CDSS for CIO 2014

Post on 22-Nov-2014

816 views 6 download

description

Clinical decision making and clinical decision support systems

Transcript of CDSS for CIO 2014

1

รศ.ทพญ.ดร. ศ�ร�วรรณ สื�บนุ�การณ�Siriwan Suebnukarn

Thammasat Universityssiriwan@tu.ac.th

Clinical Decision Making Clinical Decision Making and Clinical Decision and Clinical Decision Support SystemsSupport Systems

Part I1. Clinical decision making

2. Clinical decision support systems

Part II3. Learning from big data

4. Advanced & alternative decision support tools

3

Clinical decision making requires the clinician to apply accumulated knowledge to a specific amount patient information to produce a result that may be a diagnosis, prognosis, course of therapy, or the selection of further tests.

Too often, the decisions are based on limited knowledge, the information is incomplete or imperfect, and the decisions must be made during a limited period of time.

Clinical decision makingClinical decision making

4

Heuristic

Analytical

Clinical decision makingClinical decision making

a patient with flank pain, nausea, vomiting, and hematuria

5

• Heuristic and analytical are used when making decision.• In the emergency case, quick action based on pattern

recognition (Type 1 process) is crucial.  • Sometimes, however, it may be wrong, particularly if

other conditions aren’t evaluated and ruled out (Type 2 process). 

• For instance, a patient with flank pain, nausea, vomiting, and hematuria demonstrates the “pattern” of a kidney stone (common), but may in fact have a dissecting aortic aneurysm (uncommon).

Clinical decision makingClinical decision making

6

Clinical decision makingClinical decision making

Clinicalpractice

Evaluation

Searchevidence

Criticalappraisal

Research

Evidencecause, diagnosis,

therapeutic, prognosis

Evidence-based medicine

Problemanalysis

Clinicalquestions andexaminations

Diagnostic,therapeutic

decision

Knowledgeexperience

Hypothesesgeneration

Problem solvingProcess

Patientcircumstances,

preferences,values

Problemidentification

Intelligence phase

Design phase

Choice phase

7

Clinical decision makingClinical decision making

Intelligence phase Finding the problem Problem classification Problem decomposition Problem ownership

Design phase

Choice phase

Analytical(Optimization)

Blindsearch

Heuristicsearch

Optimal (best) Optimal (best) Good enough

All possible alternatives

are checked.

Only somealternatives

are checked.

Use algorithmsto generate

improved solutions

Find the rulesto guide

the search.

Good enough

8

Market basket analysis (Wal-Mart)– Customer identification e.g. loyalty card identifier and

or name and address – Purchase transactions e.g. what was purchased, by

who, when and the value – Product identification e.g. type or category of product

9

Clinical Decision Support SystemsClinical Decision Support Systems

CDSS is an interactive computer program

that are designed to provide expert support

for health professional

making decision

and improved health care.

WhatHowto Whom

Definition

Expert System (ES)• Software that emulates functions of an expert (is a type of artificial

intelligence [AI])

Decision Support System (DSS)• A tool that helps a user to reach a decision• Term is particularly applicable for complex semi-structured decisions

involving several factors– e.g., Should an automotive company introduce a new sport utility vehicle,

– What’s the competitive doing?,

– What would the design and re-tooling costs?,

– What do people want?

– How’s that vary as a function of petrol price?

• ES technology may be all or part of a DSS

10

Clinical Decision Support SystemsClinical Decision Support Systems

11

What is Artificial Intelligence?What is Artificial Intelligence?

is the science and engineering of making intelligent machines, especially intelligent computer programs.

12

Does AI aim at human-level intelligence?Does AI aim at human-level intelligence?

The ultimate effort is to make computer programs that can solve problems and achieve goals in the world as well as humans (Turing test).

13

Is AI about simulating human intelligence?Is AI about simulating human intelligence?

Sometimes but not always. AI researchers use methods that are not observed in people or that involve much more computing than people can do.

Why?

People fall out-of-date• Even if you read all the time, the medical literature is

growing too fast to keep up with

People have cognitive limits• Can’t accurately integrate large numbers of factors

People makes mistakes• Both lapses and due to ignorance

People aren’t always there• Hard (and expensive) to monitor 24/7

• Don’t always have the subspecialist you want

14

Clinical Decision Support SystemsClinical Decision Support Systems

15

Electronic Health Record

http://www.meaningfulusecriteria.org/

Signs, symptoms, laboratory results

Clinical Decision Support SystemsClinical Decision Support Systems

Diagnostic, therapeutic recommendations

The heart of a clinical decision support module is a method of transforming input parameters to a patient-specific

output.

