation of complex plans with branching logic as well as simpler alerts, and therapeutic as well as diagnostic guidelines

GLIF: Necessary Extensions • A formal syntax for conditions (currently strings) • Ability to represent complex temporal expressions and to query patient records for these expressions • Ability to handle uncertainty regarding patient data • Clarification of the application semantics • As in other frameworks: ability to ground the medical concepts within an established, standard vocabulary

GLIF: Current Status • Several guidelines are encoded in paper (breast cancer workup, breast cancer therapy, cholesterol screening, influenza) • A formal Arden-like syntax for conditions is being developed • An interpreter for the conditions is being developed in BWH (Harvard), an expression evaluator (EV), that can be used for determination of conditions such as eligibility • The EV has been integrated into a WWW-based front-end, that "drives" a user through a guideline

A GLIF 3 Flowchart

The Active. Guideline Architecture in Epic. Care. TM (Tang & Young, Proc. AMIA 2000)

Using A Depression Active. Guideline Within Epic. Care. TM

The Prodigy III Scenarios: A High-Level View of a Hypertension Guideline (Johnson et al. , Proc. AMIA 2000)

The Asgaard Project (Shahar, Miksch, and Johnson, AIM 1998) • A task-specific framework for the representation, application, critiquing and quality assessment of time-oriented clinical guidelines • Uses the Asbru guideline-specification language, which includes expressive semantics for sequential, parallel, and periodic actions • Enables explicit representation of intentions as temporal patterns to achieve, avoid, or maintain • Focuses on the critiquing and quality-assessment tasks • Develops algorithms for recognizing and explaining care-provider intentions given their actions, the intentions of the guideline they are applying, and a domain-specific knowledge base

The BGU/Stanford/Vienna/UK Asgaard Project (Shahar, Miksch, and Johnson, AIM 1998) • A task-specific framework for the representation, application, and quality assessment of time-oriented clinical guidelines • Uses the highly expressive Asbru guideline-specification language • Enables explicit representation of process and outcome intentions • The quality-assessment algorithms try to explain care-provider intentions given their actions, the intentions of the guideline they are applying, and a domain-specific knowledge base • Includes a Web-based guideline server at BGU on which an Asbru-based guideline library resides

Summary • Multiple approaches to guideline representation • Prescriptive versus critiquing approaches • major issues: – Grounding of guidelines in the terms of shared vocabularies – Clear semantics – Authoring and maintenance: Knowledge reuse and sharing – Site-specific instantiation, sensitive to local constraints – Improved temporal representations (both for EMRs and for guideline specification languages) – Sufficient expressiveness to capture the intentions of the guideline designers in a machine-readable fashion