Creating Clinical Fuzzy Automata with Fuzzy Arden - PowerPoint Presentation

Slide1

Creating Clinical Fuzzy Automata with Fuzzy Arden SyntaxUsing an ARDS detection automaton as a working exampleJeroen S. DE BRUIN1,2, Heinz STELTZER3, Andrea RAPPELSBERGER1, and Klaus-Peter ADLASSNIG1,21 Section for Artificial Intelligence and Decision Support, Center for Medical Statistics, Informatics, and Intelligent Systems, Medical University of Vienna, Spitalgasse 23, 1090 Vienna, Austria;2 Medexter Healthcare, Borschkegasse 7/5, 1090 Vienna, Austria; 3 Trauma Hospital Vienna South, Kundratstrasse 37, 1120 Vienna, Austria

AMIA 2017 Annual Symposium — November 04‒08, 2017 Washington DC

Medical conditions as automataWith certain chronic or progressive medical conditions (e.g., acute respiratory distress syndrome, or ARDS), it is intuitive to model a disease’s progression in terms of statesStates, or state names, are medical, semantically rich, linguistic conceptsAs the patient gets better or worse, he moves between disease states. This dynamic behavior of health can be modeled by transitions.BenefitsIntuitive, easy to understand, easier to visualizeDrawbacks

States are too rigid, often mutually exclusive, model is oftenoversimplified

Linguistic uncertaintyIn medicine, linguistic concepts used for classification of diseases, symptoms or syndromes are inherently unsharp with respect their boundariesFuzzy sets and logic can be used to reflect this linguistic uncertaintyFuzzy sets express the relationship between linguistic terms and measured or observed data as a degree of compatibilityDegree is calculated by a membership functionFuzzy logic models propositional uncertainty due to

incomplete knowledge of relationships between clinical linguistic conceptsApproximate reasoning instead of exact rule inference

Fuzzy automataA system (patient) can be in more than one state at the same timeMembership of each state is expressed as a degree of compatibility Combinations of state values make the model more expressiveMultiple simultaneous state transitions possibleTransitions between linguistic concepts more gradual and thus more intuitive

Result: Fuzzy automata well-suited for the use in (automated) clinical monitorsEvaluation and interpretation of streams of patient data in brief time intervals Reduction of dimensionality of input dataPresentation of outcomes in semantically meaningful, clinically relevant linguistic concepts

Standard for medical knowledge representation: Arden SyntaxA standard language for writing situation-action rules, procedures, or knowledge bases that trigger results based on clinical events detected in patient dataEach module, referred to as a medical logic module (MLM), contains sufficient knowledge to make at least a single medical decisionExtended by medical knowledge packages (MKPs) consisting of interconnected MLMs for complex clinical decision supportSince version 2.9 – formal support for constructs based on fuzzy set theory and fuzzy logicHealthcare industry and academic users

Fuzzy Arden SyntaxPurposeIntroduce fuzziness into clinical decision making (as a virtue not as a deficiency!) Main conceptsExtension of the truth value model, defining a truth value over a continuous spectrum in a range [0, 1] rather than a dichotomous “true/false” or “yes/no”, resp., model Introduction of the fuzzy set data type to model the

unsharpness of boundaries in definitions of linguistic conceptsIntroduction of three basic propositional fuzzy logic operations – conjunction, disjunction, and negation – which are equipped to handle all truth values in the specified

range

[0, 1

]

Introduction of

parallel,

weighted

program branches

to handle

conditional

statements

where the condition is neither true nor false

Fuzzy Arden Syntax: An exampleNative support for fuzzy set declarationNative support for compatibility calculation

FuzzyArden ARDSKnowledge-based decision supportmonitoring patients with acute respiratory distress syndrome (ARDS) early detection of ARDStherapy advice in ARDS casesInternational study (Vienna, Berlin, Marburg, Paris, Milan)to improve ARDS definitionto compare therapy entry criteria

