Clinical applications and Arden Syntax

Educational material, part 5

Medexter HealthcareBorschkegasse 7/5 A-1090 Vienna

www.medexter.com

www.meduniwien.ac.at/kpa (academic)

Better care, patient safety, and quality assurance by Medexter, Vienna, Austria

The ArdenSuite

EHR applications

Integration into i.s.h.medat the

Vienna General Hospital

SOP checkingin melanoma patients

receiving chemotherapy

University of Colorado Health—with Epic EHR

• patient follow-up and authorization of additional inpatient services (e.g., occupational and physical therapy)

© 2014 Epic Systems Corporation. Used with permission.

Example of e-mail from HFRRS MLM to HF nurse practitioners:

Heart failure readmission risk score

Input:

• vital signs

• lab data

• demographics

• ATD info

• ICD codes

Intensive care application

CDS integration at the cardiac ICU of the Universitätsklinikum

Erlangen

• Commercial ArdenSuite CDS engine by Medexter Healthcare connected with Dräger ICM

• Data, time, and user controlled MLM processing

from Stefan Kraus, Erlangen, Germany

from Stefan Kraus, Erlangen, Germany

cAlerts—Context-sensitive, laboratory-data-based

clinical alerts

Clinical alerts at Demo EMR Demo

Context-specific alerts: C-reactive protein

Rule ID Message Context Rule

CRP_slight Slightly increased CRPPrevious value in the normal range or no previous value

20 mg/l ≤ CRP < 50 mg/l

CRP_mod Moderately increased CRPPrevious value in the normal range or no previous value

50 mg/l ≤ CRP < 100 mg/l

CRP_sign Significantly increased CRPPrevious value in the normal range or no previous value

CRP ≥ 100 mg/l

CRP_cont Further increase of CRP

Patient has no leukemia and4th+ day of infection andvalue available between 12 and 36 hours prior to the current value

CRP ≥ 20 mg/l and

CRP−yesterday′s CRPyesterday′s CRP

≥ 0.20

Note: CRP, C-reactive protein concentration in mg/l

Context-specific alerts: Leukocyte count

Rule ID Message Context Rule

LpenLeukocyte value indicates leukopenia

- 0.5 G/l < leukocyte count ≤ 3 G/l

Lcyt_leukLeukocyte value indicates leukocytosis. Admission diagnosis: leukemia

Leukemia Leukocyte count ≥ 12 G/l

Lcyt_sinc Significant increase of leukocytesLeukocyte count−previous leukocyte count

previous leukocyte count≥ 0.4

Lcyt_sdecSignificant decrease of leukocytes

Leukocyte count−previous leukocyte countprevious leukocyte count

≤ −0.4

Lcyt_incIncrease of leukocytes compared to previous finding

No leukemia andno previous infection andprevious value in the normal range.

Leukocyte count ≥ 12 G/l

Lcyt_pers Persistent leukocytosis

No leukemia and4th+ day of infection andvalue available between 12 and 36 hours prior to the current value.

Leukocyte count ≥ 12 G/l and

Leukocyte count−yesterday′s leukocyte countyesterdays′s leukocyte count

≥ −0.10

Note: Leukocyte count expressed in grams (G) per liter

Inflammation monitoring and alerts

Hepaxpert—Interpretation of hepatitis serology

test results

Automated interpretation of

hepatitis serology test results

• includes frequent, rare, as well as inconsistent combinations

• complete coverage of the problem domains

• e.g., hepatitis B serology: about 150 rules in 3 layers for 61,440 possible combinations

test results

interpretation

Integration ofHepaxpert into a

laboratory information system

Hepatitis serology interpretation integrated into an EHR

Rheumexpert—Interpretive tool for

rheumatology

Differential diagnostic supportin

rheumatology

CDS Hooks and the ArdenSuite

CDS Hooks

e.g.,ArdenSuite

e.g.,Rheumexpert

Hepaxpert

adapted from: www.cds-hooks.org; light green boxes by Medexter

e.g.,databasesopenEHR

…

openEHRand the ArdenSuite

OpenEHR connector

• extension for ArdenSuite server

• RESTful connection to openEHR server

• receive openEHR data

• use data in MLM

EhrScape

• 1st MLM • 2nd MLM

Send query & return the data • calls the 1st MLM

• analyzes blood pressure for orthostatic hypotension

Screenshot from ArdenSuite IDE

Activiti BPMN and the ArdenSuite

Clinical guideline for “Hepatitis B in pregnancy: outpatient visit”

