Transcript Surveillance System Component Tasks

Statistical Methods for Alerting
Algorithms in Biosurveillance
Howard S. Burkom
The Johns Hopkins University Applied Physics Laboratory
National Security Technology Department
Washington Statistical Society Seminar
February 3, 2006
National Center for Health Statistics
Hyattsville, MD
ESSENCE Biosurveillance Systems
• ESSENCE: An Electronic
Surveillance System for
the Early Notification of
Community-based Epidemics
• Monitoring health care data
from ~800 military treatment
facilities since Sept. 2001
• Evaluating data sources
– Civilian physician visits
– OTC pharmacy sales
– Prescription sales
– Nurse hotline/EMS data
– Absentee rate data
• Developing & implementing
alerting algorithms
Outline of Talk
• Prospective Syndromic Surveillance:
introduction, challenges
• Algorithm Evaluation Approaches
• Statistical Quality Control in Health
Surveillance
• Data Modeling and Process Control
• Regression Modeling Approach
• Generalized Exponential Smoothing
• Comparison Study
• Summary & Research Directions
Required Disciplines: Medical/Epi
Medical/Epidemiological
• filtering/classifying clinical
records => syndromes
• interpretation/response to
system output
• coding/chief complaint
interpretation
Required Disciplines: Informatics
Information Technology
• surveillance system
architecture
• data ingestion/cleaning
• interface between health
monitors and system
Required Disciplines: Analytics
Analytical
• Statistical hypothesis tests
• Data mining/automated
learning
• Adaptation of methodology to
background data behavior
Essential Task Interaction
in Volatile Data Background
Medical/Epidemiological
• filtering/classifying clinical
records => syndromes
• interpretation/response to
system output
• coding/chief complaint
interpretation
Information Technology
• surveillance system
architecture
• data ingestion/cleaning
• interface between health
monitors and system
Analytical
• Statistical hypothesis tests
• Data mining/automated learning
• Adaptation of methodology to
background data behavior
The Multivariate Temporal
Surveillance Problem
Varying Nature of the Data:
Multivariate Nature of Problem:
•
•
•
•
Scale, trend, day-of-week, seasonal
behavior depending on grouping:
Many locations
Multiple syndromes
Stratification by age, gender, other
covariates
Surveillance Challenges:
• Defining anomalous behavior(s)
– Hypothesis tests--both appropriate
and timely
• Avoiding excessive alerting due to
multiple testing
– Correlation among data streams
– Varying noise backgrounds
• Communication with/among users at
different levels
• Data reduction and visualization
Data issues affecting monitoring
Most suitable for
– Statistical properties
modeling without
• Scale and random dispersion
data-specific
– Periodic effects
information
• Day-of-week effects, seasonality
– Delayed (often variably) availability in monitoring system
– Trends: long/short term: many causes, incl. changes in:
• Population distribution or demographic composition
• Data provider participation
• Consumer health care behavior
• Coding or billing practices
– Prolonged data drop-outs, sometimes with catch-ups
– Outliers unrelated to infectious disease levels
• Often due to problems in data chain
• Inclement weather
• Media reports (example: the “Clinton effect”)
Forming the Outcome Variable:
Binning by Diagnosis Code
Rash Syndrome Grouping
of Diagnosis Codes
www.bt.cdc.gov/surveillance/syndromedef/word/syndromedefinitions.doc
Rash ICD-9-CM Code List
ICD9CM
050.0
050.1
050.2
050.9
051.0
051.1
052.7
052.8
052.9
057.8
057.9
695.0
695.1
695.2
695.89
695.9
ICD9DESCR
SMALL POX, VARIOLA MAJOR
SMALL POX, ALASTRIM
SMALL POX, MODIFIED
SMALLPOX NOS
COWPOX
PSEUDOCOWPOX
VARICELLA COMPLICAT NEC
VARICELLA W/UNSPECIFIED C
VARICELLA NOS
EXANTHEMATA VIRAL OTHER S
EXANTHEM VIRAL, UNSPECIFI
ERYTHEMA TOXIC
ERYTHEMA MULTIFORME
ERYTHEMA NODOSUM
ERYTHEMATOUS CONDITIONS O
ERYTHEMATOUS CONDITION N
692.9
782.1
DERMATITIS UNSPECIFIED CA
RASH/OTHER NONSPEC SKIN E
2
2
026.0
026.1
026.9
051.2
051.9
053.20
SPIRILLARY FEVER
STREPTOBACILLARY FEVER
RAT-BITE FEVER UNSPECIFIED
DERMATITIS PUSTULAR, CONT
PARAVACCINIA NOS
HERPES ZOSTER DERMATITIS E
HERPES ZOSTER WITH OTHER SPECIF
COMPLIC
H.Z. W/ UNSPEC. COMPLICATION
HERPES ZOSTER NOS W/O COM
ECZEMA HERPETICUM
HERPES SIMPLEX W/OTH.SPEC
HERPES SIMPLEX, W/UNS.COM
3
3
3
3
3
3
053.79
053.8
053.9
054.0
054.79
054.8
Consensus
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
3
3
3
3
3
3
Chief Complaint Query
Simulated Data
Dynamic Detection
Dynamic Detection
Simulated Data
Example with Detection Statistic Plot
Injected Cases
Presumed
Attributable to
Outbreak Event
Threshold
Comparing Alerting Algorithms
Criteria:
• Sensitivity
– Probability of detecting an outbreak signal
– Depends on effect of outbreak in data
• Specificity ( 1 – false alert rate )
– Probability(no alert | no outbreak )
– May be difficult to prove no outbreak exists
• Timeliness
– Once the effects of an outbreak appear in
the data, how soon is an alert expected?
