Vous êtes sur la page 1sur 46

Population health thinking with Bayesian networks

Decision-making, causal inference, and probabilistic reasoning

Tomás J. Aragón, MD, DrPH


UCSF Seminar for EPI T32 Scholars, San Francisco, CA, June 6, 2018
Health Officer, City & County of San Francisco
Director, Population Health Division (PHD)
San Francisco Department of Public Health
https://taragonmd.github.io/ (GitHub page)

PDF slides produced in Rmarkdown LATEX Beamer—Metropolis theme

1
1. Background and introduction

2. Some concepts and examples

3. Ideas for the future

2
Acknowledgments (in alphabetical order by first name)

1. Decision Quality: Value creation from better business


• Carl Spetzler, Hanna Winter, decisions. Wiley, 2016.
Jennifer Mayer1 2. Thinking, Fast and Slow. New York: Farrar, Straus and
• Daniel Kahneman2 Giroux, 2011.
• Elizabeth Feldman Barrett3 3. How Emotions are Made: The secret life of the brain.
• Judea Pearl5,6 Mariner Books;, 2018.
4. Bayesian Networks: With Examples in R. CRC Press,
• Kenneth Wachter7
2015
• Madelyn Glymour6
5. The Book of Why: The new science of cause and effect.
• Marco Scutari, Jean-Baptiste
Basic Books, 2018
Denis4 6. Causal Inference in Statistics: A Primer. Wiley, 2016.
• Nicholas Jewell6,10 7. Demography training, UC Berkeley
• Norman Fenton, Martin Neil8 8. Risk Assessment and Decision Analysis with Bayesian
• Richard Brand10 Networks. CRC Press; 2012.
• Ronald A. Howard, Ali Abbas9 9. Foundations of Decision Analysis. Pearson, 2015.
10. Biostatistics training, UC Berkeley School of Public
• Steve Selvin10
Health
3
1. Background and introduction

4
My career journey

• Masters and doctorate, Epidemiology (MPH, DrPH)


• Resident and Fellow, Internal medicine and infectious diseases
• Medical epidemiologist and deputy health officer
• Principal investigator, CDC training and systems research center
• Instructor, Applied Epidemiology Using R (2004–present)
• Health officer, City and County of San Francisco
• Director, Population Health Division, SFDPH
• Certificate, Stanford University, Strategic decision analysis
• Student, Cal State University Dominguez Hills, online MBA courses
• Current focus: health equity, performance improvement, collective impact,
leadership development, and population health data science

5
My technical journey

• Epidemiology (+ biostatistics, demography, some math modeling)


• Multi-criteria decision methods (analytic hierarchy process, etc.)
• Decision analysis (decision trees)
• Business analytics (linear and integer programming, etc.)
• Bayesian networks (subset of probabilistic graphical models)
• Decision networks (hybrid Bayesian networks)1
• Causal models (causal graph, structural causal model, causal loop diagram)

1
Also called influence diagrams, relevance diagrams, or value maps (business)
6
2. Some concepts and examples

7
Some definitions

Health is a state of complete physical, mental and social well-being and not merely the
absence of disease or infirmity (World Health Organization, 1946).
Public health is what we, as a society, do collectively to assure the conditions in which
people can be healthy (Institute of Medicine, 1988).
Population health2 is a systems3 framework for studying and improving the health of
populations through collective action and learning (Aragon & Garcia, 2017).4

2
Essential population health goals include (1) protecting and promoting health and equity, (2)
transforming people and place, (3) ensuring a healthy planet, and (4) achieving health equity.
3
Complex, adaptive, socioecological systems
4
Download https://escholarship.org/uc/item/825430qn
8
My view of population health: The radical health development model
Model helps us to understand and connect
1. historical, structural, institutional, and
community trauma and toxic stress;
l)
2. inter-generational transmission of
ic a

Historical,
trauma effects (epigenetics and social
NT
ys

structural, community,

(in
ph

and institutional trauma

div
BE
E

risk) to offspring;
ic,

idu
NM
om

Hl,Agroup
a
on
RO
ec

VI , inst
3. life course neurodevelopment
al,

ORituio
, c VI
ur
ult

affecting a child’s brain, learning,


EN

S na
Toxic stress and
ial

complex trauma
c

behavior and health for life; and


(so

l)
inter-generational
Epigenetics and

transmission

4. industry exploitation of our


Resilience?
http://taragonmd.us 2018-03-28
neuro-vulnerabilities to design and
(brain, cognition, emotions, body)
BIOLOGY market products for addiction and
Adapted from UCSF/UC Berkeley health model
overconsumption 9
An overview

What we do naturally, but with many pitfalls5

1. Decision making6
2. Causal inference
3. Probabilistic reasoning

Some tools to support thinking and analysis

1. Decision network (hybrid Bayesian network, influence or relevance diagram)


2. Causal graph (causal Bayesian network, directed acyclic graph)
3. Bayesian network (conditional probabilistic graphical models)
5
See “Understanding intuitive decision-making” in https://escholarship.org/uc/item/825430qn
6
An irrevocable allocation of resources among two or more alternatives
10
Causal thinking with Bayesian networks
Causal (predictive), evidential (diagnostic), and inter-causal (explaining away) reasoning

