Presentation is loading. Please wait.

Presentation is loading. Please wait.

GLARE (GuideLine Acquisition Representation and Execution) Paolo Terenziani Dipartimento di Informatica, Universita’ del Piemonte Orientale “Amedeo Avogadro”,

Similar presentations


Presentation on theme: "GLARE (GuideLine Acquisition Representation and Execution) Paolo Terenziani Dipartimento di Informatica, Universita’ del Piemonte Orientale “Amedeo Avogadro”,"— Presentation transcript:

1 GLARE (GuideLine Acquisition Representation and Execution) Paolo Terenziani Dipartimento di Informatica, Universita’ del Piemonte Orientale “Amedeo Avogadro”, Alessandria, Italy - Introduction - Representation formalism - Kernel Architecture: acquisition and execution modules - Decision making - Temporal reasoning - General Architecture (advanced features of GLARE) - Verification

2 Introduction Many different computer systems managing clinical guidelines (e.g., Asgaard, GEM, Gliff, Guide, PROforma,…) Different roles: -support -critique -evaluation -education -…... Clinical guidelines are a means for specifying the “best” clinical procedures and for standardizing them Adopting (computer-based) clinical guidelines is advantageous

3 GLARE (GuideLine Acquisition Representation and Execution) -Joint project with: Gianpaolo Molino, Mauro Torchio Laboratorio di Informatica Clinica, Azienda Ospedaliera S. Giovanni Battista, Molinette, Torino, Italy Stefania Montani, Alessio Bottrighi, Luca Anselma, Gianluca Correndo -Domain independent (e.g., bladder cancer, reflux esophagitis, heart failure) -User-friendly (limited number of primitives)

4 GLARE (Guideline Acquisition, Representation and Execution) Representation Formalism A B C B1B1 B2B2 B 1.1 B 1.2 B 2.1 B 2.2 B 2.3 CG B B1B1 B2B2 D Tree of graphs

5 Representation Formalism A B C B1B1 B2B2 B 1.1 B 1.2 B 2.1 B 2.2 B 2.3 CG B B1B1 B2B2 D Tree of graphs Atomic actions

6 Representation Formalism A B C B1B1 B2B2 B 1.1 B 1.2 B 2.1 B 2.2 B 2.3 CG B B1B1 B2B2 D Tree of graphs Atomic actions Composite actions (plans)

7 Representation Formalism Tree of graphs Atomic actions Composite actions (plans) Control relations between actions: -sequence A B C B1B1 B2B2 B 1.1 B 1.2 B 2.1 B 2.2 B 2.3 CG B B1B1 B2B2 D

8 Representation Formalism Tree of graphs Atomic actions Composite actions (plans) Control relations between actions: -sequence -“controlled” (e.g., during) A B C B1B1 B2B2 B 1.1 B 1.2 B 2.1 B 2.2 B 2.3 CG B B1B1 B2B2 D

9 Representation Formalism Tree of graphs Atomic actions Composite actions (plans) Control relations between actions: -sequence -“controlled” -alternative A B C B1B1 B2B2 B 1.1 B 1.2 B 2.1 B 2.2 B 2.3 CG B B1B1 B2B2 D

10 Representation Formalism Tree of graphs Atomic actions Composite actions (plans) Control relations between actions: -sequence -“controlled” -alternative -repetition (e.g. “3 times each 2 days for a month”) A B C B1B1 B2B2 B 1.1 B 1.2 B 2.1 B 2.2 B 2.3 CG B B1B1 B2B2 D

11 Representation Formalism Temporal constraint associated with atomic actions and control relations (see below)

12 Representation Formalism Hierarchy of Action Types Action Plan Query Work action ConclusionDecision Clinical action Pharmacol. prescription Diagnostic decision Therapeutic decision

13 Representation Formalism description of a clinical action

14 Therapeutic decisions Fixed set of parameters (effectiveness, cost, side-effects, compliance, duration) Treatment choice for symptomless gallbladder stones Treatment choice Surgical treatment Expectant management Litholitic therapy

