The Model-based Approach to Medical Decision Support Why and How - - PowerPoint PPT Presentation

The Model-based Approach to Medical Decision Support

Why and How ...

Peter Lucas

peterl@cs.ru.nl

Model-based System Development Section Institute for Computing and Information Sciences Radboud University Nijmegen

CIHC 21-9-2010 – p. 1/47

Medicine ∼ Engineering?

Bridge building Engineering principles Consequence of failure Medicine Clinical principles Consequence of failure

CIHC 21-9-2010 – p. 2/47

Is Support Needed?

Obstetric clinics at Vienna General Hospital mid 1800s Doctors (1st clinic) versus midwives (2nd clinic): Ignaz Semmelweis (1818–1865): infection after child birth can be drastically cut by hand washing

CIHC 21-9-2010 – p. 3/47

Today · · ·

Hand hygiene in the intensive care unit: prospective observations of clinical practice

Pol Arch Med Wewn, 2008; 118 (10): 543-547 Ismael A. Qushmaq, Diane Heels-Ansdell, Deborah J. Cook, Mark B. Loeb, Maureen O. Meade

• Abstract. INTRODUCTION: Adherence to hand hygiene

recommendations in the intensive care unit (ICU) is variable and moderate, at best. OBJECTIVES: To measure adherence to hand hygiene recommendations among ICU clinicians in a prospective

• bservational study in 6 multidisciplinary ICUs among 4
• hospitals. . . . RESULTS: The rate of adherence to current

recommendations was 20%. . . .

CIHC 21-9-2010 – p. 4/47

Other Facts

From a recent study (Arch Intern Med. 2010; 170(12): 1015-1021): Diagnostic errors often result in patient harm Structured review study of 7926 patient records

• f 21 hospitals across the Netherlands

Results: diagnostic adverse events occurred in 0.4% of hospital admissions (6.4% of all adverse events) 83.3% were judged to be preventable. Human failure was the main cause (96.3%) the consequence was a higher mortality than for other adverse events (29.1% vs 7.4%)

CIHC 21-9-2010 – p. 5/47

Protocols and Guidelines

2002 Centers for Disease Control and Prevention Guidelines for the prevention of intravascular catheter-related infections:

Wash your hands before inserting a central venous catheter Clean the skin with chlorhexidine Use of full-barrier precautions during CVC insertion Avoid the femoral site Remove unnecessary central venous catheters ⇒ This guideline is clearly ignored

CIHC 21-9-2010 – p. 6/47

Clinical Guidelines

Definition: clinical (practice) guidelines: systematically developed statements to assist practitioners and patients decisions about appropriate health care in specific clinical circumstances Characteristics: Guidelines are based on scientific evidence (results from RCTs for example — evidence-based medicine) In conjunction with considerations such as safety, availability, and cost effectiveness Aim: improving health-care outcomes and reduce costs of care

CIHC 21-9-2010 – p. 7/47

NICE (www.nice.org.uk)

National Institute for health and Clinical Excellence

CIHC 21-9-2010 – p. 8/47

Example: NICE DM2 Guideline

DM2 GL: ORAL GLUCOSE CONTROL THERAPIES (2): Thiazolidinediones (glitazones) R40 If glucose concentrations are not adequately controlled (to HbA1c <7.5% or other higher level agreed with the individual), consider, after discussion with the person, adding a thiazolidinedione to: the combination of metformin and a sulfonylurea where insulin would otherwise be considered but is likely to be unacceptable or of reduced effectiveness because of: employment, social or recreational issues related to putative hypoglycaemia barriers arising from injection therapy or . . . ... a sulfonylurea if metformin is not tolerated

CIHC 21-9-2010 – p. 9/47

Which Decision Support is Best?