16

Electronic Health Record

http://www.meaningfulusecriteria.org/Diagnostic, therapeutic recommendations

Signs, symptoms, laboratory results

KnowledgeBase

IF-THEN rules

Inferencemechanisms

Machine learning

From previous experience,

Find patterns in clinical data

Decision Support ModuleDecision Support Module

17

Data mining and knowledge discoveryData mining and knowledge discovery

data are a set of facts (for example, cases in a database),

extracting a pattern is fitting a model to data; finding structure from data; or, in general, making any high-level description of a set of data.

Manual probing of a data set is slow, expensive, and highly subjective.

18

Machine learning techniquesMachine learning techniques

MYCIN was the first CDSS to perform a significant task with performance comparable to a human expert

• Objective: identify bacteria causing severe infections, such as bacteremia and meningitis, and to recommend antibiotics, with the dosage adjusted for patient’s body weight

• PhD thesis of Shortliffe in 1970s• Never used in practice: ethical issues – what if it

made an incorrect recommendation?

19

MYCINMYCIN

20

MYCINMYCIN

de Dombal’s System (Adams et al., 1986; de Dombal et al., 1993) dealt with 16,737 patients with acute abdominal pain

• Delivered substantial evidence in 8 UK hospitals of massive savings in resources (over 8000 bed nights in two years) couple with improvement in performance (less perforated appendices, less negative laparotomy, less mortality

• Bayesian model

21

de Dombal’s Systemde Dombal’s System

22

HELPHELP

• Health Evaluation through Logical Processing (HELP) was the first hospital information system to collect patient data needed for clinical decision-making.

• The original system was developed at the LDS Hospital in Salt Lake City (UT, USA) since 1967 at the Department of Medical Informatics at the University of Utah.

• LDS Hospital is a 520 bed private acute care hospital affiliated with a parent organization known as Intermountain Health Care (IHC).

• A key feature of the system is the electronic health record (EHR).

23

HELPHELP

24

HELPHELP

• HELP can ‘data drive’ and use the knowledge base to make decisions from the data as it is stored.

• For example, physicians receive alerts directly using provider order entry (POE) of potential adverse drug events; drug-drug, drug-allergy, drug-laboratory, drug-disease, drug-dose, drug-diet and drug-interval.

• A serum potassium of 6.2 meq/l will trigger an elevated potassium alert to the nurse caring for that patient via a digital pager.

25

• A recent experiment in diagnostic data mining involved the electronic records of more than 32,000 Emergency Department patients.

• A Bayesian Network based approach was used to assemble a diagnostic system for community acquired pneumonia. The resulting system showed high accuracy with a specificity of 92.3% at a sensitivity of 95%.

HELPHELP

26

HELPHELP

The ‘Commandments’ (Bates et al., 2003) for effective CDSS:1. Speed is everything

2. Deliver in real time

3. Fit into the user’s workflow

4. Little things can make a big difference: ‘usability’

5. Recognize that physicians will strongly resist: ‘offer an alternative’

6. Changing direction is easier than stopping

7. Simple intervention works best

8. Ask for additional info only when you really need it

9. Monitor impact, get feedback, and respond

10.Manage and maintain you knowledge-based systems

27

CDSS - Success CasesCDSS - Success Cases

Alert fatigue– When clinicians are exposed to too many clinical decision

support alerts they may eventually stop responding to them.

• A threat to patient safety

– Alert fatigue is caused by poor signal-to-noise ratio because

• The alert was not serious, was irrelevant, or was shown repeatedly

– Alert fatigue can be mitigated by

• Reducing the number of alerts, prioritizing alerts, filtering superfluous alerts

28

CDSS - Success CasesCDSS - Success Cases

29

CDSS is most likely to success where it

•is well integrated with the clinical information systems infrastructure•provides useful recommendations•is supported by the organization

– Sponsored by clinical leadership– Users adequately trained– CDSS knowledge-based is maintained – current and

accurate– Running on responsive system architecture

Conclusions of Part IConclusions of Part I

30

CDSS is most likely to success where it

•is well integrated with the clinical information systems infrastructure

•provides useful recommendations

•is supported by the organization

– Sponsored by clinical leadership

– Users adequately trained

– CDSS knowledge-based is maintained

– Running on responsive system architecture

Conclusions of Part IConclusions of Part I

Part I1. Clinical decision making

2. Clinical decision support systems

Part II3. Learning from big data

4. Advanced & alternative decision support tools

32

Electronic Health Record

http://www.meaningfulusecriteria.org/

Signs, symptoms, laboratory results

Learning from Big DataLearning from Big Data

Diagnostic, therapeutic recommendations

Knowledge acquisitionKnowledge representation

33

Knowledge Acquisition and RepresentationKnowledge Acquisition and Representation

1. Human-intensive techniques

by watching and analyzing human experts

2. Data-intensive techniques

through machine learning

34

Human-intensive techniquesHuman-intensive techniques

Shortliffe and Patel, 2007

• Cognitive task analysis– Interview, semi-structured interview– Focus group, brain storming– Observation of expert in simulated or real