FuzzyArden ARDS automatonstartnormalhypoxicimproved afterhand bagging

not improved afterhand baggingrespondingto high FiO2not respondingto high FiO2

State

Interpretation

Start

Initial state, undecided

Normal

Oxygenation is satisfactory without additional effort

Hypoxic

Oxygenation is too low

Responding to high FiO

2

Oxygenation was positively affected by high FiO

2

Not responding to high FiO

2

High FiO

2

did not have a desired effect

Improved after

hand bagging

Manual oxygenation through hand bagging has improved oxygenation

Not improved after

hand bagging

Hand bagging did not have the desired effect

FuzzyArden ARDS fuzzy setsConditionFuzzy set definitionGraphical representationadequate oxygenationSaO2 above 97% (93%) for 5 minuteshypoxemiaSaO2 between 90% and 93% (87% and 97%) for 2 minuteshigh FiO2FiO2 above 60% for 30 secondslow FiO2FiO2

below 60% for 30 secondsrapidly improving oxygenationSaO2 increasing from 87–95% oxygenation (85–99%) to 97–100% (93–100%) within 30–90 secondsslowly decreasing oxygenationSaO2 above 96% (91%) steady or oxygenation decreasing to 94% (89%) within 25 minutes

FuzzyArden ARDS fuzzy set definition in Arden Syntax3D fuzzy sets with the same pre- and postconditions can be directly defined3D fuzzy sets with different pre- and postconditions

require explicit definition of both

FuzzyArden ARDS fuzzy set definition in Arden SyntaxDuration modeling natively supported by Arden Syntax3D fuzzy set will yield a degree of compatibility equal to the maximum of all calculations(sup-min composition)

FuzzyArden ARDS linguistic state transitionsBegin stateTransition conditionEnd stateStartAdequate oxygenation

NormalStartHypoxemiaHypoxicNormalHypoxemia

Hypoxic

Hypoxic

Low

FiO

2

∧ adequate oxygenation

Normal

Hypoxic

High

FiO

2

rapidly improving oxygenation

Responding to high FiO

2

Hypoxic

High

FiO

2

hypoxemia

Not responding to high FiO

2

Responding to high FiO

2

Low

FiO

2

slowly decreasing oxygenation

Improved after hand bagging

Responding to high FiO

2

Low

FiO

2

hypoxemia

Not improved after hand bagging

Not responding to high FiO

2

Low

FiO

2

hypoxemia

Hypoxic

Not responding to high FiO

2

High FiO

2

∧ adequate oxygenation

Responding to high FiO

2

Improved after hand bagging

Adequate oxygenation

Normal

Improved after hand bagging

Hypoxemia

Hypoxic

Not improved after hand bagging

Hypoxemia

Hypoxic

Begin state

Transition condition

End state

Start

Adequate oxygenation

Normal

Start

Hypoxemia

Hypoxic

Normal

Hypoxemia

Hypoxic

Hypoxic

Low

FiO

2

∧ adequate oxygenation

Normal

Hypoxic

Responding to high FiO

2

Hypoxic

Not responding to high FiO

2

Responding to high FiO

2

Improved after hand bagging

Responding to high FiO

2

Not improved after hand bagging

Not responding to high FiO

2

Hypoxic

Not responding to high FiO

2

High FiO

2

∧ adequate oxygenation

Responding to high FiO

2

Improved after hand bagging

Adequate oxygenation

Normal

Improved after hand bagging

Hypoxemia

Hypoxic

Not improved after hand bagging

Hypoxemia

Hypoxic

FuzzyArden ARDS state transitions in Fuzzy Arden SyntaxCreating enumeration objects for better readability

FuzzyArden ARDS state transitions in Fuzzy Arden SyntaxExplicit indication a value is a truth value, to trigger calculation using fuzzy logic (fuzzy and, fuzzy or)

Using the sup-min composition again to determine the new degree of compatibility of the “normal” state

DiscussionFuzzy Arden Syntax for fuzzy automataRules closely resemble natural language, clinicians can verify the implemented knowledge in MLMs easily without in-depth knowledge of modern programming languagesPresentation of automaton configuration in medical, semantically rich, linguistic concepts, thus easier interpretable by cliniciansNative support for fuzzy sets and fuzzy logic in natural language concepts, improves readabilityLimitationsRepresentation of 3D fuzzy sets not (yet) intuitiveNo prospective tests so far, thus real-time performance is still unknown