Modeling the Signal
as Epicurve of Primary Cases
Observed vs Modeled Incubation Period
Distribution: Sverdlovsk 1979 Outbreak
12
Number of Cases
• Need “data epicurve”:
time series of
attributable counts
above background
• Plausible to assume
proportional to
epidemic curve of
infected
• Sartwell lognormal
model gives idealized
shape for a given
disease type
observed
modeled
10
8
6
4
2
0
0
10
20
30
40
Days after Exposure
Sartwell, PE. The distribution of incubation periods of infectious
disease. Am J Hyg 1950; 51:310:318
50
Signal Modeling: Realizations
of Smallpox Epicurve
“maximum
likelihood”
epicurve
Each symptomatic
case a random draw
Assessing Algorithm
Performance
Summary processing: measure dependence of sensitivity or
timeliness on false alert rate (ROC or AMOC curves or key
sample values at practical rates)
Sensitivity/Specificity as
a function of threshold:
Receiver Operating Characteristic
(ROC)
Timeliness/Specificity as
a function of threshold:
Activity Monitor Operating
Characteristic
(AMOC)
False Alert Rate
(1 – specificity)
False Alert Rate
(1 – specificity)
Detection Performance Comparison
Fever_Labbaji, lognormal signal
1.00
EWMA
0.90
EARS C2
EARS C3 (CUSUM)
Detection Probability
0.80
0.70
0.60
0.50
0.40
0.30
0.20
0.10
0.00
0
10
20
30
40
50
60
Background Recurrence (days)
70
80
90
Quality Control Charts and
Health Surveillance
Benneyan JC, Statistical Quality Control Methods in Infection Control and
Hospital Epidemiology, Infection and Hospital Epidemiology, Vol. 19,
(3)194-214
Part I: Introduction and Basic Theory
Part II: Chart use, statistical properties, and research issues
• 1998 Survey article gives 135 references
• Many applications: monitoring surgical wound infections, treatment
effectiveness, general nosocomial infection rate, …
Monitoring process for “special causes” of variation
• Organize data into fixed-size groups of observations
• Look for out-of-control conditions by monitoring mean, standard
deviation,…
• General 2-phase procedure:
Phase I: Determine mean m, standard deviation s of process from historical
“in-control” data; control limits often set to m  3s
Phase II: Apply control limits prospectively to monitor process graphically
Adaptation of Traditional Process Control to
Early Outbreak Detection
On adapting statistical quality control to biosurveillance:
Woodall , W.H. (2000). “Controversies and Communications in
Statistical Process Control”, Journal of Quality Technology 32, pp.
341-378.
• “Researchers rarely…put their narrow contributions into the context
of an overall SPC strategy. There is a role for theory, but theory is
not the primary ingredient in most successful applications.”
Woodall , W.H. (2006, in press). “The Use of Control Charts in Health
Care Monitoring and Public Health Surveillance”
•
•
“In industrial quality control it has been beneficial to carefully distinguish
between the Phase I analysis of historical data and the Phase II monitoring
stage”
“It is recommended that a clearer distinction be made in health-related SPC
between Phase I and Phase II…”
Does infectious disease surveillance require an “ongoing Phase I”
strategy to maintain robust performance?
Statistical Process Control
in Advanced Disease Surveillance
Key application issues:
• Background data characteristics change
over time
– Hospital/clinic visits, consumer purchases not governed
by physical science, engineering
– But monitoring requires robust performance: algorithms
must be adaptive
• Target signal: effect of infectious disease outbreak
– Transient signal, not a mean shift
– May be sudden or gradual
The Challenge of Data Modeling
for Daily Health Surveillance
• Conventional scientific application of regression
– Do covariates such as age, gender affect treatment?