15 Local information associated with treatment choice (in the symptomless gallbladder stones guideline)

16 Diagnostic Decisions *Decision parameters *Decision criteria score-based mechanism For each alternative For each parameter  score  (additive) threshold range

17 Diagnostic Decisions (Gastro Esophageal Reflux Disease) PARAMETERS: heartburn absent (“no-hb”), heartburn lasted not more than 3 months (“hb= 3m”); dysphagia absent (“no-dys”); dysphagia present (“dys”); occurrence of weight loss (“wl”) or non-occurrence (“no-wl”); hemathemesis absence (“no- hem”); hemathemesis presence (“hem”); postural reflux absent (“no-ref”), postural reflux lasted not more than 3 months (“ref= 3m”). THRESHOLD: >9. (One should conclude “no GERD” only if heartburn, dysphagia, weight loss, hematemesis and postural reflux are all absent.)

18 Acquisition Strict interaction with DB’s Clinical DB  hierarchically organized vocabulary  Standardization  Data sharing  Support for semantic checks (e.g., legal attribute values) NOTE: the organization (schema) of Patients DB is equal to the one of the Clinical DB  During acquisition, GLARE gets the information used at execution-time to retrieve automatically the patient’s data (via automatically generated dynamic-SQL queries)

19 Acquisition “Intelligent” helps (syntactic & semantic checks) - “legal” names & “legal” values for attributes - “logical” design criteria (no unstructured cycles, well-formed alternatives & decisions) - “semantic” checks: consistency of temporal constraints

20 Digression 1 We only allow “structured” guidelines Each block (including cycles) has just one entry and one exit a crucial step in order to enforce the semantic clearness of a representation formalism, and, thus, its understandability and practical usefulness (see the discussion about programming languages in the 60’)

21 Architecture of the system (Acquisition part) Clinical DB Pharmac. DB Resource DB ICD-9-CM DB Expert Physician Acquisition Interface Guidelines DB Knowledge Manager

22 Acquisition Graphical Interface

23 Clinical Guidelines Execution Module “Instantiate” a general guideline on a specific patient  Independent of the guideline  Independent of the task (general purpose)  Providing support to decisions Basic Requirements

24 Architecture of the system Clinical DB Pharmac. DB Resource DB ICD DB Expert Physician Acquisition Interface Guidelines DB Knowledge Manager User Physician Execution Interface Guidelines Instantiation DB Patient DB Execution Module

25 Agenda-based execution In the agenda -next actions to be executed -execution time (earliest and latest e.t.)  “On-line” execution: wait until the next e.t. (support physician in clinical activity)  “Simulated” execution: jump to the next e.t. (education, critique, evaluation)

26 Executing atomic actions Work actions: -evaluate pre-conditions -execute action within its range of time -delete from the agenda Query actions: -retrieve data from patient’s DB -wait for data not already in the DB, or for the update of “expired” data Conclusion actions: -insert conclusion into the patient’s DB Decision actions: (see alternatives)

27 Executing composite actions Sequence: -evaluate next action e.t. (given current time and delay) -execute it Concurrent actions: -execute actions according with the temporal constraints Decision + alternative actions (e.g., diagnostic decision): -evaluate parameter values for each alternative, using patient’s DB -determine the score for each alternative -compare the score with the threshold -show the alternatives to the user-physician (distinguishing between “suggested” and “not suggested” ones, and showing parameters and scores) -execute the alternative chosen by the physician (warning available)

28 Executing Clinical Guidelines: other issues  Exits  Failures return to previous decisions (chronological vs. guided backtracking)  Repeated actions and user-defined periodicities -expressive language for user-defined periodicities [IEEE TKDE, 97] -computing next execution time  A user-friendly graphical interface

29 GLARE: Architecture of the system GLARE KERNEL Decision-Making Module Temporal Reasoning Module Update Module Contextualization Module Model-Checking Module SPIN

