Clinical decision support
Klaus-Peter Adlassnig
Section for Medical Expert and Knowledge-Based SystemsCenter for Medical Statistics, Informatics, and Intelligent SystemsMedical University of ViennaSpitalgasse 23, A-1090 Vienna
www.meduniwien.ac.at/kpa
Einführung in die Medizinische Informatik, WS 2015/16, 04. November 2015
Computers in clinical medicine—steps of natural progression
• step 1: patient administration– admission, transfer, discharge, and billing
• step 2: documentation of patients’ medical data– electronic health record: all media, distributed, life-long (partially fulfilled)
• step 3: patient and hospital analytics– data warehouses, quality measures, reporting and research databases,
patient recruitment… population-specific
• step 4: clinical decision support (applying knowledge to data)– safety net, quality assurance, evidence-based
… patient-specific
Medical research
biomolecular researchmedical statistics
clustering and classification data and knowledge mining
consensus conferences
factual/causal, definitional, statistical, and heuristic knowledge
molecular biomedicine
medical knowledge
facts consensus generalization
medical studies
evidence-based medicine
Patient care
decision-oriented analysis and interpretation
of patient data
factual/causal, definitional, statistical, and heuristic knowledge
human-to-human human-to-humanclinical decision support
patient-physician encounter
patient-physician follow-up
medical knowledge modules
medical knowledge
Medical information and knowledge-based systems
symptomssigns
test resultsclinical findings
biosignalsimages
diagnosestherapies
nursing data
•••
standardizationtelecommunication
chip cards
anatomybiochemistryphysiology
pathophysiologypathologynosology
therapeutic knowledgedisease management
•••
subjective experienceintuition
knowledge-based systems
patient’s medical data physician’s medical knowledge
medical statisticsclustering & classificationdata & knowledge mining
machine learning
clinical decision supportmedical expert systems
manypatients
single patient
diagnosistherapy
prognosismanagement
generalknowledge
•••
generalknowledge
telemedicine telemedicineintegration
information systems
induction
deduction
Clinical medicine
medical guidelines
history data
symptoms
signs
laboratorytest results
biosignals
images
genetic data
prognosispatient patient
symptomatic therapy
differentialtherapy
differentialdiagnosis
symptoms
signs
test results
findings
examination subspecialities clinic
medication history
radiological diagnosis
laboratorydiagnosis
…
Clinical medicine
medical guidelines
history data
symptoms
signs
laboratorytest results
biosignals
images
genetic data
prognosispatient patient
symptomatic therapy
differentialtherapy
differentialdiagnosis
symptoms
signs
test results
findings
examination subspecialities clinic
medication history
radiological diagnosis
laboratorydiagnosis
…
QMRDxPlainCADIAG
MYCIN
…
ANNs
SVMs
…
…
personalizedmedicine
+
+
Clinical medicine: high complexity• sources of medical knowledge
‒ factual/causal‒ definitional‒ statistical‒ heuristic
• layers of medical knowledge‒ observational and measurement level‒ interpretation, abstraction, aggregation, summation‒ pathophysiological states‒ diseases/diagnoses, therapies, prognoses, management decisions
• imprecision, uncertainty, and incompleteness‒ imprecision (=fuzziness) of medical concepts
* due to the unsharpness of boundaries of linguistic concepts‒ uncertainty of medical conclusions
* due to the uncertainty of the occurrence and co-occurrence of imprecise medical concepts‒ incompleteness of medical data and medical theory
* due to only partially known data and partially known explanations for medical phenomena• “gigantic” amount of medical data and medical knowledge
‒ patient history, physical examination, laboratory test results, clinical findings‒ symptom-disease relationships, disease-therapy relationships, …‒ terminologies, ontologies: SNOMED CT, LOINC, UMLS, …
specialization, teamwork, quality management, computer support
Clinical medicine: Hidden challenges
• holistic diagnosis– logical conclusions, evidence-based knowledge, and practical experience– intuition and pattern matching– patient‘s non-formalizable/non-digitizable data
• probable vs. possible diagnoses– suspected diagnosis, clinical diagnosis, pathological diagnosis– most probable diagnosis vs. possible diagnoses– limits of investigation, invasiveness, costliness
• terminology in context– not every diagnostic term is a diagnosis– surveillance vs. alert vs. clinical diagnosis– clinical diagnosis vs. discharge diagnosis
Clinical decision support: Definitions
• Foundational: Key origin of field of biomedical informatics– AIM = artificial intelligence in medicine– computer-based diagnosis in the heyday of AI
• Now: Intelligent assistant– support/assist human decision makers, not supplant them
• Core: Applying knowledge to data
Miller RA. Medical diagnostic decision support systems—past, present and future: a threaded bibliography and brief commentary. Journal of the American Medical Informatics Association 1994;1:8-27.