Protocols and guidelines: Evidence based (reflect scientific evidence) Have been shown to have a positive effect on quality of care in some cases Non-interactive, often very lengthy textual documents (with fixed structure) Are hard to personalise Decision-support systems: Interactive, and allow exploring clinical problems

• f individual patients

Offer one or more problem solving modes Can incorporate ideas from guidelines

CIHC 21-9-2010 – p. 10/47

Computer-based Guidelines

CIHC 21-9-2010 – p. 11/47

Start: Clinical Reasoning

signs diagnostic process test

therapy selection

patient data diagnosis

medical knowledge

patient data therapy

prediction disease progress

prognosis

CIHC 21-9-2010 – p. 12/47

Its Computerisation: Not Easy

Early academic AI attempts, e.g.: Diagnosis and treatment of sepsis using rule-based system: MYCIN (1974–1979) Diagnosis of disorders in internal medicine (e.g., gastrointestinal, rheumatoid, endocrine disorders): INTERNIST-I (1975–1985) Diagnosis of glaucoma by Causal ASsociationel NETwork: CASNET (1971–1978) Commercial AI attempts: Quick Medical Reference (QMR) – based on INTERNIST-I (discontinued 2001) DXplain (1984–) – http://dxplain.org

CIHC 21-9-2010 – p. 13/47

Why Failure?

Focus on diagnostic systems: after entering set of findings ⇒ differential diagnosis First generation programs: immature technology, PhD projects Don’t offer the support clinicians want to have Computational infrastructure too primitive until 2000 Clinicians had little computer literacy until ±1995 No integration with electronic patient record systems (still not generally available) Bad computer inferface

CIHC 21-9-2010 – p. 14/47

Knowledge Formalisation

Ingredients (knowledge representation): Uncertainty (probability theory) and decision theory Intuitive qualitative notions, such as: causal relations associations actions

• utcomes

justification · · · ⇒Probabilistic graphical models, such as Bayesian networks, and influence diagrams offer a good start

CIHC 21-9-2010 – p. 15/47

The Model-based Approach

Management (diagnosis, treatment, prognosis) can be formalised: meta-model, e.g., What is a diagnosis? What is a prognosis, etc. Medical knowledge is also modelled (object model) Deployment of: probabilistic graphical models, in particular Bayesian networks logical methods

CIHC 21-9-2010 – p. 16/47

Problem Solving – Probabilistic

A diagnosis d∗ is maximum a posteriori assignment d∗ = argmaxdP(d | e), where e

• bserved evidence (symptoms, test results)

Prognostic reasoning; determine outcome o: P(o | e, a), with a a sequence of treatment actions Optimal treatment: a∗ ∈ argmaxa

• P(o | e, a)u(a, o, e)

Pretreatment

• bservations

Treatments Outcome Pretreatment

• bservations

Treatments Outcome U

CIHC 21-9-2010 – p. 17/47

Have you got Mexican Flu?

P(m, c, s) = 0.009215 P(m, ¯ c, s) = 0.000485 P(m, c, ¯ s) = 0.000285 P(m, ¯ c, ¯ s) = 1.5 · 10−5 P( ¯ m, c, s) = 9.9 · 10−6 P( ¯ m, ¯ c, s) = 0.0098901 P( ¯ m, c, ¯ s) = 0.0009801 P( ¯ m, ¯ c, ¯ s) = 0.97912 M: mexican flu; C: chills; S: sore throat Probability of mexican flu and sore throat? Probability of mexican flu given sore throat?

CIHC 21-9-2010 – p. 18/47

Have you got Mexican Flu?

P(m, c, s) = 0.009215 P(m, ¯ c, s) = 0.000485 P(m, c, ¯ s) = 0.000285 P(m, ¯ c, ¯ s) = 1.5 · 10−5 P( ¯ m, c, s) = 9.9 · 10−6 P( ¯ m, ¯ c, s) = 0.0098901 P( ¯ m, c, ¯ s) = 0.0009801 P( ¯ m, ¯ c, ¯ s) = 0.97912 M: mexican flu; C: chills; S: sore throat Probability of mexican flu and sore throat? 0.0097 Probability of mexican flu given sore throat? 0.495

CIHC 21-9-2010 – p. 18/47

Probabilistic Reasoning

Joint probability distribution P(X1, X2, . . . , Xn) marginalisation: P(Y ) =

• Z

P(Y, Z), with X = Y ∪ Z conditional probabilities: P(Y | Z) = P(Y, Z) P(Z) Bayes’ theorem: P(Y | Z) = P(Z | Y )P(Y ) P(Z)

CIHC 21-9-2010 – p. 19/47

Probabilistic Reasoning (cont)

Examples:

