Applying AI in preventative health interventions: Algorithms, deployments and fairness
MILIND TAMBE
Director Center for Research on Computation and Society
Harvard University
&
Director “AI for Social Good”
Google Research India
Date: 12/14/2019 2
AI and Multiagent Systems Research for Social Impact
Public Safety
and Security
ConservationPublic Health
Viewing Social Problems as Multiagent Systems
Key research challenge across problem areas:
Optimize Our Limited Intervention Resources
when
Interacting with Other Agents
Date: 12/14/2019 3
Multiagent Systems Reasoning:
Social Networks, Game Theory
Reasoning with Social Networks
▪ Social networks to enhance intervention, e.g., HIV information
▪ Real-world pilot tests: Homeless youth shelters in Los Angeles
Public HealthOptimizing Limited Intervention (Health Worker) Resources
Date: 12/14/2019 4
▪ Security games and adversary (poacher) behavior prediction
▪ Real-world: National parks in Uganda, Malaysia…
Green Security Games
Conservation/Wildlife ProtectionOptimizing Limited Intervention (Ranger) Resources
Date: 12/14/2019 5
▪ Game Theory for security resource optimization
▪ Real-world: US Coast Guard, US Federal Air Marshals Service…
Public Safety and SecurityOptimizing Limited Intervention (Security) Resources
Stackelberg Security Games
Date: 12/14/2019 6
Common ThemesInterdisciplinary Partnerships, Multiagent Systems, Data-to-deployment pipeline
Date: 12/14/2019 7
Field tests&
deployment
Prescriptivealgorithm
Multiagent Reasoning; Intervention
Immersion
Data Collection
Date: 12/14/2019 8
Predictivemodel
Learning/Expert input
Common ThemesInterdisciplinary Partnerships, Multiagent Systems, Data-to-deployment pipeline
Fairness Challenges in AI for Public Health
Field tests
Prescribe
Select key influencers
Data
PartialSocial
network
Date: 12/14/2019 9
Predict
Link Prediction
Fairness challenges in intervention; not just prediction
Fairness Challenges in AI for Public Health
Field tests
Prescribe
Select key influencers
Data
PartialSocial
network
Date: 12/14/2019 10
Predict
Link Prediction
Fairness challenges in intervention; not just prediction
Fairness Challenges in AI for Public Health
Field tests
Prescribe
Select key influencers
Data
PartialSocial
network
Date: 12/14/2019 11
Predict
Link Prediction
Fairness challenges in intervention; not just prediction
Address fairness: Need collaboration of AI & public health stakeholders
Fairness & AI for Public Health
HIV prevention
Date: 12/14/2019 12
TB Prevention
Suicide Prevention
AI Interventions in Public health
Fairness challenges in intervention
PhD student photos in top right corner
Public HealthOptimizing Limited Intervention (Social Worker) Resources
13Date: 12/14/2019
Preventing HIV in homeless youth: Rates of HIV 10 times housed population
➢ Shelters: Limited number of peer leaders to spread HIV information in social networks
➢ “Real” social networks gathered from observations in the field; not facebook etc
Influence Maximization Background
▪ Given:
▪ Social network Graph G
▪ Choose K “peer leader” nodes
▪ Objective:
▪ Maximize expected number of influenced nodes
▪ Assumption: Independent cascade model of information spread
14Date: 12/14/2019
Independent Cascade Model and Real-world Physical Social Networks
15Date: 12/14/2019
A BP(A,B)=0.4
C D
0 1μ = 0.5
C D
0 1μ ∈ [0.3, 0.7]
Robust, Dynamic Influence Maximization
▪ Worst case parameters: a zero-sum game against nature
▪ Payoffs: (performance of algorithm)/OPT
Nature
Chooses parameters
μ,σvs
Algorithm
Chooses policy, i.e.,
Chooses Peer leaders
16Date: 12/14/2019
Wilder
Params #1 Params #2
Policy #1 0.8, -0.8 0.3, -0.3
Policy #2 0.7, -0.7 0.5, -0.5
HEALER Algorithm [2017]Robust, Dynamic Influence Maximization