• studies in Colorado and Utah and in New York (1997)
– errors in the delivery of health care leading to the death of as many as 98,000 US citizens annually
• causes of errors– error or delay in diagnosis– failure to employ indicated tests– use of outmoded tests or therapy– failure to act on results of testing or
monitoring– error in the performance of a test, procedure,
or operation– error in administering the treatment – error in the dose or method of using a drug– avoidable delay in treatment or in responding
to an abnormal test– inappropriate (not indicated) care– failure of communication– equipment failure
• prevention of errors– we must systematically design safety into
processes of care
errors
prevention
Clinical decision support for patient safety and quality assurance
patients’ structured medical data: EHRs (local, national), Apps, …
hospital management and quality benchmarking• evidence-based reminders and processes• computerized clinical guidelines, protocols, standard
operating procedures• healthcare-associated infection surveillance
prognostic prediction• illness severity scores, prediction rules• trend detection and visualization
therapy advice• drug alerts, reminders, calculations
– indication, contraindications, redundant medications, substitutions
– dosage calculations, drug-drug interactions, adverse drug events
• management of antimicrobial therapies, susceptibility and resistance rates
• open- and closed-loop control systems
diagnostic support• alerts, reminders, to-do lists• clinical interpretation, (tele)monitoring• differential diagnostics
– rare diseases, rare syndromes– further or redundant examinations – diagnostic completeness (multi-morbidity)
• consensus-criteria-based evaluations– disease classification and surveillance criteria
structured medical knowledge: rules, tables, trees, guidelines, scores, algorithms, …
T
Clinical decision support: Five rights
• Framework for approaching & configuring CDS interventions
• “Rights” − right information delivered to the− right person in the− right intervention format through the− right channel at the− right point in workflow
Interpretation of
hepatitis serology test results
test results
interpretation
ORBIS Experter: 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 more than 61,000 possible combinations
Differentialdiagnose rheumatischerErkrankungen
ComputergestützteEntwöhnung vom
Respirator
University of Colorado Health—with Epic EMR
• 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 (HFRRS)
Input:
• vital signs• lab data• demographics• ATD info• ICD codes
Integration into i.s.h.medat the
Vienna General Hospital
SOP checkingin melanoma patients
receiving chemotherapy
ArdenSuite integration with ICM by Dräger
• Data-, time-, and user-controlled execution of MLMs
• Application-specific viewer
by Stefan Kraus
by Stefan Kraus
Use Case: Hypoglycemia
DATA:
LET glucose BE READ {…glucose…};
LET physician_DECT BE DESTINATION {sms:26789};
LOGIC:
IF LATEST glucose IS LESS THAN 50 THEN
CONCLUDE true;
ENDIF;
ACTION:
WRITE „Warning…“ AT physician_DECT; by Stefan Kraus
CONCLUDE TRUE Do something
Hypoglycemia alert via DECT cordless telecommunications
Event monitors are
“tireless observers,constantly monitoringclinical events”
George Hripcsak
by Stefan Kraus
Arden Syntax: HL7- and ANSI-approved
• A standard language for writing situation-action rules, procedures, or knowledge bases that compute results based on clinical events detected in patient data
continuous development since 1989
• Each module, referred to as a medical logic module (MLM), contains sufficient knowledge to make a single decision
• Medical knowledge packages (MKPs) consist of interconnected MLMs for complex clinical decision support
• The Health Level Seven Arden Syntax for Medical Logic Systems, version 2.9—including fuzzymethodologies—was approved by Health Level Seven (HL7) International and the American National Standards Institute (ANSI) in 2013
• Version 2.10—including ArdenML, an XML-based representation of Arden Syntax MLMs—was approved in 2014
⇒ healthcare industry and academic users
General MLM LayoutMaintenance Category Library Category Knowledge Category Resources Category
Identify an MLMData TypesOperators
Basic OperatorsCurly Braces List OperatorsLogical OperatorsComparison OperatorsString OperatorsArithmetic OperatorsOther Operators
Control StatementsCall/Write Statements and Trigger
Sample MLM (excerpt)
ArdenSuite: Arden-Syntax-based genuine technology platform for clinical decision support (CDS)
Interface possibilities:1) Web services for MLM calling and for data transfer2) Web services for MLM calling and server/database
connector for data access3) Data warehouse + ArdenSuite server =
autonomous CDS system
ArdenSuite server and software components• web-services-based
ArdenSuite server including‒ ArdenSuite engine‒MLM manager‒ XML-protocol-based
interfaces, e.g., SOAP, REST, and HL7