Department of Obstetrics and Gynecology, Vienna General Hospital/Medical University of Vienna, Austria

Hepatitis B breastfeeding recommendations for HBsAg positive patients

Test results Hepatitis B PCRpositive1

Hepatitis B PCRnegative2

Hepatitis B PCR unknown

HBeAg positive Don’t breastfeed or wean from breastfeeding

Breastfeeding after immunization possible

Don’t breastfeed or wean from breastfeeding

HBeAg negative Don’t breastfeed or wean from breastfeeding

Breastfeeding after immunization possible

Breastfeeding after immunization possible

HBeAg unknown Don’t breastfeed or wean from breastfeeding

Breastfeeding after immunization possible

No recommendation possible until data is available

1 Positive result corresponds to >107 genomes/ml2 Negative result corresponds to ≤107 genomes/mlNote: HBeAg, hepatitis B envelope antigen; PCR, polymerase chain reaction

Activiti BPMN workflow and MLM calls

… part of pediatrician–mother conversation

PMML and the ArdenSuite

Predictive Model Markup Language (PMML)

• XML-based file format

• representation of statistical and data mining models

• vendor-independent method to exchange models

• supported by over 20 vendors (e.g., IBM, KNIME, R)

A. Guazzelli, What is PMML?, IBM (2010). http://www.ibm.com/developerworks/library/ba-ind-

PMML1/index.html.

Data Mining Group – PMML version 4.3, from: http://dmg.org/pmml/pmml-v4-3.html

PMML to Arden Syntax transformer

• decision trees (PMML) -> Arden Syntax (MLM)

• plugin for the ArdenSuite IDE

• input: PMML file (PMML v. 4.3)

• output: MLM file (Arden Syntax v. 2.9)

some further processing with ArdenSuite IDE, e.g., add institution

Momo—Monitoring of microorganisms

Vienna General HospitalA 1,900-bed tertiary care and teaching hospital

microbiology laboratory information system

MOLIS

reports, surveillance, and analytics

MOMO58 structured parameters

pdfs structured results

• hospital information system i.s.h.med

• clinical users• laboratory & QM• hospital hygiene & infection control• hospital administration• medical research• external EHR & EMR systems

Fuzzy Arden Syntax

From: Dupré, B. The sorites paradox. In Dupré B. (2007) 50 philosophy ideas you really need to know. QuercusPublishing PLc, London, p. 122.

Lotfi

Lotfi A. Zadeh with Klaus-Peter Adlassnig, visit at Lotfi’s home in Berkeley, CA, U.S.A., on 22 Nov 2015

Welcome!

Why Fuzzy Arden Syntax? – part I

• Modeling linguistic uncertainty by fuzzy sets

‒ due to the unsharpness (fuzziness) of boundaries in linguistic concepts; gradual transition from one concept to another

‒ modeled by fuzzy sets (e.g., fever, increased glucose level, hypoxemia)

‒ a fuzzy set calculates a degree of compatibility in the range [0,1] between raw data and a linguistic clinical concept

• Modeling propositional uncertainty by fuzzy logic

‒ due to the incompleteness of medical conclusions; uncertainty in definitional, causal, statistical, and heuristic relationships

‒ modeled by fuzzy logic truth values between zero and one (e.g., 0.6, 0.9)

– to evaluate logical combinations of clinical concepts to draw conclusions about higher-level concepts; repeated use propagates results through an inference network

Why Fuzzy Arden Syntax? – part II

• Modeling two-dimensional fuzzy sets and fuzzy automata

‒ linguistic fuzzy states represent physiological or pathophysiological states

‒ state transitions are described by linguistic instructions

‒ two-dimensional fuzzy sets include time dimension

• Modeling fuzzy control

– heuristic, linguistic control rules

– here: open-loop control cycle

Fuzzy sets

• Crisp boundary

‒ Defines a sharp threshold.

‒ Checking if a given value is greater or less

than the defined crisp threshold results in

either true or false.

‒ Borderline cases are not detected.