Does treatment success of differ among sites if we
control for covariates?
– Studies use static data sets with exploratory analysis
• In surveillance, we model to predict data levels in
the absence of the signal of interest
– Need reliable estimates of expected levels to recognize
abnormal levels
– Data sets dynamic—covariate relationships change
The Challenge of Data Modeling for
Daily Health Surveillance, cont’d
Modeling to generate expected data levels
–
–
Predictive accuracy matters, not just strength of
association or overall goodness-of-fit
For a gradual outbreak, recent data can “train” model
to predict abnormal levels
Alerting decisions based on model residuals
Residual = observed value – modeled value
Conventional approach:
–
–
assume residuals fit a known distribution (normal,
Poisson,…)
hypothesis test for membership in that distribution
For surveillance, can also apply control-chart
methods to residuals
Monitoring Data Series
with Systematic Features
• Problem: How to account for short-term
trends, cyclic data features in alerting
decisions?
• Approaches
– Data Modeling
• Regression: GLM, ARIMA, others & combinations
– Signal Processing
• LMS filters and wavelets
– Exponential Smoothing: generalizes EWMA
Example: OTC Purchasing Behavior
Influenced by Many Factors
Loglinear Regression
Example: Tracking Daily Sales of Flu Remedies
Log(Y) = b0 + b1-6d + b7t + b8-9h +b10w + b11p + e
daily count
of anti-flu
sales
day of week linear
(6 indicators) trend
harmonic
(seasonal)
weather
(temp.)
sales
promotion
(indicator)
deviation
(Poisson
dist.)
Recent Surveillance Method
Based on Loglinear Regression
Modeling emergency department visit patterns for infectious disease
complaints: results and application to disease surveillance
Judith C Brillman , Tom Burr , David Forslund , Edward Joyce , Rick
Picard and Edith Umland
BMC Medical Informatics and Decision Making 2005, 5:4, pp 1-14
http://www.biomedcentral.com/content/pdf/1472-6947-5-4.pdf
Modeling visit counts on day d:
Let S(d) = log ( visits(day d) + 1 ), the “started log”
S(d) = [Σi ci × Ii(d)] + [c8 + c9 × d] + [c10 × cos(kd) + c11 × sin(kd)],
k = 2π / 365.25
c1-c7
day-of-week effects
c9
long-term trend
c10-c11 seasonal harmonic terms
Training period: 3036 days ~ 8.33 years
Test period:
1 year
Brillman et. al. Figure 1
EWMA Monitoring
• Exponential Weighted
Moving Average
• Average with most weight
on recent Xk:
Sk = wS k-1 + (1-w)Xk,
where 0 < w < 1
• Test statistic:
Sk compared to
expectation from sliding
baseline
Basic idea: monitor
(Sk – mk) / sk
Exponential Weighted Moving Average
60
Daily Count
Smoothed
50
40
30
20
10
0
02/25/94
•
•
03/02/94
03/07/94
03/12/94
03/17/94
03/22/94
03/27/94
Added sensitivity for gradual events
Larger w means less smoothing
04/01/94
EWMA Concept & Smoothing Constant
Brown, R.G. and Meyer, R.F. (1961), "The Fundamental
Theorem of Exponential Smoothing," Operations
Research, 9, 673-685.
• Exponential smoothing represents “an elementary
model of how a person learns”:
xk = xk-1 + w (xk - xk-1)
where 0 < w < 1
• For the smoothed value Sk,
Sk = wS k-1 + (1-w)Xk ,
The variance of Sk is sS = [w / (2 - w)] sX
• So a smaller w is preferred because it gives a more
stable Sk; values between 0.1 and 0.3 often used
• But Chatfield: changes in global behavior will result in
a larger optimal w
Generalized Exponential Smoothing
Holt-Winters Method: modeling level, trend, and seasonality
http://www.statistics.gov.uk/iosmethodology/downloads/
Annex_B_The_Holt-Winters_forecasting_method.pdf
Forecast Function:
yˆ
n  k |n
=
(m  k b ) (c
n
n
n-sk
)
where: mj = level at time j,
bj = trend at time j,
cj = periodic multiplier at time j
s = periodic interval
k = number of steps ahead
and mj, bj, cj are updated by exponential smoothing
Holt-Winters Updating Equations
Updating Equations, multiplicative method:
Level at time t:
yt
mt = 
 (1 -  ) ( mt -1  bt -1 ) ,
ct - s
0 <  <1
bt = b ( mt - mt - 1 )  (1- b ) bt - 1 ,
0 < b <1
yt
Periodic multiplier ct = 
 (1 -  ) ct -s ,
0 <  <1
mt
at time t:
Slope at time t:
And choice of initial values m0, b0, c0,…cs-1 should be
calculated from available data
Forecasting Local Linearity:
Automatic vs Nonautomatic Methods
Chatfield, C. (1978), "The Holt-Winters Forecasting Procedure," Applied
Statistics, 27, 264-279.