P(m, s)=P(m, c, s)+P(m, ¯ c, s)=0.009215+0.000485=0.0097 P(m | s)=P(m, s)/P(s)=0.0097/0.0196=0.495

Note that: Mainly interested in conditional probability distributions: P(Z | E) = P E(Z) for (possibly empty) evidence E (instantiated variables) Tendency to focus on conditional probability distributions of single variables Many efficient reasoning algorithms exist

CIHC 21-9-2010 – p. 20/47

Bayesian Networks

flu (FL) (yes/no) SARS (RS) (yes/no) fever (FE) (yes/no) dyspnoea (DY) (yes/no) TEMP (≤ 37.5/> 37.5) VisitToChina (CH) (yes/no) P(CH, FL, RS, DY, FE, TEMP) P(FL = y) = 0.1 P(CH = y) = 0.1 P(RS = y | CH = y) = 0.3 P(RS = y | CH = n) = 0.01 P(FE = y | FL = y, RS = y) = 0.95 P(FE = y | FL = n, RS = y) = 0.80 P(FE = y | FL = y, RS = n) = 0.88 P(FE = y | FL = n, RS = n) = 0.001 P(DY = y | RS = y) = 0.9 P(DY = y | RS = n) = 0.05 P(TEMP ≤ 37.5 | FE = y) = 0.1 P(TEMP ≤ 37.5 | FE = n) = 0.99

CIHC 21-9-2010 – p. 21/47

Evidence Propagation

Nothing known:

NO YES

FLU

no yes

FEVER

no yes

SARS

no yes

VisitToChina

no yes

DYSPNOEA

<=37.5 >37.5

TEMP

Temperature >37.5 ◦C:

NO YES

FLU

no yes

FEVER

no yes

SARS

no yes

VisitToChina

no yes

DYSPNOEA

<=37.5 >37.5

TEMP CIHC 21-9-2010 – p. 22/47

Project I: Pneumonia in ICU

ICU at Utrecht MC Diagnosis and antimicrobial treatment

• f patients with

ventilator-associated pneumonia (VAP) About 15-20% of ICU patients develop VAP Mortality rate: up to 40% Up to 50% of antibiotics in ICUs are prescribed for airway infections

CIHC 21-9-2010 – p. 23/47

Software Infrastructure

DHTML Variable−value pairs triples Data SQL Variable−value−probability

Bayesian Network

PHP Module HTTP Server

Apache

Web Browser

CPR Reasoning System

CIHC 21-9-2010 – p. 24/47

Global Model Pneumonia

hospitalisation colonisation aspiration mechanical ventilation immunological status pneumonia symptoms signs, lab antimicrobial therapy side effects

• rganism

susceptibility coverage

CIHC 21-9-2010 – p. 25/47

Detailed Pneumonia Network

colonisation pneumonia colonisation PA colonisation HI colonisation SP pneumonia PA pneumonia HI pneumonia SP pneumonia symptoms signs lab hospitalisation aspiration mechanical ventilation immunological status coverage

• verall

coverage Antibiotics

CIHC 21-9-2010 – p. 26/47

Prediction

p s m c ne1e2 s h p l hosp.

• mech. vent.

<5 days IC 0-24 hours 0.00 0.25 0.50 0.75 1.00 p s m c ne1e2 s h p l hosp.

• mech. vent.

<5 days IC 24-48 hours 0.00 0.25 0.50 0.75 1.00 p s m c ne1e2 s h p l hosp.

• mech. vent.

<5 days IC 48-96 hours 0.00 0.25 0.50 0.75 1.00 p s m c ne1e2 s h p l hosp.

• mech. vent.

≥5 days IC 0-24 hours 0.00 0.25 0.50 0.75 1.00 p s m c ne1e2 s h p l hosp.

• mech. vent.

≥5 days IC 24-48 hours 0.00 0.25 0.50 0.75 1.00 p s m c ne1e2 s h p l hosp.

• mech. vent.

≥5 days IC 48-96 hours 0.00 0.25 0.50 0.75 1.00 : Pr(pneumonia) = 1.0 : Pr(pneumonia) = ?