30 GLARE: Contextualization GLARE KERNEL Temporal Reasoning Module Contextualization Module Update Module Model-Checking Module Decision-Making Module SPIN

31 GLARE: Contextualization Gap between GL generality and the peculiarity of specific contexts of execution is one of the major obstacles to GL dissemination & exploitation In GLARE: software module to automatically adapting (general) GL considering locally available resources * INPUT: a “general” GL + a list of locally available resources * OUTPUT: a “locally-adapted” GL: only paths for which resources are locally available are retained

32 GLARE: Model Checking GLARE KERNEL Temporal Reasoning Module Contextualization Module Update Module Model-Checking Module Decision-Making Module SPIN

33 GLARE: Model Checking After the acquisition of a GL, it is important to verify its properties (e.g., correctness, patient class eligibility, patient applicability) GLARE is loosely coupled with the model-checker SPIN General-purpose solution! Given any GLARE GL, any property that can be expressed in LTL can be checked!

34 GLARE: Update GLARE KERNEL Temporal Reasoning Module Contextualization Module Update Module Model-Checking Module Decision-Making Module SPIN

35 GLARE: Update GL content has to be adapted to local contexts (e.g., country rules) and continuously revised to be up-to-date To grant for quality and EBM, a team of responsible has to evaluate update proposals In GLARE (ongoing work): bitemporal Database support for managing: -proposals of update - acceptance\rejection

36 GLARE: Temporal Reasoning GLARE KERNEL Temporal Reasoning Module Contextualization Module Update Module Model-Checking Module Decision-Making Module SPIN

37 GLARE: Temporal Reasoning Temporal constraints (e.g., delays between actions) play a fundamental role within GL. - A formalism to represent them is needed! - Correct & complete algorithms to propagate the constraints are also needed! GLARE’s Temporal Reasoner can: - check the consistency of constraints (acquisition phase) - check whether actions are scheduled\executed accordingly to the GL constraints, and schedule next actions (execution phase)

38 GLARE: Decision-making GLARE KERNEL Temporal Reasoning Module Contextualization Module Update Module Model-Checking Module Decision-Making Module SPIN

39 GLARE: Decision-making Decision making is a core issue in the clinical practice. In particular, therapy selection is a critical issue. Different parameters (cost, effectiveness, expected utility) must be taken into account In GLARE: a semi-automatic decision support system based on -simulation capabilities -Decision Theory

40 GLARE: current status A Java prototype implements the kernel of the system -The prototype interacts with the (current) patient records (stored in Cache) during execution -All software developed by students: no “fully engineered” and “supported” sowtware available yet Different extensions of the prototype to demonstrate the feasibility of contextualization, decision making, model-checking, and temporal reasoning facilities Ready for the step towards a commercial product

41 Remarks Computer-based approaches to GL can provide crucial adavntages for -Organizations (quality, service optimization, standardization) -Physicians (decision support, quality) -Patients (quality & standardization of the treatment, avoiding errors) A strict and long-term cooperation of Physicians and Computer Scientists has lead to the development of GLARE GLARE embodies advanced Artificial Intelligence techniques, providing automatic support for contextualization, decision making, model-checking, update management, and temporal reasoning

42 GLARE: Decision-making “local information”: considering just the decision criteria associated with the specific decision at hand “global information”: information stemming from relevant alternative pathways in the guideline

43 “What if” facility Facility for gathering the chosen parameter (e.g, resources, costs, times) from the “relevant” alternative paths on the guideline It provides an idea of what could happen in the rest of the guideline if the physician selects a given alternative for the patient, and supports for comparisons of the alternatives

44 Syntomless gallbladder stones treatment choice: “global information” Treatment choice Surgical treatment Expectant management Litholitic therapy Litholitic treatment Choice of surgical appr. Laparoscopy Laparotomy Laparoscopy Laparotomy

45 Syntomless gallbladder stones treatment choice: “global information” Treatment choice Surgical treatment Expectant management Litholitic therapy Litholitic treatment Choice of surgical appr. Laparoscopy Laparotomy Laparoscopy Laparotomy Durationmin:2 daysMax:3 days