▪ Equilibrium strategy despite exponential strategy spaces: Double oracle
Influencer’s oracle
Nature’s oracle
Params #1 Params #2 Params #3
Policy #1 0.8, -0.8 0.3, -0.3 0.4, -0.4
Policy #2 0.7, -0.7 0.5, -0.5 0.6, -0.6
Policy #3 0.6, -0.6 0.4, -0.4 0.7, -0.7
Influ
en
ce
r
Nature
17Date: 12/14/2019
Params #1 Params #2 Params #3
Policy #1 0.8, -0.8 0.3, -0.3 0.4, -0.4
Policy #2 0.7, -0.7 0.5, -0.5 0.6, -0.6
Policy #3 0.6, -0.6 0.4, -0.4 0.7, -0.7
Wilder
Theorem: Converge with approximation guarantees
\ Params #1 Params #2
Policy #1 0.8, -0.8 0.3, -0.3
Policy #2 0.7, -0.7 0.5, -0.5
Policy #3 0.6, -0.6 0.4, -0.4
Challenge: Multi-step Policy
K = 4
1st time step
18Date: 12/14/2019
Params #1 Params #2 Params #3
Policy #1 0.8, -0.8 0.3, -0.3 0.4, -0.4
Policy #2 0.7, -0.7 0.5, -0.5 0.6, -0.6
Policy #3 0.6, -0.6 0.4, -0.4 0.7, -0.7
K = 4
2nd time step
WilderYadav
12/14/2019 19
HEALER: POMDP Model for Multi-Step PolicyRobust, Dynamic Influence Maximization
Action
Choose nodes
Observation: Update
propagation probability
POMDP
Policy
HIDDEN STATE
Yadav
Params #1 Params #2 Params #3
Policy #1 0.8, -0.8 0.3, -0.3 0.4, -0.4
Policy #2 0.7, -0.7 0.5, -0.5 0.6, -0.6
Policy #3 0.6, -0.6 0.4, -0.4 0.7, -0.7
POMDP
partitions
Pilot Tests with HEALER
with 170 Homeless Youth [2017]
Recruited youths:
HEALER HEALER++ DEGREE CENTRALITY
62 56 55
20Date: 12/14/2019
12 peer leaders
Yadav Wilder
Results: Pilot Studies [2017]
0
20
40
60
80
100
HEALER HEALER++ Degree
Percent of non-Peer Leaders
Informed Not Informed
0
20
40
60
80
100
HEALER HEALER++ Degree
Informed Non-Peer Leaders Who Started Testing for HIV
Testing Non-Testing
21Date: 12/14/2019
Yadav Wilder
More details: Journal of Society of Social Work & Research (Nov 2018)
Data to Deployment Pipeline:Network Sampling to avoid Data Collection Bottleneck
22Date: 12/14/2019
Data collection costly Sample 18%
Sampling from largest
communities
Wilder
New experiment With 60 homeless youth
12 peer leaders
Results: Pilot Studies
with New Sampling Algorithm [2018]
0
20
40
60
80
100
SAMPLE HEALER Degree
Percent of non-Peer Leaders
Informed Not Informed
0
20
40
60
80
100
SAMPLE HEALER Degree
Informed Non-Peer Leaders Who Started Testing for HIV
Testing Non-Testing
23Date: 12/14/2019
Wilder
AI Assistant: HEALER
24Date: 12/14/2019
Continuing Research on HIV prevention [2019]
▪ Completing 900 youth study at three homeless shelters
12/14/2019 25
Public Health: Optimizing Limited Social Worker ResourcesPreventing Tuberculosis in India [2019]
26Date: 12/14/2019
Tuberculosis (TB): ~500,000 deaths/year, ~3M infected in India
➢ Patient in low resource communities: Non-adherence to TB Treatment
➢ Digital adherence tracking: Patients call phone #s on pill packs; many countries
➢ Predict adherence risk from phone call patterns? Intervene before patients miss dose
TB Treatment Adherence but Limited Resources:
Intervening Selectively before patients miss doses
Date: 12/14/2019 27
Data Collect
Phone logs
Predicthigh risk patients
RF or LSTM
Prescription
ConstraintTop K
Field
Killian
➢ 15K patients, 1.5M calls
Increasing TB Treatment Adherence:
Intervening before patients miss doses [2019]
Date: 12/14/2019 28
Killian
Data from
State ofMaharashtra
India
107
120
144
97
0
20
40
60
80
100
120
140
160
180
True Positives False Positives
Nu
mb
er o
f Pa
tien
ts
Best Model vs. Baseline: Prediction High Risk Patients
Baseline Best Model
+35% -19%
Improving TB interventions
Decision-Focused Methods (simultation study)
Date: 12/14/2019 29
Wilder
0.4
0.42
0.44
0.46
0.48
0.5
Interventions: Decision-Focused Better
Stage by Stage Decision Focused
Decision focused learning improves TB interventions
Integrating with Everwell’s Platform
Date: 12/14/2019 30
Killian
This work has a lot of potential to save lives.