• Fuzzified boundary

‒ Defines a gradual transition

‒ Checking if a given value is greater or less

than the defined fuzzified boundary results in a

truth value between 0 and 1

‒ Borderline cases are detected

‒ Weighted results for borderline cases, all other

are as usual

Fuzzy sets ‒ Arden vs. Fuzzy Arden

• Classical Arden Syntax

fever_limit := 38;

temperature := 37.9;

message := "patient has no fever";

IF temperature > fever_limit THEN

msg := "patient has fever";

END IF

• Result message: “patient has no fever”

• Borderline case is not detected

• Fuzzy Arden Syntax

fever_limit := FUZZY SET (37.5,0), (38,1);

temperature := 37.9;

message := "patient has no fever";

IF temperature > fever_limit THEN

msg := "patient has fever";

END IF

• Result message: “patient has fever”

(with applicability 0.8)

• Borderline case is detected

Four clinical concepts in Moni-ICU

Clinical Concept (Unit) Fuzzy Set

Normal Range Borderline Range Pathological Range

Increased body temperature (fever) (°C)

< 37.5 37.5 – 38.0 1) > 38.0 2)

Increased C-reactive protein (CRP) (mg/dl)

< 1.0 1.0 – 6.0 3) > 6.0 3)

Leukopenia (WBC/mm3) > 5,000 4,000 – 5,000 4) < 4,000 2)

Leukocytosis (WBC/mm3) < 11,000 11,000 – 12,000 4) > 12,000 2)

1) as defined by clinicians2) as defined by CDC/NHSN, ECDC, and KISS for retrospective surveillance purposes3) as defined by clinicians; CRP is an early phase protein, useful as an “infection radar” for prospective purposes4) as defined by clinicians; white blood cell count (WBC) is a slowly reacting indicator, important for surveillance

purposes

Frequency distributions: four clinical concepts as well as the topmost HAI definitions (24,325 patient days)

Clinical Concept Absent

n (%)

Borderline

n (%)

Present

n (%)

Increased body temperature (fever) 16,074 (66.1) 3,421 (14.0) 4,830 (19.9)

Increased C-reactive protein (CRP) 4,383 (18.0) 5,841 (24.0) 14,101 (58.0)

Leukopenia 22,991 (94.5) 668 (2.8) 666 (2.7)

Leukocytosis 15,169 (62.4) 1,544 (6.3) 7,612 (31.3)

BSI or1) CRI2 or UTI-A or UTI-B 20,687 (85.0) 606 (2.5) 3,032 (12.5)

1) inclusive disjunction with precedence of “present” over “borderline” over “absent”

Two different hyperglycemia definitions

hyperglycemia (surveillance) is true is 1.00.hyperglycemia (alerting) is true is 0.75.

Moni—Monitoring of hospital-acquired infections

Moni-ICU and Moni-NICU

Monitoring (for surveillance and alerts) of healthcare-associated infections (HAIs) in

ICUs with adult patients and in NICUs with neonatal patients

Characteristics

(1) PDMSs and LISs as electronic data sources provide structured medical data

(2) medical knowledge bases containing computerized knowledge of every clinical entity involved

(3) processing algorithms evaluate, aggregate, and interpret clinical data stepwise until data can be mapped into the given HAI definitions

patient-specific alerts

infection control

natural-language definitions of nosocomial

infections

Fuzzy theories

Artificial intelligence

Monitoringof

nosocomial infections

knowledge-based systems

fuzzy sets and logic

ICUICU

microbiology

cockpit surveillance remote

clinical data

Medicine

data on microorganisms

cockpit surveillance at ward

ICU

Adlassnig, K.-P., Blacky, A. & Koller, W. (2009) Artificial-Intelligence-Based Hospital-Acquired Infection Control. In Bushko, R.G. (Ed.) Strategy for the Future of Health, Studies in Health Technology and Informatics 149, IOS Press, Amsterdam, 103–110.

A bloodstream infection—with clinical signs and growth of same skin

contaminant from two separate blood samples

BSI-A2

1

clinical_signs_of_BSI (t-1d, t, t+1d)

same_skin_contaminant_from_two_separate_blood_samples

Linguistic uncertainty defined by fuzzy sets—example: fever

feverT (t-1d) ...

fever

feverT (t)

thermoregulation applied

feverT (t+1d) ...

data import

intensive care unit

maximum value

of the day

e.g., 38.5 CC

1

037 37.5 38 38.5

linguistic HAI definitions

basic concepts:symptoms, signs, test results, clinical findings

intermediate concepts:pathophysiological states

abstraction:rules, type-1 & type-2 fuzzy sets, temporal abstraction

feature extraction:mean values, scores, …

preprocessing: missing data, plausibility, …

ICU, NICU, and microbiology patient data bases

y inference stepsreasoning

symbols

data-to-symbolconversion

raw data

x inference steps

layer n (goal)

layer 0 (start)

… ……

patient-specific cockpit&

legal reporting&

quality benchmarking

Knowledge pyramid

Advanced HAI cockpit surveillance

Second study

93 ICU patient admissions; 882 patient days; 30 HAI episodes over complete or

partial duration of stay; 76 stays with no HAI episodes

De Bruin, J.S., Adlassnig, K.-P., Blacky, A., Mandl, H., Fehre, K. & Koller, W. (2013) Effectiveness of an Automated Surveillance System for Intensive Care Unit-Acquired Infections. Journal of the American Medical Informatics Association 20(2), 369–372.

