Optimizing Disease Outbreak Detection Methods Using Reinforcement Learning Masoumeh Izadi Clinical & Health Informatics Research Group Faculty of Medicine,

Slides:



Advertisements
Similar presentations
Dialogue Policy Optimisation
Advertisements

Markov Decision Process
1 University of Southern California Keep the Adversary Guessing: Agent Security by Policy Randomization Praveen Paruchuri University of Southern California.
SA-1 Probabilistic Robotics Planning and Control: Partially Observable Markov Decision Processes.
CSE-573 Artificial Intelligence Partially-Observable MDPS (POMDPs)
2005 Syndromic Surveillance1 Estimating the Expected Warning Time of Outbreak- Detection Algorithms Yanna Shen, Weng-Keen Wong, Gregory F. Cooper RODS.
 2005 Carnegie Mellon University A Bayesian Scan Statistic for Spatial Cluster Detection Daniel B. Neill 1 Andrew W. Moore 1 Gregory F. Cooper 2 1 Carnegie.
Early Statistical Detection of Bio-Terrorism Attacks by Tracking OTC Medication Sales Galit Shmueli Dept. of Statistics and CALD Carnegie Mellon University.
Overview of Uses for Public Health Surveillance Daniel M. Sosin, M.D., M.P.H. Division of Public Health Surveillance and Informatics Epidemiology Program.
Bayesian Biosurveillance Gregory F. Cooper Center for Biomedical Informatics University of Pittsburgh The research described in this.
Empirical/Asymptotic P-values for Monte Carlo-Based Hypothesis Testing: an Application to Cluster Detection Using the Scan Statistic Allyson Abrams, Martin.
Planning under Uncertainty
Decision Theoretic Analysis of Improving Epidemic Detection Izadi, M. Buckeridge, D. AMIA 2007,Symposium Proceedings 2007.
 2004 University of Pittsburgh Bayesian Biosurveillance Using Multiple Data Streams Weng-Keen Wong, Greg Cooper, Denver Dash *, John Levander, John Dowling,
What’s Strange About Recent Events (WSARE) v3.0: Adjusting for a Changing Baseline Weng-Keen Wong (Carnegie Mellon University) Andrew Moore (Carnegie Mellon.
KI Kunstmatige Intelligentie / RuG Markov Decision Processes AIMA, Chapter 17.
Planning in MDPs S&B: Sec 3.6; Ch. 4. Administrivia Reminder: Final project proposal due this Friday If you haven’t talked to me yet, you still have the.
Results 2 (cont’d) c) Long term observational data on the duration of effective response Observational data on n=50 has EVSI = £867 d) Collect data on.
Bayesian Biosurveillance Using Causal Networks Greg Cooper RODS Laboratory and the Laboratory for Causal Modeling and Discovery Center for Biomedical Informatics.
1 Hybrid Agent-Based Modeling: Architectures,Analyses and Applications (Stage One) Li, Hailin.
Conclusions On our large scale anthrax attack simulations, being able to infer the work zip appears to improve detection time over just using the home.
Markov Decision Processes
Planning to learn. Progress report Last time: Transition functions & stochastic outcomes Markov chains MDPs defined Today: Exercise completed Value functions.
Lecture 10 Comparison and Evaluation of Alternative System Designs.
Population-Wide Anomaly Detection Weng-Keen Wong 1, Gregory Cooper 2, Denver Dash 3, John Levander 2, John Dowling 2, Bill Hogan 2, Michael Wagner 2 1.
More RL. MDPs defined A Markov decision process (MDP), M, is a model of a stochastic, dynamic, controllable, rewarding process given by: M = 〈 S, A,T,R.
Bayesian Network Anomaly Pattern Detection for Disease Outbreaks Weng-Keen Wong (Carnegie Mellon University) Andrew Moore (Carnegie Mellon University)
1 Bayesian Network Anomaly Pattern Detection for Disease Outbreaks Weng-Keen Wong (Carnegie Mellon University) Andrew Moore (Carnegie Mellon University)
Sample Size Determination
CS Reinforcement Learning1 Reinforcement Learning Variation on Supervised Learning Exact target outputs are not given Some variation of reward is.
MANSOUR SHARAHILI 1 Modelling and optimising the sport and exercise training process With thanks to professor Philip Scarf.