46 Syntomless gallbladder stones treatment choice: “global information” Treatment choice Surgical treatment Expectant management Litholitic therapy Litholitic treatment Choice of surgical appr. Laparoscopy Laparotomy Laparoscopy Laparotomy Durationmin:6 daysMax:8 days

47 Syntomless gallbladder stones treatment choice: “global information” Treatment choice Surgical treatment Expectant management Litholitic therapy Litholitic treatment Choice of surgical appr. Laparoscopy Laparotomy Laparoscopy Laparotomy Durationmin:1 dayMax: all life long

48 Syntomless gallbladder stones treatment choice: “global information” Treatment choice Surgical treatment Expectant management Litholitic therapy Litholitic treatment Choice of surgical appr. Laparoscopy Laparotomy Laparoscopy Laparotomy Durationmin:2 monthsMax:1 year

49 Local information associated with treatment choice (in the symptomless gallbladder stones guideline)

50 Digression 2 Why don’t we put “global info (about paths)” locally in the decision actions? Given “local info” in each node, collecting & storing might be automatic HOWEVER: -exponential space in each node -data duplication (consistency after updates?) -not user friendly (too many data!) -not all aternatives are “relevant” -data not always necessary >>global data only at execution time, on request

51

52 Enhancing Decision Making through Decision Theory Supporting therapy selection is a critical objective Selection parameters (cost, side-effects) have to be balanced by the physician Several decisions in sequence (dynamic decision problem) IDEA: Decision theory: a natural candidate for decision support PROBLEM: Mapping concepts between the two areas is not straightforward

53 Justification of the approach Shared assumption: –An (implicit or explicit) query action is always found before a decision action Each decision is based on a data collection completed at decision time Does not depend on the previous history –Markov assumption –Discrete time process –Completely observable Markov Decision Process (MDP)  Decision Theory

54 Correspondence between GL and DT concepts State: parameters that describe the patient’s clinical condition –Query action: means for observing the state State transition: changes in the patient’s state variables –Through GL actions (or exogenous changes) –Observation through query actions –Diagnostic decisions are observable but not needed –The effect of a single work action is typically not observable *State transition produced by the set of actions between two therapeutic decisions Utility: life expectancy corrected by QALYs –Information provided by the medical literature or by phisician interviews Cost: not only monetary expenses –Resources –Time

55 Selecting and adapting DT algorithms to cope with GL Goals: 1.Generation of the Markov model of GLs 2.aOptimal policy evaluation 2.bPath utility evaluation

56 Generation of the Markov model -States: therapeutic decisions -Transitions: macro-actions representing all actions between two consecutive therapeutic decisions -Probability of transitions: literature or interview

57 Selecting and adapting DT algorithms to cope with GL Focusing Path selection to focus on paths of interest Concurrent actions Selection of one possible order of execution Repeated actions different strategies, depending on whether -the number of repetitions in known a priori - the number of repetitions in not known a priori

58 Evaluation of the optimal policy 1The number of repetitions is known - finite time horizon approaches  dynamic programming 2The number of repetitions is not known - infinite time horizon approaches  value iteration

59 Evaluation of path utility Fixing a path = applying a specific policy dynamic programming without maximizing the expected value of the cumulative utility function wrt the different actions value iteration without maximizing wrt the different actions

60 Conclusions Decision theory can provide crucial advantages for supporting therapeutic decision making in GL systems 1Knowledge representation level: analysing the correspondences between DT and GL concepts 2Algorithmic level: selecting and adapting DT algorithms to fit the GL context General methodology (GLARE as an example)

61 Conclusions (Decision Theory-based approach) Decision theory can provide crucial advantages for supporting therapeutic decision making in GL systems 1Knowledge representation level: analysing the correspondences between DT and GL concepts 2Algorithmic level: selecting and adapting DT algorithms to fit the GL context General methodology (GLARE as an example)