CIHC 21-9-2010 – p. 27/47

Specification of Interactions

Compact specification: probability tables P(Xi | pa(Xi)) can still be large even when taking into account independence information Easy way to map domain knowledge to entries into a probability table Way to use qualitative knowledge about interactions as constraints to probabilistic information Might be useful in developing applications

CIHC 21-9-2010 – p. 28/47

Qualitative Modelling

Infection Body response to A Body response to B Body response to C Colonisation by bacterium A Colonisation by bacterium B Colonisation by bacterium C Fever WBC ESR

People become colonised by bacteria when entering a hospital, which may give rise to pneumonia

CIHC 21-9-2010 – p. 29/47

Bayesian-network Modelling

Qualitative causal modelling Cause → Effect

Inf BRA BRB BRC

Quantitative interaction modelling P(Inf | BRA, BRB, BRC)

BRA t f BRB BRB t f t f BRC BRC BRC BRC Inf t f t f t f t f t 0.8 0.6 0.5 0.3 0.4 0.2 0.3 0.1 f 0.2 0.4 0.5 0.7 0.6 0.8 0.7 0.9

CIHC 21-9-2010 – p. 30/47

Causal Independence

C1 C2 . . . Cn I1 I2 . . . In E f conditional independence interaction function

P(e | C1, . . . , Cn) =

• I1,...,In

P(e | I1, . . . , In) ×

n

• k=1

P(Ik | Ck) =

• f(I1,...,In)=e

n

• k=1

P(Ik | Ck) Note: P(ik | ¯ ck) = 0 – absent causes don’t contribute

CIHC 21-9-2010 – p. 31/47

Symmetric Boolean Functions

Order of arguments doesn’t matter; defined in terms

• f exact function ek:

f(I1, . . . , In) =

n

• k=0

ek(I1, . . . , In) ∧ γk where γk are Boolean constants only dependent of the function f Example: threshold function τl: τl(I1, . . . , In) =

n

• k=l

ek(I1, . . . , In)

CIHC 21-9-2010 – p. 32/47

Decomposition by Counting

Threshold function τ3:

no yes

O3

1 2 3

O2

1 2

O1

no yes

S2

no yes

S1

no yes

S3

no yes

S4

no yes

O3

1 2 3

O2

1 2

O1

no yes

S2

no yes

S1

no yes

S3

no yes

S4

CIHC 21-9-2010 – p. 33/47

Qualitative Modelling

Overall Coverage

• f A

Coverage

• f B

Coverage

• f C
• Col. by A
• Col. by B
• Col. by C

Antibiotic

By antibiotic treatment M clinicians try to cover O at most 2 of the bacteria giving rise to pneumonia P(O | C1, . . . , Cn, M)

CIHC 21-9-2010 – p. 34/47

Overall Susceptibility

Pτk(o|C1, . . . , Cn, M)=

• k≤l≤n
• el(S1,...,Sn)

n

• j=1

P(Sj | Cj, M)

C1 C2 S1 S2 O M

Cj: causal factor j Sj susceptibility to medication M: treatment by antimicrobial medication O: overall outcome

CIHC 21-9-2010 – p. 35/47

Various Models

Conditional probability distributions: P(Sj | Cj, M) susceptibility I model: P(sj | Cj, M) = 0 if Cj = yes, M = no 1 otherwise susceptibility II model: P(si | ¬ci, ¬m) = 1, whereas P(si | ¬ci, m) = 0 susceptibility III model: P(sj | Cj, M) = 1 if Cj = yes, M = yes 0 otherwise