Predictive State Representation Masoumeh Izadi School of Computer Science McGill University UdeM-McGill Machine Learning Seminar.
The Snow Project Methods April 2008 Johan G. Bellika ab A Department of Computer Science, University of Tromsø B Norwegian Centre for Telemedicine, University.
Using Disease Surveillance and Response to Facilitate Adaptation to Climate- Related Health Risks Kristie L. Ebi, Ph.D., MPH Development Day at COP-11.
Extracting Places and Activities from GPS Traces Using Hierarchical Conditional Random Fields Yong-Joong Kim Dept. of Computer Science Yonsei.
SPONSOR JAMES C. BENNEYAN DEVELOPMENT OF A PRESCRIPTION DRUG SURVEILLANCE SYSTEM TEAM MEMBERS Jeffrey Mason Dan Mitus Jenna Eickhoff Benjamin Harris.
MAKING COMPLEX DEClSlONS
Epidemiology The Basics Only… Adapted with permission from a class presentation developed by Dr. Charles Lynch – University of Iowa, Iowa City.
Reinforcement Learning
Hypothesis Testing: One Sample Cases. Outline: – The logic of hypothesis testing – The Five-Step Model – Hypothesis testing for single sample means (z.
Ohio Digital Government Summit Disease Surveillance (Homeland Security session) October 5, 2004 Rana Sen Deloitte Consulting LLP.
What’s Strange About Recent Events (WSARE) Weng-Keen Wong (University of Pittsburgh) Andrew Moore (Carnegie Mellon University) Gregory Cooper (University.
Cluster Detection Comparison in Syndromic Surveillance MGIS Capstone Project Proposal Tuesday, July 8 th, 2008.
Syndromic Surveillance in Montreal: An Overview of Practice and Research David Buckeridge, MD PhD Epidemiology and Biostatistics, McGill University Surveillance.
Analyzing over-the-counter medication purchases for early detection of epidemics and bio-terrorism by Anna Goldenberg Advisor: Rich Caruana Note: Sponsored.
Using the Repeated Two-Sample Rank Procedure for Detecting Anomalies in Space and Time Ronald D. Fricker, Jr. Interfaces Conference May 31, 2008.
Utilities and MDP: A Lesson in Multiagent System Based on Jose Vidal’s book Fundamentals of Multiagent Systems Henry Hexmoor SIUC.
TM Emerging Health Threats and Health Information Systems: Getting Public Health and Clinical Medicine to Real Time Response John W. Loonsk, M.D. Associate.
A Passive Approach to Sensor Network Localization Rahul Biswas and Sebastian Thrun International Conference on Intelligent Robots and Systems 2004 Presented.
Project Sentinel Collaboratory Georgetown University Medical Center Washington Hospital Center Seong K. Mun, PhD Funded By National Library of Medicine.
1 Random Disambiguation Paths Al Aksakalli In Collaboration with Carey Priebe & Donniell Fishkind Department of Applied Mathematics and Statistics Johns.
BME 353 – BIOMEDICAL MEASUREMENTS AND INSTRUMENTATION MEASUREMENT PRINCIPLES.
Reinforcement Learning AI – Week 22 Sub-symbolic AI Two: An Introduction to Reinforcement Learning Lee McCluskey, room 3/10
U of Minnesota DIWANS'061 Energy-Aware Scheduling with Quality of Surveillance Guarantee in Wireless Sensor Networks Jaehoon Jeong, Sarah Sharafkandi and.
Reinforcement Learning Guest Lecturer: Chengxiang Zhai Machine Learning December 6, 2001.
Bayesian Biosurveillance of Disease Outbreaks RODS Laboratory Center for Biomedical Informatics University of Pittsburgh Gregory F. Cooper, Denver H.
Partially Observable Markov Decision Process and RL
Reinforcement Learning (1)
Intelligent Systems (AI-2) Computer Science cpsc422, Lecture 7
Bayesian Biosurveillance of Disease Outbreaks
Biomedical Data & Markov Decision Process
One Health Early Warning Alert
Estimating the Expected Warning Time of Outbreak-Detection Algorithms
Gerald Dyer, Jr., MPH October 20, 2016
Reinforcement Learning Dealing with Partial Observability
Reinforcement Nisheeth 18th January 2019.
Reinforcement Learning (2)
Intelligent Systems (AI-2) Computer Science cpsc422, Lecture 7
Reinforcement Learning (2)
Intelligent Systems (AI-2) Computer Science cpsc422, Lecture 7
Presentation transcript:

Optimizing Disease Outbreak Detection Methods Using Reinforcement Learning Masoumeh Izadi Clinical & Health Informatics Research Group Faculty of Medicine, McGill

Overview Motivation Problem formulation Basic definitions The suggested method Experimental results Concluding remarks

The Surveillance Cycle Event Reports Individual Event Definitions Population Pattern Definitions Event Detection Algorithm Pattern Report Population Under Surveillance Intervention Decision Intervention Guidelines Public Health Action Data Describing Population Pattern Detection Algorithm 1. Identifying individual cases 2. Detecting population patterns 3. Conveying information for action (Buckeridge DL & Cadieux G, 2007)

Surveillance Research Achieving the National Electronic Disease Surveillance System (NEDSS) architecture Data fusion (linkage) New data sources Case definitions (automation/validation) Geographic Information System (GIS) indices Forecasting Evaluation and quality control

The Surveillance Cycle Event Reports Individual Event Definitions Population Pattern Definitions Event Detection Algorithm Pattern Report Population Under Surveillance Intervention Decision Intervention Guidelines Public Health Action Data Describing Population Pattern Detection Algorithm 1. Identifying individual cases 2. Detecting population patterns 3. Conveying information for action Decision Algorithm Knowledge 2. Using RL to identify optimal policies for responding to statistical alarms. 1. Accounting for population mobility in detecting spatial disease clusters. 3. Simulation modeling to evaluate outbreak detection. (Buckeridge DL & Cadieux G, 2007)

Outbreak Detection Knowledge Data Detection Method Environment Warning???

Data Sources

Outbreak Problems Large scale bioaerosol (e.g., Anthrax) Communicable (e.g., SARS) Waterborne Building contamination Foodborne Continuous release Sexual/blood borne

Detection Methods Define a threshold. Signal an alarm when the # of ED visits per day exceeds the threshold.

Anthrax Cases in DC Flu Flu Flu Anthrax Attacks Data courtesy of Medstar & Georgetown University

Existing Detection Methods  Temporal methods e.g. Moving average  Spatio-temporal methods e.g. Space-time scan

Features Shared by Most Detection Methods Design a baseline. Define an important event when the p-value of a statistic is less than an expected value by the baseline.

Obtaining Baseline Data Baseline All Historical Data Today’s Environment 1.Learn Bayesian Network using Optimal Reinsertion [Moore and Wong 2003] 2. Generate baseline given today’s environment Bayesian Biosurveillance of Disease Outbreaks [UAI04 Cooper et al]

Important Events determine which of these p-values are significant for a specific problem. Idea: use association rules to define cases

Key Observations  There is a great amount of uncertainty about suspicious events.  An action has to be taken in response to any suspicious change in the environmental patterns.  Surveillance systems faced by high-risk decision problems under uncertainty.