Bill ThiesCo-founder, Everwell Health Solutions
.
Suicide Prevention in Marginalized Populations: Choose Gatekeepers in social networks
11/11/201931
Rahmattalabi
Suicide rate is significantly high among youth experiencing homelessness (~7%-26%)
More than 50% have either had thoughts of suicide or have made an attempt
Suicide Prevention in Marginalized Populations: Choose Gatekeepers in social networks
11/11/201932
▪ Worst case availability: constant-sum game against nature
Rahmattalabi
Nature
Chooses some
gatekeepers to not
participate
vs
Algorithm
Chooses K gatekeepers
Train: Available: Covered:
Fairness & AI for Public Health
HIV prevention
Date: 12/14/2019 33
TB Prevention
Suicide Prevention
AI and Public health
Fairness challenges in intervention
Suicide Prevention in Marginalized Populations: Choose Gatekeepers in social networks (NeurIPS 2019)
11/11/201934
Rahmattalabi
Is there any Inequality in coverage of marginalized populations?
Suicide Prevention in Marginalized Populations: Choose Gatekeepers in social networks (NeurIPS 2019)
11/11/201935
Rahmattalabi
▪ Inequality in coverage of marginalized population
Maximin Fairness:Towards addressing Fairness Challenges
11/11/2019 36
Rahmattalabi▪ Maximin fairness constraints: Every group at least W coverage
HIV Prevention in Marginalized Populations: Choose Peer Leaders (IJCAI 2019)
11/11/201937
• Intervention: Influence maximization
• Are there any fairness violations?
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
NativeAmerican
Asian Black NativeHawaiian
White Latino
Fairness violation
Greedy
HIV Prevention in Marginalized Populations: Choose Peer Leaders (IJCAI 2019)
11/11/201938
• Are there any fairness violations?
Budget: k = 5
Optimal value with 𝑘 = 3: 4.25
Constraint:f 𝑆 ≥ 4.25
𝑝 = 0.25
Optimal value with 𝑘 = 2: 2.5
Constraint:f 𝑆 ≥ 2.5
k = 3 k = 2
9
15⋅ 5
6
15⋅ 5
Multiobjective Influence MaximizationTowards Addressing Fairness Challenges
11/11/201939
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
NativeAmerican
Asian Black NativeHawaiian
White Latino
Fairness violation
Greedy FairInfluence
• Four networks from HIV prevention interventions:~70 nodes each
• Compare standard greedy algorithm to both fairness concepts
• Is more attention or less attention fair?
• What are fairness concerns in south Asian context?
• Is interpretability a requirement for fairness?
Date: 12/14/2019 40
Fairness in TB Treatment
AI & Multiagent Systems for Social Impact
Fairness and AI for Public HealthChallenges in Intervention
Date: 12/14/2019 41
Fairness challenges in intervention; important to step beyond prediction
No consensus on exact fairness approach; rising concerns AI & fairness
Public health officials/stakeholders involved from start in fair solutions
Collaboration between AI systems & Stakeholders to address fairness
Thank you!
42Date: 12/14/2019