CIHC 21-9-2010 – p. 36/47

Model I, Colonised by 1

NO YES

Col_B

NO YES

Suscep_B

YES NO

Medication

NO YES

Suscep_C

NO YES

Col_C

NO YES

Coverage_OR

NO YES

Suscep_A

NO YES

Col_A

NO YES

Coverage_2

NO YES

Coverage_AND

NO YES

Col_B

NO YES

Suscep_B

YES NO

Medication

NO YES

Suscep_C

NO YES

Col_C

NO YES

Coverage_OR

NO YES

Suscep_A

NO YES

Col_A

NO YES

Coverage_2

NO YES

Coverage_AND

CIHC 21-9-2010 – p. 37/47

Model II, Colonised by 1

NO YES

Col_B

NO YES

Suscep_B

YES NO

Medication

NO YES

Suscep_C

NO YES

Col_C

NO YES

Coverage_OR

NO YES

Suscep_A

NO YES

Col_A

NO YES

Coverage_2

NO YES

Coverage_AND

NO YES

Coverage_AND

NO YES

Suscep_C

YES NO

Medication

NO YES

Suscep_B

NO YES

Col_B

NO YES

Coverage_2

NO YES

Suscep_A

NO YES

Col_A

NO YES

Coverage_OR

NO YES

Col_C

CIHC 21-9-2010 – p. 38/47

Model III, Colonised by 1

NO YES

Col_B

NO YES

Suscep_B

YES NO

Medication

NO YES

Suscep_C

NO YES

Col_C

NO YES

Coverage_OR

NO YES

Suscep_A

NO YES

Col_A

NO YES

Coverage_2

NO YES

Coverage_AND

NO YES

Col_B

NO YES

Suscep_B

YES NO

Medication

NO YES

Suscep_C

NO YES

Col_C

NO YES

Coverage_OR

NO YES

Suscep_A

NO YES

Col_A

NO YES

Coverage_2

NO YES

Coverage_AND

CIHC 21-9-2010 – p. 39/47

Model II, Colonised by 2

NO YES

Col_C

NO YES

Suscep_C

YES NO

Medication

NO YES

Suscep_B

NO YES

Col_B

NO YES

Coverage_2

NO YES

Suscep_A

NO YES

Col_A

NO YES

Coverage_OR

NO YES

Coverage_AND

NO YES

Col_C

NO YES

Suscep_C

YES NO

Medication

NO YES

Suscep_B

NO YES

Col_B

NO YES

Coverage_2

NO YES

Suscep_A

NO YES

Col_A

NO YES

Coverage_OR

NO YES

Coverage_AND

CIHC 21-9-2010 – p. 40/47

Project II: Medical Images

national breast cancer screening programme decision-making under uncertainty interpretation of image features in terms of probabilistic graphical models from single- to multi-view interpre- tation

CIHC 21-9-2010 – p. 41/47

Singleview CAD System

Region features: contrast, size, location, margin, spiculation, etc. Advantage: a good detection rate per image Shortcoming: unsatisfactory performance at a patient level because views are treated independently

CIHC 21-9-2010 – p. 42/47

Multiview Interpretation

Mediolateral

• blique view

Craniocaudal view

View–A

View–B A1 B2 B1 A2 L11 L12 L22 L21

CIHC 21-9-2010 – p. 43/47

Multiview Bayesian Network

Ai / Bj = (x1, x2, …, xn)

A1 L11 CA1 A2 B1 B2 CA1 IA1 IA2 View-A IB1 IB2 a) RegNet b) ViewNet L12 L21 L22 CB1 CB2 CA2 CA2 CB1 CB2 View-B

Interpretation of regions of interest (real-valued feature vector): logistic regression Combination of region and view information: causal independence

CIHC 21-9-2010 – p. 44/47

Project III: Self Management

PATIENT PHYSICIAN MEASURE

BLOOD PRESSURE

COMPLETE

PATIENT’S DATA

BAYESIAN

NETWORK MODEL

AUTOMATIC

PATIENT’S DATA UPLOAD

COMMUNICATE

CURRENT PATIENT’S RISK FOR PE

HOSPITAL SERVER BAYESIAN

NETWORK MODEL

COMPUTE AND REPORT

CURRENT PATIENT’S RISK FOR PE

HOSPITAL

CONTROL

STORE CONTROL

MEASUREMENTS

AUTOMATIC

RETRIEVAL OF MEASUREMENTS VIA SMS

COMPUTE AND REPORT

CURRENT PATIENT’S RISK FOR PE

CIHC 21-9-2010 – p. 45/47

Android Smart Phone

Starting screen BP input BN output

CIHC 21-9-2010 – p. 46/47

Conclusions & Future Research

Clinicians and patient need support to explore problems: what if the patient is treated in this way? what if this diagnostic test is omitted? · · · Reasoning should include uncertainty (= available scientific evidence from data and literature) & logic (plans) Bayesian networks are a good start; probabilistic logics offers an even richer set of languages Prognosis is our current application area

CIHC 21-9-2010 – p. 47/47