world environments– Think-aloud protocols: to gain insight into their

mental process

• Concept analysis– Concept mapping: labeled node-link

structures, like semantic networks

35

Human-intensive techniquesHuman-intensive techniques

36

Human-intensive techniquesHuman-intensive techniques

37

Electronic Health Record

http://www.meaningfulusecriteria.org/Diagnostic, therapeutic recommendations

Signs, symptoms, laboratory results

KnowledgeBase

IF-THEN rules

Inferencemechanisms

Machine learning

From previous experience,

Find patterns in clinical data

Knowledge Acquisition and RepresentationKnowledge Acquisition and Representation

38

The Logic represents knowledge of its domain, its goals and the current situation by sentences in logic and decide what to do by inferring that a certain action or course of action is appropriate to achieve its goals using backwards reasoning.

If Situation1 and … and Situationn then Goal

If X is green and eats flies, then X is a frog

What is X?

Knowledge-based CDSSKnowledge-based CDSS

39

Knowledge-based CDSSKnowledge-based CDSS

40

Data-intensive techniquesData-intensive techniques

41

Data-intensive techniquesData-intensive techniques

42

Data-intensive techniquesData-intensive techniques

43

Machine Learning is a branch of Artificial Intelligence concerned with the design and development of algorithms that allow computers to prediction, based on known properties learned from the training data.

Machine learningMachine learning

44

Bayesian probabilistic modelBayesian probabilistic model

A Bayesian network represents domain knowledge qualitatively by the use of graphical diagrams with nodes and arrows that represent variables and the relationships among the variables.Quantitatively, the degree of dependency isexpressed by probabilistic terms.

45

BN structure and parameter (1) Human experts provide the nodes, the arcs,

and the conditional probabilities. (2) Human experts provide the causal

relationships, the network structure is designed using this information, and the parameters can be learned from data.

(3) All machine-learned: using one of the available Bayesian network structure learning algorithms, the network structure can be learned from data as well as the parameters.

Bayesian probabilistic modelBayesian probabilistic model

46

Bayesian probabilistic modelBayesian probabilistic model

47

Bayesian probabilistic modelBayesian probabilistic model

48

Bayesian probabilistic modelBayesian probabilistic model

49

Bayesian probabilistic modelBayesian probabilistic model

50

Neural models of intelligence emphasize the brain's ability to adapt to the world in which it is situated by modifying the relationships between individual neurons.

Rather than representing knowledge in explicit logical sentences, they capture it implicitly, as a property of patterns of relationships.

Artificial Neural NetworkArtificial Neural Network

51

Artificial Neural NetworkArtificial Neural Network

ANN was carried out on a data sheet of patients presenting to an emergency department with flank pain suspicious for renal colic (Eken et al., 2009).

52

• The artificial neural network consists of interconnected nodes that form a network with variable weights between connections.

• The relationship between the input and the output of the neuron can be described as

• where xi is a input signal, wi is the weight, y is the output, b is the threshold, and f is the activation function.

Artificial Neural NetworkArtificial Neural Network

53

Papnet is a commercial neural network-based computer program for assisted screening

of Pap (cervical) smears.

Traditionally, Pap smear testing relies on the human eye to look for abnormal cells under a microscope. In fact, even the best laboratories can miss from 10% - 30% abnormal cases

Papnet-assisted reviews of cervical smears result in a more accurate screening process than the current practice.

Artificial Neural NetworkArtificial Neural Network

54

A CBR system stores cases, into a case base. Each case is composed of: -description, -solution, -outcome of applying

that solution to the problem.When a new problem is encountered, the system searches its case base for the most similar past case or cases. The solution to a similar past problem forms the basis for developing a solution to the current problem.

Case-based reasoningCase-based reasoning

55

Case-based reasoningCase-based reasoning

Case description includes – self-monitoring of blood glucose, a diabetes history, o

ccupational information, insulin sensitivity, …

Knowledge engineers met with physicians to review the patient data, identify cases, recommend therapy.

Use of case-based reasoning to enhance intensive management of patients on insulin pump therapy (Schwartz et al., 2008)

56

Decision TreesDecision Trees

A decision tree takes input as an object described by a set of properties, and outputs a yes/no decision.

57

Decision TreesDecision Trees

A decision tree prediction of the presence of major depressive disorder (Batterham et al., 2009).

Major depressive Disorder

58

Pattern recognition is the research area that studies the operation and design of systems that recognize patterns in data.

Important application areas are image analysis, character recognition, speech analysis, person identification.