Bayesian Artificial Intelligence,


Second Edition (2011). By
Kevin B. Korb and Ann E.
Nicholson. CRC Press

11
Causal thinking with Bayesian networks
Causal (predictive), evidential (diagnostic), and inter-causal (explaining away) reasoning

Bayesian Artificial Intelligence,


Second Edition (2011). By
Kevin B. Korb and Ann E.
Nicholson. CRC Press

12
Bayesian networks with ‘bnlearn’: Set up nodes and levels

library(bnlearn)
dag <- empty.graph(nodes = c('P', 'S', 'C', 'X', 'D'))
dag <- set.arc(dag, from = 'P', to = 'C')
dag <- set.arc(dag, from = 'S', to = 'C')
dag <- set.arc(dag, from = 'C', to = 'X')
dag <- set.arc(dag, from = 'C', to = 'D')
P.lv <- c('Low', 'High')
S.lv <- c('Yes', 'No')
C.lv <- c('Yes', 'No')
X.lv <- c('Pos', 'Neg')
D.lv <- c('Yes', 'No')

13
Bayesian networks with ‘bnlearn’: View graph

graphviz.plot(dag)

P S

X D 14
Bayesian networks with ‘bnlearn’: Set up conditional probability tables

P.prob <- array(c(.9, .1), dim=2, dimnames=list(P = P.lv))


S.prob <- array(c(.3, .7), dim=2, dimnames=list(S = S.lv))
C.prob <- array(c(.03,.97,.05,.95,.001,.999,.02,.98), dim=c(2,2,2),
dimnames = list(C=C.lv,P=P.lv, S=S.lv))
X.prob <- array(c(.9,.1,.2,.8), dim=c(2, 2),
dimnames = list(X = X.lv, C = C.lv))
D.prob <- array(c(.65,.35,.3,.7), dim=c(2, 2),
dimnames=list(D = D.lv, C = C.lv))
cpt <- list(P=P.prob, S=S.prob, C=C.prob, X=X.prob, D=D.prob)
bn <- custom.fit(dag, cpt)

15
Bayesian networks with ‘bnlearn’: Display one node

> bn$C
Parameters of node C (multinomial distribution)
Conditional probability table:
, , S = Yes
P
C Low High
Yes 0.030 0.050
No 0.970 0.950
, , S = NO
P
C Low High
Yes 0.001 0.020
No 0.999 0.980
16
Decision (priority-setting) challenges

The roads we take are more important than the goals we announce. Decisions
determine destiny. . . . Frederick Speakman

• Multiple, completing objectives


• Future uncertainty
• Time constraints
• Multiple stakeholders
• Multi-party decisions
• Organizational and environmental complexity
• Analytical complexity
• High stake and high cost implications

17
Structured decision making: node definitions

A great way to structure a decision is with a type of Bayesian network. Bayesian


networks that include decision, uncertainty, calculation, and value nodes are called
relevance diagrams (also called influence diagrams or decision networks).

Decision Uncertainty Calculation Value

Figure 1: Node definitions for relevance diagrams

18
Decision archetype: Bayesian decision network (top) vs. decision tree (bottom)

Test Treat
(evidence) (decision)
Disease Utility
(hypothesis) (value)

U[D+,T+]
pos

Test neg U[D+,T-]


yes
U[D-,T+]
pos
no
Disease Test neg
treat U[D-,T+]
Treat
(decision) no treat yes pos U[D+,T+]
Disease Test
neg
no U[D+,T-]

Test pos U[D-,T+]


neg
19
U[D-,T-]
Example: Univeral challenge: How do we set budget priorities?
Priority-setting and resource allocation

Key approaches

• Program budgeting and marginal analysis (PBMA),


• Multi-criteria decision analysis (MCDA), and
• Accountability for reasonableness (A4R)

Marginal analysis
Involves considering a mix of services and analyzing changes in that mix. The best mix
of servcies is determined by examing the relative costs and benefits of the various
options at the margins

20
Structured decision making: relevance (influence) diagram7

HI criteria j Health impact

Benefit
SA criteria k Strategic alignment
value i

Program OI criteria l Organizational impact


option i

Expenditure m
Budget
Financial impact value i
Revenue n

Figure 2: Multi-criteria (or multi-objective) decision analysis (MCDA)


21
Fundamental objectives and criteria with weights

Fundamental objectives Criteria (to be weighted)

Health Impact Effectiveness


(50) Equity
Primary prevention (social determinants)
Innovation
Alignment with community/client preferencesxs
Strategic Alignment Alignment with PHD stated goals
(30) Alignment with external entities (political feasibility)
Ability to meet accreditation (includes QI practice)
Impact on legal or regulatory mandate
Collaboration/ partnership
Alignment with PHD goals
Organizational Impact Workforce issues: morale, opportunity for development
(20) Operational efficiency
Sustainability
Financial Impact Associated Revenue
Downstream impact on service utilization

22
Example: Causal graph for a public health hazard

Hazard Exposure Consequences

23
Causal graph for a public health hazard with interventions

Hazard Risk Consequence


mitigation mitigation management
(primary (secondary (tertiary
prevention) prevention') prevention)