Moni-ICU

goldstandard

I+ I-

I+ 26 1

I- 4 75

30 76

HAI episodes correctly / falsely identified or missed by Moni-ICU

sensitivity = 87%• 3 false-negative pneumonias + 1 false-negative CVC-related

infection due to missing microbiology

specificity = 99%• 1 false-positive CVC-related infection because of a present

concomitant leukemia (with leukocytosis)

Acceptability

Barriers to CDS

• mental

– necessity or imperative not recognized (fatalistic attitude towards risk/suffering) no innovation

– factual incomprehension (don’t understand it) no innovation

– emotional or rational refusal (don’t want it) against innovation

– insufficient endorsement (don’t do it) against innovation

• clinical

– too simplistic or insufficient quality (lack of content quality)

– lack in workflow integration (lack of process quality)

• technical

– lack in structured patient data (documentation)

– insufficient data/semantic interoperability (data and terminology standards)

• financial

– insufficient funds (often not true!)

How to overcome these barriers?

By clinically useful solutions and—last but not least—recognizing the importance

of CDS.

Combined reasons for Moni’s success

• clinical– no diagnoses, but graded compliance with definitions

– no need for additional data entry

– two-step reporting: (1) automated generation and (2) expert verification

• methodological– pure knowledge-based system with explanatory component

– consensual classification criteria

– hierarchical layers of data and knowledge

– fuzzy set theory and logic

• technical

– separation of intensive care data collection, microbiology data collection, CDS rule engine server, knowledge packages, and web-based infection control cockpit

– integration of different hospital IT systems (intensive care IT, microbiology IT, CDS IT)

• administrative– structured intensive care data and structured data from microbiology (through Momo)

– support by hospital administration

– several clinical lead users

Arden Syntax “Lego” MLMs

Faster to your destination with Arden Syntax “Lego” MLMs

clinical and microbiological building blocks

• healthcare-associated infection monitoring and surveillance

• clinical event monitoring

Clinical Arden Syntax MLM building blocks

Clinical concept Definition

Fever Body temperature > 38 °C

Leukopenia < 4,000 WBC/mm3 blood

Leukocytosis ≥ 12,000 WBC/mm3 blood

Elevated CRP CRP >10 mg/dl blood

ShockSystolic blood pressure

Heart rate<1

Drop in BPBP value in the 37.5% percentile of all averages between systolic and diastolic BP over the last 3 days

Note: WBC, white blood cell; CRP, C-reactive protein; BP, blood pressure.

Microbiological Arden Syntax MLM building blocks

Note: CFU, colony forming unit; GNB, Gram-negative bacteria.1) Specimens taken and tested in parallel; CFU ratio between catheter and peripheral sample must be > 52) Specimens taken and tested in parallel; catheter sample must test positive at least 2 hours before withdrawal of

the peripheral blood sample.

Information element Description

Positive urine culture Urine culture with ≥ 105 CFU/ml of at most two different microorganisms

Conditional positive urine culture

1. At least two urine cultures with repeated isolation of the sameuropathogen (GNB or S. saprophyticus) with ≥ 102 colonies/ml urinein non-voided specimens

2. ≤ 105 colonies/ml of a single uropathogen (GNB or S. saprophyticus)in a patient being treated with an effective antimicrobial agent for aurinary infection

Positive catheter culture 1. Quantitative culture of a catheter tip (≥ 103 CFU / ml)2. Semiquantitative culture of a catheter tip (> 15 CFU / ml)3. Quantitative positive culture of blood drawn from a central venous

catheter; with CFU/ml reported1)

4. Quantitative positive culture of blood drawn from a peripheral vein;with CFU/ml reported1)

5. Qualitative positive culture of blood drawn from a central venouscatheter2)

6. Qualitative positive culture of blood drawn from a peripheral vein2)

7. Positive culture of pus from the insertion site

The future

Begin of exponential growth

We are here.

Kurzweil, R. The singularity is near. Viking, 2005.

Licensing options—I

Licensing options—II

Licensing options—III