‒ a project-specific dataand knowledge servicescenter
• Java libraries ‒ ArdenSuite compiler ‒ ArdenSuite engine
• ArdenSuite integrated development and test environment (IDE) including‒Medical logic module
(MLM) editor and authoring tool
‒ ArdenSuite compiler(syntax versions 2.1, 2.5, 2.6, 2.7, 2.8, 2.9, and 2.10)
‒ ArdenSuite engine‒MLM test environment‒MLM export component
• command-line ArdenSuite compiler data warehouse
– selected data and results, e.g., ICU & NICU, microbiology, MONI– reporting, quality measures, and benchmarking– study support and recruitment– App docking station (e.g., through FHIR server) – data and knowledge mining (big data)
Fuzzy Arden Syntax: Modeling uncertainty in medicine
• linguistic uncertainty
‒ due to the unsharpness (fuzziness) of boundaries of linguistic concepts; gradual transition from one concept to another
‒ modeled by fuzzy sets (e.g., fever, increased glucose level, hypoxemia)
• propositional uncertainty
‒ due to the incompleteness of medical conclusions; uncertainty in definitional, causal, statistical, and heuristic relationships
‒ here: modeled by truth values between zero and one (e.g., 0.6, 0.9)
Crisp sets vs. fuzzy sets
age
1
χY young
0 threshold
0 age
1
µY young
threshold0
0
“arbitrary” yes/no decisions• cause of unfruitful
discussions• often simply wrong
“intuitive” gradual transitions
Examples of fuzzy sets as they are applied in Moni-ICU
DoCfever
0.81.0
37.5 38.0 ºC0
37.9
DoC
4,000 5,000 11,000 12,000 WBC/mm3
leukopenia leukocytosis
0
1.0
11,500
0.5
DoC
1.0 1.3 shock index =systolic blood pressure/
heart rate
shock present1.0
0
0.9
Clinical concepts and relationships between them
( ) Dt3S2S1S →¬∨∧
truth value
degree of compatibility
𝑆𝑆1: fever𝑆𝑆2: hypotension𝑆𝑆3: leukopenia𝑆𝑆4: leukocytosis𝑆𝑆5: increased CRP𝑆𝑆6: inflammatory signs (with sepsis)
Uncertainty in conclusions I: through linguistic uncertainty in premises
Example:
𝑆𝑆1 ∨ 𝑆𝑆2 ∨ 𝑆𝑆3 ∨ 𝑆𝑆4 ∨ 𝑆𝑆51.0
𝑆𝑆6 is true is 0.8.
38.037.9
37.5
fever
°C
1.0
0.8
0.0
degreeof
compatibility
Uncertainty in conclusions II: through uncertainty in propositions
Example 1:
𝑆𝑆10.8
𝐷𝐷1 𝑆𝑆1 : highly increased amylase𝐷𝐷1: acute pancreatitis
Example 2:
𝐼𝐼10.8
𝑆𝑆1 𝐼𝐼1 : thermoregulation (cooling)𝑆𝑆1: fever
Example 3:
at least 4/11: 𝑆𝑆1 … 𝑆𝑆110.75
𝐷𝐷1 𝑆𝑆1: morning stiffness lasting at least one hour
𝑆𝑆5: symmetric joint involvement
𝑆𝑆9: positive serum rheumatoid factor
𝐷𝐷1: rheumatoid arthritis
……
…
Two different hyperglycemia definitions
Hyperglycemia (surveillance) is true is 1.00.Hyperglycemia (alerting) is true is 0.75.
Towards a science of clinical medicine
patient’s medical data and
healthcare processes for
machine processing
patient’s medical data and
healthcare processes for
human processing
≠
“Measure what is measurable, and make measurable what is not so.”Galileo Galilei
1564–1642
Crucial point in clinical medicine:
“Digitize what is digitizable, and make digitizable what is not so.”Klaus-Peter Adlassnig
observations measurementse.g., temperature chart e.g., CRP
skin color (jaundice, livid, …) color measurement… …
Challenges to clinical decision support (CDS)
• mental– necessity or imperative not recognized (fatalistic attitude towards risk/suffering)– factual incomprehension (don’t understand it)– emotional refusal (don’t want it)– insufficient endorsement (don’t do it)
• 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.
Literature on “Clinical Decision Support”: 36,211 publications
0
500
1000
1500
2000
2500
3000
3500
4000
1960 1970 1980 1990 2000 2010 2020Max Plischke, 2015
Challenge for CDS: Explosion in data + knowledge
Stead WW, Searle JR, Fessler HE et al. Biomedical informatics: Changing What Physicians Need to Know and How They Learn. Academic Medicine 2011; 86(4):429-434.
A “holy grail” of clinical informatics is scalable, interoperable clinical decision support.
according to
Kensaku Kawamoto
HL7 Work Group Meeting,
San Diego, CA, September 2011
Clinical decision support: Ten commandments
• Speed is everything• Anticipate needs & deliver in real time (alert)• Fit into user workflow (five rights)• Little things make a big difference (e.g., screen design, used terms)• Recognize that MDs will resist stopping (e.g., medication)• Changing direction is easier than stopping (e.g., dosing)• Simple interventions work best• Ask for information only if you really need it (avoid additional data input)• Monitor impact, get feedback and respond (keep the user informed)• Manage & maintain your KBS (continues improvement)
Bates DW, Kuperman GJ, Wang S et al. Ten Commandments for Effective Clinical Decision Support: Making the Practice of Evidence-Based Medicine a Reality. Journal of the American Medical Informatics Association 2003;10:523-30.
Regulatory framework for clinical decision support software: Present uncertainty and prospective proposition
From Y. Tony Yang and Bradley Merrill Thompson (2015) Journal of the American College of Radiology.