Pattern RecognitionPattern Recognition

59

Pattern RecognitionPattern Recognition

60

Pattern RecognitionPattern Recognition

Support Vector Machine Classifier

Part I1. Clinical decision making

2. Clinical decision support systems

Part II3. Learning from big data

4. Advanced & alternative decision support tools

• Biosignal processing

• Business Intelligence

• Virtual reality in medicine

62

Advanced & Alternative DS ToolsAdvanced & Alternative DS Tools

Patient > Signals > Processing > Decision• Physiological instruments measure

– Health rate– Blood pressure– Oxygen saturation levels– Blood glucose– Nerve conduction

•Provide physicians with real-time data and greater insights to aid in clinical assessments

63

BBiosignal processingiosignal processing

64

Peer et al., 2003•Atrial fibrillation (AF) is the most frequent heart arrhythmia and-moreover-one of the most important risk factors for the occurrence of stroke.•In clinical diagnoses paroxysmal atrial fibrillation (PAF) often remains unrecognized if the arrhythmia is not manifest in the recorded electrocardiogram (ECG). •The system is based on an automatic ECG processing algorithm to identify patients prone to PAF.•Cardiologists use this Internet-based telemedicine application to transmit ECG signals to the analysis center, where they are processed automatically.

65

easyGeasyG-electrocardiogram analysis system Graz-electrocardiogram analysis system Graz

• RFID tags are now incredible cheap and small and can be put in anything (even in pills or flesh)– Passive RFID is activated by radio energy of a ready

(active RFID unit) and responds with it’s code number (not so different from an optical bar code)

• Can greatly improve inventory tracking– E.g. on wheelchairs, surgical instruments– Can combine with other sensors – e.g. do staff [with

RFID badges] wash hands ([approach sink’s RFID reader], depress [monitored] soap dispenser) every time coming and going from bed of patient with fever [from patient’s worn temp sensor])?

66

Radio Frequency Identifier (RFID)Radio Frequency Identifier (RFID)

67

Radio Frequency Identifier (RFID)Radio Frequency Identifier (RFID)

Kodak has filed a patent fora digestible RFID for trackingmedication compliance or digestive diagnosis

A digestible RFID in a pillhas been paired with ashoulder-worn patch.

RFID is run by electric chargefrom stomach acid to registerits ingestion.

Shoulder patch logs event and send on to remote care providers.

68

Business intelligence is the ability of HIS to quickly provide answers to questions posed by business users about the current status of the business, the business and economic trends, and the potential impact of changes in strategy and in the environment.

Business intelligenceBusiness intelligence

69

Getting data in from operational DBMS and other external sources, extracted, transformed and loaded (ETL) into the data warehouse.

Getting data out involves online analytical processing (OLAP) that provides graphical, multidimensional views for users to analyze, query, and mine the data.

Business intelligenceBusiness intelligence

70

WarehouseWarehouse

71Duke's medical records warehouse

Data warehouseData warehouse

72

A TEXTUAL DATA WAREHOUSE

There is some small amount of data in the healthcare environment that is transaction oriented and structured (payments and insurance coverage and claims).

The healthcare data warehouse consists primarily of text. The text must be integrated before being placed into the data warehouse in order for the data warehouse to make sense and be usable.

In order to execute these activities, it is necessary to have a common healthcare and medical vocabulary, ontology.

Data warehouseData warehouse

73

• Many BI systems are using – ‘scorecard’ and ‘dashboard’ methodologies and – developing Web-based query and reporting tools to

optimize delivery of services as well as improve their own data warehouse projects.

• to improve – workflow efficiency,– monitor quality and improve outcomes,– develop best practices,– optimize insurance procedures, and – uncover patterns of increased expenditures.

OOnline analytical processingnline analytical processing

74

Scorecard provides managers with the information they need to spot trends, forecast future performance, estimate whether they are on target to achieve organizational goals, and address situations before they impact the bottom line.

75

Dashboards monitor business activities in real time and allow organizations to maintain, optimize, improve business processes.

• Computer-assisted diagnosis• Treatment planning and support• Simulation and training• Tele-medicine where desktop or immersive VR

image is viewed or operated at remote location

76

Virtual Reality (VR)Virtual Reality (VR)

77

da Vinci® – Minimal invasive robotic surgery

Virtual Reality (VR)Virtual Reality (VR)

virtual reality applications for motor rehabilitation after strokeCochrane review found over 20 studies on VR and interactive video gaming as a therapy approach for stroke rehabilitation with some resulting in better arm function, improve walking speed

78

Virtual Reality (VR)Virtual Reality (VR)

79

Virtual Reality (VR)Virtual Reality (VR)

80

Conclusions•Appreciation of the areas of, and conditions for success in CDSS development•Understanding of the methods for automatic learning of decision rules and associations from large databases•Awareness of a range of decision support tools including biosignaling, BI and VR

82

ReferencesReferences