62 GLARE: Temporal Reasoning GLARE KERNEL Temporal Reasoning Module Contextualization Module Update Module Model-Checking Module Decision-Making Module SPIN

63 Temporal Constraints in Clinical Guidelines Temporal constraints are an intrinsic part of clinical knowledge (e.g., ordering of the therapeutic actions) Different kinds of temporal constraints, e.g., -duration of actions (min / max) -qualitative constraints (e.g., before, during) -delays (min / max) -periodicity constraints on repeated actions

64 Temporal Constraints in Clinical Guidelines repetitions TheThe therapy for multiple mieloma is made by six cycles of 5-day treatment, each one followed by a delay of 23 days (for a total time of 24 weeks). Within each cycle of 5 days, 2 inner cycles can be distinguished: the melphalan treatment, to be provided twice a day, for each of the 5 days, and the prednisone treatment, to be provided once a day, for each of the 5 days. These two treatments must be performed in parallel.

65 Managing Temporal Constraint: the Problem Temporal Constraints without Temporal Reasoning (constraint propagation) -are useless -clash against users’ intuitions/expectations Both representation and inference are NEEDED

66 Managing Temporal Constraints: the Problem Implied constraint (temporal reasoning): (1.6) C ends between 30 and 60 m after the start of A ABC Correct (consistent) assertion: (1.7) C ends between 30 and 50 m after the start of A Not correct (inconsistent) assertion: (1.8) C ends more than 70 m. after the start of A However: Temporal Reasoning is NEEDED in order to support such an intended semantics!

67 Managing Temporal Constraints: the Problem DESIDERATA for Temporal Reasoning Algorithms - tractability  “reasonable” response time - correctness  no wrong inferences - completeness  reliable answers DESIDERATA for the Representation formalism - expressiveness  capture most temporal constraints in GL TRADE-OFF! SPECIALIZED APPROACHES (since ’80 in AI literature)

68 Digression: Why Completeness is fundamental? Implied constraint (temporal reasoning): (1.6) C ends between 30 and 60 m after the start of A ABC Suppose that temporal reasoning is NOT complete, so that (1.6) is not inferred The answer to query (Q1) might be: YES (Q1) Is it possible that C ends more than 70 m. after the start of A? Complete Temporal Reasoning is NEEDED in order to grant correct answers to queries!

69 Temporal Constraint Treatment WHEN Temporal Reasoning is useful in Guidelines? ACQUISITION -to check consistency EXECUTION -to compare the duration of paths, in hypothetical reasoning (simulation) facilities -to check that the time of execution of actions on patients is consistent with the constraints in the guideline -to schedule next actions

70 Temporal Representation and Temporal Reasoning for Clinical Guidelines Different kinds of temporal constraints No current approach in the AI literature covers all of them Our proposal: an extension of the “consensus” STP approach [Dechter et al., 91] Our goal: expressiveness + correct, complete and tractable inferences

71 GLARE’s approach (representation) Labeled tree of STPs (STPs-tree) Tree of STPs for the multiple mieloma chemotherapy guideline. The overall therapy (node N1) is composed by 6 cycles of 5 days plus a delay of 23 days. In each cycle (node N2), two therapies are executed in parallel: Alkeran (node N3: Sa and Ea are the starting and ending nodes), to be repeated twice a day, and Deltacorten (node N4: Sd and Ed are the starting and ending nodes), to be repeated once a day. Arcs between any two nodes X and Y in a STP (say N2) of the STP-tree are labeled by a pair [n,m] representing the minimal and maximal distance between X and Y.

72 Consistency checking on STPs-trees ALGO1: temporal consistency of guidelines Top-down visit of the nodes in the STPs-tree For each node in the STPs-tree: 1)the consistency of the constraints used to specify the repetition taken in isolation is checked; 2)the “extra” temporal constraints regarding the repetition are mapped onto STP constraints; 3)Floyd-Warshall’s algorithm is applied to the constraints in the STP plus the “extra” STP constraints determined at step 2. Property 1. ALGO1 is correct, complete, and tractable (it operates in O(N 3 ), where N is the number of actions in the guideline).