Hazard Exposure Consequences

24
Draft decision networks for bioterrorism attack (anthrax)

Release Exposure Death


Disease
Medical Value
Care
Specificity Test 1 Closure Costs

Sensitivity PEP

Sensitivity PEP

Test 1 Test 2
Specificity
Closure

Costs
Release Exposure Medical
Disease
Care Value
Death

25
Example: Trauma/toxic stress affects neurodevelopment & executive function
Left figure: System 1 and System 2; Right figure: Neuroplasticity and executive function

Executive function skill proficiency

26
The Hacking of the American Mind*
The science behind the corporate takeover of our bodies and brains (Lustig, 2017)

Industry goals: increase normalization, demand, and consumption

Social Marketing Executive Function Price Supply Location

Social Norms Perceptions Resources Availability

Pleasure* Demand Consumption (age of initiation,


(dopamine) potency, quantify, and frequency)

neg

Happiness* Problem Use Toxicity Intoxication


(serotonin) (dependence, abuse) (end-organ) (risky behavior)
27
Example: Complaints of syringes and feces rise dramatically in SF
SF Gate, Wednesday, November 2, 2016

28
Strengthening syringe access to protect the public’s health8

PWID migration Primary drug Drug treatment


into and out of SF. prevention on demand

DPH Syringe Access Syringes from


and Disposal Program other sources No. of PWID

Safe injection
Syringe disposal services DPW site 311 complaints No. of syringes
and recovery clean-ups and response distributed
(proposed)

DPH & DPW No. of improperly BASELINE PREVALENCE Other PWID risk component: Syringe sharing
Worker safety discarded syringes (Initial probability of Transmission probability contact rate
bloodborne infections) from syringe sharing

SCM Public Other SCM risk components:


(1) PWID infection
needlestick needlestick 1. Transmission probability risk from sharing
contact rate contact rate 2. Post-exposure prophylaxis

UPDATED PREVALENCE Other HCW risk components:


(4) Public infection Probability of having 1. Needlestick contact rate
risk from needlestick bloodborne infections 2. Transmission probability
3. Post-exposure prophylaxis

(3) SCM infection (2) HCW infection risk


risk from needlestick from needlestick

8 29
Garcia BA, Aragon TJ, et al. (2018). Available from https://escholarship.org/uc/item/2r7298jr.
The Book of Why: The new science of cause and effect (Pearl, 2018)

30
Causal graph: Archetype 1 confounding — Backdoor criterion

31
Causal graph: Archetype 2 confounding w/ mediation — Frontdoor criterion

32
Causal graph: Archetype 3 confounding — Instrumental variable

33
Exercise preparation: based on “How Emotions are Made”
The secret life of the brain (Barrett, 2018)

1. Classical view of emotions


2. Brain’s mission (20/80 rule)
3. Prediction (concepts; System 1)
4. Sensory input (incl. interoception)
5. Prediction error (± learning; System 2)
6. Experience/meaning (categorization)
7. Affect and affective realism
8. Social reality and evolution of cultures

34
35
Affective circumplex: valence (pleasant, unpleasant) vs. arousal (low, mid, high)

Unpleasant High arousal Pleasant


High arousal High arousal
(upset, distressed) (elated, thrilled)

Mid arousal Mid arousal


(miserable, displeased) (gratified, pleased)

Low arousal Low arousal


(lethargic, depressed) (serene, calm)

Low arousal
36
Figure 3
37
38
39
40
Mental phenomena represented as combinations of prediction and sensory input

100%
Sensory input
Prediction
0%
Daydreaming Learning Autism
Memory Meditative states
Delusions Experiential blindness
Hallucinations LSD trip
Placebo effect
Imagination
Optical illusions
Predictions that match sensory input

Figure 4
41
Exercise: Generate alternative causal graph

Here is simple causal graph with cause, mediation, and effect.

Sensory Affect Emotion


input (feeling) (meaning)

Consider a direct physical threat from an angry person (sensory input), leading to
unpleasant arousal (affect), leading to fear and anger (emotion). Use concepts from
previous slides and your knowledge and experience . . .

1. to construct more detailed causal graph, and then


2. to add an intervention node.

How can you transform your causal graph into a Bayesian decision network?
42
3. Ideas for the future

43
Population health data science climbing Pearl’s “Ladder of Causation”

Become the best at getting better!


1. Decision making
2. Causal inference
3. Probabilistic reasoning
Teach epidemiology by building upon
population health decision making
1. Decision networka
2. Causal graphb
3. Bayesian network
a
hybrid Bayesian network, influence diagram
b
causal Bayesian network, directed acyclic graph

44
Population health data science landscape (source: http://www.bayesia.com/)
The "Causal Revolution" will fill out Quadrant 3.

45
Questions

Our comforting conviction that the world


makes sense rests on a secure foundation:
our almost unlimited ability to ignore our
ignorance.
. . . Daniel Kahneman9
9
Thinking, Fast and Slow. New York: Farrar, Straus and Giroux, 2011.
46

Vous aimerez peut-être aussi