Data Assimilation Methods in
Parameter Estimation: An Application
to Tuberculosis Transmission Model
IIT Mandi, Himachal Pradesh
Pankaj Narula and Arjun Bhardwaj
Supervisors:
Dr. Sarita Azad
Dr. Ankit Bansal
International Conference on MathematicalTechniques in Engineering
Applications (ICMTEA 2013)
Outline
 Epidemiology of Tuberculosis (TB)
 Model Formulation and Parameters
 Research Interests
 PreviousWork onTB
 Estimation Methods
 Results
Epidemiology of Tuberculosis
 TB is one of the most widespread
infectious diseases, and a leading cause of
global mortality.
 Particularly,TB in India accounts for 25%
of the world’s incident cases.
 RNTCP is being implemented by
Government of India in the country with
DOTS strategy.
The SIR Epidemic Model
S
Susceptible:
can catch the disease
I
Infectious:
have caught the disease and can spread it
to susceptible
R
Recovered:
have recovered from the disease and
are immune.
dS
dt
= – b S I
dI
dt
= b S I – γI
dR
dt
= γ I
S + I + R = 1
Parameters of the Model
 = The infection rate
 = The Removal rate
 = Fraction of infectious persons.
 Basic reproduction number obtained as:
 Average secondary number of infections
caused by an infective in total susceptible
population. An epidemic occur if .
 Fraction of population needs to be vaccinated
1 − 1/R0.
0
p
R
b


b

p
0 1R 
Model Formulation
 SILS Model ofTB
Recovery rate is assumed to be 0.85.
Aim is to estimate β and p.
( )
(1 )
dS IS
I L
dt N
dI pIS
I tL
dt N
dL p IS
I tL
dt N
b

b

b


  

  
 
  
Research Interests
 Mathematical models, deterministic or
statistical, are important tools to
understand TB dynamics and analyse
voluminous data collected by various
agencies likeWHO, RNTCP.
 Challenge is to accurately estimate
model parameters.
 Parameters like infection rate measure
the disease burden and evaluate the
measures for control.
Previous work On TB
 Parameter Estimation of Tuberculosis
Transmission Model using Ensemble
Kalman Filter; Vihari et al. (2013)
 Bayesian Melding Estimation of a
Stochastic SEIR Model, Hotta et al. (2010)
 Tuberculosis in intra-urban settings: a
Bayesian approach; Souza et al. (2007)
Methods of Parameter Estimation
 Least Square
 Maximum Likelihood Method
 Ensemble Kalman Filter
 Bayesian Melding
Maximum Likelihood Method
 To estimate a density function
whose parameters are
 as an ML estimate of
( )p x
1
( ) ( / )
n
i
i
L P x 

 



1 2( , ,....., )t
m   
argmax[ ( )]L 


Ensemble Kalman Filter (EnKf)
 The EnKf is a MC approximation of the
Kalman filter.
 It avoids evolving the covariance matrix of
the pdf of the state vector.
 The basic idea is to predict the values first
and then to adjust it by actual value.
Ensemble Kalman Filter
 Forecast Step
 Ensemble Mean
 Error Matrices
 Analysis Step
1
j j jp p
t t tk k    1,2,3,....,j m
1 1
1
1 j
m
pp
t t
j
k k
m
 

 
1
1
1 1 1 1
1 1 1 1
[ ....... ]
[ ....... ]
q
t
q
t
ppp p p
k t t t t
ppp p p
y t t t t
E k k k k
E y y y y
   
   
  
  
( [ ] )k p j fj j jt t tt t t
k k K y v y   
Bayesian Melding Method
 Bayesian melding which observes the
existence of two priors, explicit and
implicit, on every input and output.
 The technique works good with
stochastic and deterministic models with
in high dimensional parameter estimation.
Bayesian Melding Method
 These priors are coupled via logarithmic
pooling.
 It calibrates the knowledge and
uncertainty of inputs and outputs of the
model.
 The technique ignores the Borel paradox.
Results
 We have used BIP. Bayes.Melding package
to estimate trend of various parameters.
 We have used Fitmodel for Monte Carlo
simulations, 2000 samples were discarded.
Prior distributions for parameters are
taken to be normal.
Results
 The parameter estimation framework
presented here captures seasonality well
in the data which could not be expected
from standard-likelihood methods.
 The estimates presented here are verified
from three different approaches.
Results
 Comparison of parameters values
 A- our results (EnKf)(2011)
 B- Our results (Bayesian Melding)
 C-Christopher Dye(2012)
 * 8 secondary infections per year.
Parameters A(2011) B C
β 1.72 1.84* 3.5
p 0.6 0.30 0.45
R0 1.29 0.69 0.78
Results
Comparison of estimated values of β for
highest infected state Manipur from three
different approaches
Bayesian Melding
Mean value = 1.90
EnKf
Mean value = 2.31
Least square
Mean value = 2.32
Results
Estimated values of parameters for India
Ro< 1 which shows the disease is endemic in
the country
Seasonal trend
in the plot of
β and Ro
00.35 0.94R 
1.3 2.54b 
Results
Ranges of R0
TB transmission across various Indian states
THANKYOU