73 Temporal reasoning on guidelines + instantiations ALGORITHM (sketch)

74 Temporal reasoning on guidelines + instantiations

75 CONCLUSIONS (Temporal Reasoning) Many approaches to time in clinical guidelines in the literature, but properties of the inferential mechanisms (if present) quite neglected GLARE (Guideline Acquisition, Representation and Execution): a proposal of solution based on an extension of well-consolidated Artificial Intelligence techniques Temporal constraints: an essential part of clinical guidelines Temporal constraints: need a principled treatment: not only representation, but also correct, complete (and tractable) inferential mechanisms

76 GLARE: Model Checking GLARE KERNEL Temporal Reasoning Module Contextualization Module Update Module Model-Checking Module Decision-Making Module SPIN

77 Goals of our model-checking work Checking properties (e.g., correctness, liveness,…) of clinical guidelines (GL) is an important task Only limited and ad-hoc verification approaches in the literature (LTL-based) model-checking techniques are successfully used in AI and CS We show how (LTL) model checking techniques can be applied to provide a general tool for the verification of clinical guidelines We extend the GLARE system by integrating it with the model checker SPIN

78 The model checking approach domain property Domain model (formalism) Property represent. (logic) In SPIN: Promela lang. (agent based) In SPIN: LTL logics Interface to logic Model checker TRUE or CONTEREXAMPLE

79 Coupling GLARE with SPIN Tool to (automatically!) translate ANY GLARE clinical GL into a Promela model Details of translation in AMIA’06 paper Flexible and general approach for automatic checking of clinical GL properties

80 programmer Property 1 Property k ………… Guideline 1Guideline n STANDARD approach (ad-hoc) user pgm1Pgm.k

81 Property 1 Property k ………… Guideline 1Guideline n STANDARD approach (ad-hoc) user Our approach (general) SPIN Glare To Spin Promela Guideline 1 Promela Guideline n

82 ………… Guideline 1Guideline n STANDARD approach (ad-hoc) Our approach (general) SPIN Glare To Spin Promela Guideline 1 Promela Guideline n user Property 1 Property k

83 General approach to property checking No need to produce an “ad-hoc” pgm for each type of property! Software tools built once and forall ANY property (that can be expressed in SPIN –i.e. in LTL) can be checked with NO additional effort

84 Model Checking: which properties? Properties concerning a guideline “per se”: one can check if the guideline contains a path of actions satisfying a given set of conditions Properties of a guideline in a given context: specific contexts of execution may impose several limitations on the executable actions of guidelines Properties of a guideline when applied to a specific patient: provided that the model checker has in input all the data in the patient record, the feasibility of a given action, or path of actions on the specific patient can be proved Integrated proofs: any combination of the above types of proofs is feasible

85 EXAMPLE: inconsistencies in a guideline If a recovery treatment has been excluded, later on the guideline cannot prescribe it Given the LTL formula: □ (conclusion ==recovery_treatment_excluded →  proc_recovery_treatment == started) SPIN produces a counterexample to this property.

86 Conclusions (Model Checking) General approach to the automatic check of properties in clinical GL exploiting AI model checking techniques Software tool (translator) built once and forall, to check ANY propery that can be expressed in SPIN (i.e., in LTL) Experiments still ongonig, to fully explore the extremely wide range of properties that can be checked automatically Open issue:enhancing SPIN’s user interface to express (LTL) properties

87 Conclusions (general) Clinical guidelines can provide crucial advantages in the clinical practice Computer-based approaches can help GLARE: a computer-based approach used to experiment the application of formal (Artificial Intelligence, Logical and Database) techniques to accomplish cue goals (e.g., decision support, verification)


Download ppt "GLARE (GuideLine Acquisition Representation and Execution) Paolo Terenziani Dipartimento di Informatica, Universita’ del Piemonte Orientale “Amedeo Avogadro”,"

Similar presentations


Ads by Google