Chatfield, C.and Yar, M. (1988), "Holt-Winters Forecasting: Some Practical
Issues, " The Statistician, 37, 129-140.
• “Modern thinking favors local linearity rather than global linear regression
in time…”
• “Local linearity is also implicit in ARIMA modelling…”
– Simple EWMA ~ ARIMA(0,1,1)
– EWMA + trend ~ ARIMA(0,2,2)
– Multiplicative Holt-Winters has no ARIMA equivalent
• “Practical considerations rule out [Box-Jenkins] if there are insufficient
observations or …expertise available”
– “Box-Jenkins… requires the user to identify an appropriate… [ARIMA]
model”
For “fair” comparison of H-W to B-J, have both automatic or nonautomatic.
Assertion: The simplicity of H-W permits easier classification, requiring less
historic data.
Can an automatic B-J give robust forecasting over a range of input series
types?
Regression vs Holt-Winters
Results for Data Set: 1; with DOW and Seasonal Variation
HW-RMSE = [57.401] RegressedRMSE = [61.1454]
600
Raw Data
Holt Winters
Regression
500
Ongoing study with
Galit Shmueli, U. of MD
Sean Murphy, JHU/APL
Counts
400
300
200
100
0
30 time series,
700 days’ data
5 cities
3 data types
2 syndromes
50
100
150
200
250
300
350
200
250
300
350
Days
Residuals
400
HoltWinters
Regression
200
Counts
Respiratory: seasonal &
day-of-week behavior
Gastrointestinal:
day-of-week effects
0
0
-200
-400
0
50
100
150
Days
Temporal Aggregation for Adaptive Alerting
Data stream(s) to monitor in time:
baseline interval
Used to get some estimate
of normal data behavior
• Mean, variance
• Regression coefficients
• Expected covariate distrib.
-- spatial
-- age category
-- % of claims/syndrome
guardband
test interval
Avoids
• Counts to be
contamination
tested for
of baseline
anomaly
with outbreak • Nominally 1 day
signal
• Longer to reduce
noise, test for
epicurve shape
• Will shorten as
data acquisition
improves
Candidate Methods
1. Global loglinear regression of Brillman et. al.
2. Holt-Winters exponential smoothing
fixed sets of smoothing parameters for data:
with both day-of-week & seasonal behavior
with only day-of-week behavior
3. Adaptive Regression
Log(Y) = b0 + b1-6d + b7t + b8hol + b9posthol + e
56-day baseline, 2-day guardband
b1-6 = day-of-week indicator coefficient
b7 = centered ramp coefficient
b8 = coefficient for holiday indicator
b9 = coefficient for post-holiday indicator
1-day ahead and 7-day-ahead predictions
Respiratory Visit Count Data
--- Data
--- Holt-Winters
--- Regression
--- Adaptive Regr.
All series display this autocorrelation;
good test for published regression model
GI Visit Count Data
--- Data
--- Holt-Winters
--- Regression
--- Adaptive Regr.
Stratified Residual Comparisons
--- Data
--- Holt-Winters
--- Regression
--- Adaptive Regr.
Mean Residual Comparison
• When mean residuals favor regression, difference is small, and this difference
results from largest residuals
• If the holiday terms in adaptive regression are removed, H-W means uniformly
smaller
Median Residual Comparison
Residual Autocorrelation Comparison
--- Data
--- Holt-Winters
--- Regression
--- Adaptive Regr.
Residual Autocorrelation Comparison
1-Day Ahead Predictions
Residual Autocorrelation Comparison
7-Day Ahead Predictions
Summary
• Data-adaptive methods are required for robust
prospective surveillance
• Appropriate algorithm selection requires an
automated data classification methodology,
often with little data history
• Statistical expertise is required to manage
practical issues to maintain required detection
performance as datasets evolve:
– stationarity (causes rooted in population behavior,
evolving informatics, others)
– late reporting
– data dropouts
Research Directions
• Classification of time series for automatic
forecasting
– Easier for Holt-Winters than for Box-Jenkins?
– Determining reliable discriminants:
• Autocorrelation coefficients
• Simple means/medians
• Goodness-of-fit measures
– How little startup data history required?
• Most effective alerting algorithm using residuals,
given signal of interest
– Apply control chart to residuals?
– Need to detect both sudden, gradual signals
– Detection performance constraints:
• Minimum detection sensitivity
• Maximum background alert rate