Bayesian Estimation of Reproductive Number for Tuberculosis in India

  • 1.
    Data Assimilation Methodsin Parameter Estimation: An Application to Tuberculosis Transmission Model IIT Mandi, Himachal Pradesh Pankaj Narula and Arjun Bhardwaj Supervisors: Dr. Sarita Azad Dr. Ankit Bansal International Conference on MathematicalTechniques in Engineering Applications (ICMTEA 2013)
  • 2.
    Outline  Epidemiology ofTuberculosis (TB)  Model Formulation and Parameters  Research Interests  PreviousWork onTB  Estimation Methods  Results
  • 3.
    Epidemiology of Tuberculosis TB is one of the most widespread infectious diseases, and a leading cause of global mortality.  Particularly,TB in India accounts for 25% of the world’s incident cases.  RNTCP is being implemented by Government of India in the country with DOTS strategy.
  • 4.
    The SIR EpidemicModel S Susceptible: can catch the disease I Infectious: have caught the disease and can spread it to susceptible R Recovered: have recovered from the disease and are immune. dS dt = – b S I dI dt = b S I – γI dR dt = γ I S + I + R = 1
  • 5.
    Parameters of theModel  = The infection rate  = The Removal rate  = Fraction of infectious persons.  Basic reproduction number obtained as:  Average secondary number of infections caused by an infective in total susceptible population. An epidemic occur if .  Fraction of population needs to be vaccinated 1 − 1/R0. 0 p R b   b  p 0 1R 
  • 6.
    Model Formulation  SILSModel ofTB Recovery rate is assumed to be 0.85. Aim is to estimate β and p. ( ) (1 ) dS IS I L dt N dI pIS I tL dt N dL p IS I tL dt N b  b  b              
  • 7.
    Research Interests  Mathematicalmodels, deterministic or statistical, are important tools to understand TB dynamics and analyse voluminous data collected by various agencies likeWHO, RNTCP.  Challenge is to accurately estimate model parameters.  Parameters like infection rate measure the disease burden and evaluate the measures for control.
  • 8.
    Previous work OnTB  Parameter Estimation of Tuberculosis Transmission Model using Ensemble Kalman Filter; Vihari et al. (2013)  Bayesian Melding Estimation of a Stochastic SEIR Model, Hotta et al. (2010)  Tuberculosis in intra-urban settings: a Bayesian approach; Souza et al. (2007)
  • 9.
    Methods of ParameterEstimation  Least Square  Maximum Likelihood Method  Ensemble Kalman Filter  Bayesian Melding
  • 10.
    Maximum Likelihood Method To estimate a density function whose parameters are  as an ML estimate of ( )p x 1 ( ) ( / ) n i i L P x        1 2( , ,....., )t m    argmax[ ( )]L   
  • 11.
    Ensemble Kalman Filter(EnKf)  The EnKf is a MC approximation of the Kalman filter.  It avoids evolving the covariance matrix of the pdf of the state vector.  The basic idea is to predict the values first and then to adjust it by actual value.
  • 12.
    Ensemble Kalman Filter Forecast Step  Ensemble Mean  Error Matrices  Analysis Step 1 j j jp p t t tk k    1,2,3,....,j m 1 1 1 1 j m pp t t j k k m      1 1 1 1 1 1 1 1 1 1 [ ....... ] [ ....... ] q t q t ppp p p k t t t t ppp p p y t t t t E k k k k E y y y y               ( [ ] )k p j fj j jt t tt t t k k K y v y   
  • 13.
    Bayesian Melding Method Bayesian melding which observes the existence of two priors, explicit and implicit, on every input and output.  The technique works good with stochastic and deterministic models with in high dimensional parameter estimation.
  • 14.
    Bayesian Melding Method These priors are coupled via logarithmic pooling.  It calibrates the knowledge and uncertainty of inputs and outputs of the model.  The technique ignores the Borel paradox.
  • 15.
    Results  We haveused BIP. Bayes.Melding package to estimate trend of various parameters.  We have used Fitmodel for Monte Carlo simulations, 2000 samples were discarded. Prior distributions for parameters are taken to be normal.
  • 16.
    Results  The parameterestimation framework presented here captures seasonality well in the data which could not be expected from standard-likelihood methods.  The estimates presented here are verified from three different approaches.
  • 17.
    Results  Comparison ofparameters values  A- our results (EnKf)(2011)  B- Our results (Bayesian Melding)  C-Christopher Dye(2012)  * 8 secondary infections per year. Parameters A(2011) B C β 1.72 1.84* 3.5 p 0.6 0.30 0.45 R0 1.29 0.69 0.78
  • 18.
    Results Comparison of estimatedvalues of β for highest infected state Manipur from three different approaches Bayesian Melding Mean value = 1.90 EnKf Mean value = 2.31 Least square Mean value = 2.32
  • 19.
    Results Estimated values ofparameters for India Ro< 1 which shows the disease is endemic in the country Seasonal trend in the plot of β and Ro 00.35 0.94R  1.3 2.54b 
  • 20.
    Results Ranges of R0 TBtransmission across various Indian states
  • 21.