Surveillance algorithms are inaccurate in practice How precisely can we detect if an outbreak is happening? (sensitivity) How early can we detect it? (timeliness) Research to address this problem –Novel or ‘improved’ data streams –Better forecasts or detection methods –Improve decision making after alarms

Our Approach Instead of trying to improve the detection method, we ‘post-process’ the signals:  Use a standard surveillance method to provide alarm signals  Feed this signal to the model of outbreak detection as a partially observed Markov decision process (POMDP)

Partially Observable MDP POMDPs are characterized by: –States:s  S –Actions: a  A –Observations: o  O –Transition probabilities: T(s,a,s’)=Pr(s’|s,a) –Observation probabilities: T(o,a,s’)=Pr(o|s,a) –Rewards: R(s,a)

Solving POMDPs To solve a POMDP is to find, for any action/observation history, the action that maximizes the expected discounted reward. V(b)= max a [Σ s R(s,a)b(s)+ Σ s’ [T(s,a,s’)O(s’,a,z)α(s’)]] OUTCOME: an optimal policy over belief space

Suitability  The ‘true’ state of the outbreak cannot be observed  Statistical algorithms provide imperfect measurements of the true state  That the probability of success of (i.e., effectiveness) of actions can be determined  The that costs of actions and of outcomes can be determined

Limitations for inhalational anthrax Limited data from actual anthrax attacks available: –Postal attacks 2001 (Only 11 people affected, not representative of a large scale attack) –Sverdlovsk 1979 But literature contains studies on the characteristics of inhalational anthrax

Background knowledge for inhalational anthrax Can coherently incorporate different types of simulation data : Progression of symptoms Incubation period Spatial dispersion pattern

The POMDP Model S - True epidemic state {No Outbreak, D1, ….} O - Output from detection algorithm {0,1} A - Possible public health actions T(s,a,s’) - Impact of actions given the state R(s,a) - Costs of actions and of epidemic states Do nothing Review records Investigate cases Declare outbreak ActionTransition (Izadi M & Buckeridge DL, 2007)

 The transition functions reflect the probability of moving to another state if an action is performed in each state of the model. Clear Day 1 Day 2 Day 3 Day 4 Detected ClearD1D2D3D4Det s s’ T: Review records T: Investigate Transition Functions

Observation Functions Observations are noisy output of the detection algorithm Alarm - sensitivity at outbreak states and 1 - specificity in the no outbreak state. No Alarm - specificity at normal states and 1 - sensitivity in each outbreak state.

Sensitivity versus Specificity

Sensitivity in Days of Outbreak Reis et al. (2003) Proc. Natl. Acad. Sci. USA 100,

Costs and Reward Costs  Investigation (false and true positive)  Intervention (false and true positive)  Outbreak by day (false negative) calculated as (# deaths* future earnings) + (# hospitalized * cost of hospitalization) + (# outpatient visits * cost of visit) Rewards  Preventable costs each day - investigation / intervention costs Sources  Investigation costs are estimated from wages  Intervention and outbreak costs from (Kaufman, 1997)

Experimental Setup  There is a constant probability of an outbreak.  Epidemic curve taken from historical outbreak.  After 4 days, the outbreak is detected clinically.  Population size is 50,000 exposed and the outbreak results in a mean increase in surveillance data of 8% or 15% POMDP solution –Point-based approximation –Ran simulation for ten years.

Things to Notice Any alerts before actual anthrax release are considered a false positive Detection time calculated as first selection of C/P action after anthrax release. Maximum detection time is 4 days.

Preliminary Results MethodPerformance Sensitivity Specificity POMDP 100- Moving Average Linear Exponential610.97

Initial Evaluation Results 8% Increase in ED visits15% Increase in ED visits Day of Outbreak  Compared POMDP operating on detection method, to detection method alone  Method was SARIMA + MA on residuals  Specificity of 0.97 for the detection method used Sensitivity

Final Words  Conclusion: POMDP improves the timeliness and the sensitivity of detection processes  Future work:  Sensitivity analysis over parameter values.  Apply to other diseases and in other settings!

Thank You