Transcript 投影片 1 - TU Dortmund

Develop a Tool for
Therapeutic Drug Monitoring
in R Using OpenBUGS
Speaker: Miao-ting Chen1, M.S.
Mentor: Yung-jin Lee2
1
Department of Hospital Pharmacy, Kaohsiung
Veteran General Hospital
2 College of Pharmacy, Kaohsiung Medical
University
Kaohsiung, Taiwan
1
Therapeutic Drug Monitoring (TDM)
To optimize individual patient’s drug
therapy through monitoring its serum
concentrations of the target drugs, as well
as the observed clinical response
Observation  estimate PK/PD parameters
 dosage adjustment
2
BUGS
The BUGS (Bayesian inference Using Gibbs Sampling):
Bayesian analysis of complex statistical models using
Markov chain Monte Carlo (MCMC) methods
likelihood
prior distribution
posterior
posterior distribution
prior
likelihood
3
Bayesian PK Hierarchical Model
(using warfarin as the example)
model {
for (i in 1:N) {
likelihood
INR[i]~dnorm(mu[i],1.0E+6)
mu[i]<-pow(a[i],(1/0.383))
a[i]<-((1/((m[i]*(cl_F[i]/v_F[i]))/(pow(kc[i],2))*(1(kc[i]*tau[i]/(1-exp(kc[i]*tau[i]))))m[i]/kc[i]*log
((D[i]/v_F[i])/(Cpmax[i]*(1-exp(cl_F[i]/v_F[i]*tau[i]
)))))+3.36)/4.368)
m[i]~dgamma(0.1,0.1)
Cpmax[i]~dgamma(0.1,0.1)
kc[i]~dgamma(0.1,0.1)
cl_F[i]~dgamma(0.1,0.1)
v_F[i]~dnorm(7.5,100)
Prior distribution
}}
4
Required programs or R packages
BRugs
OpenBUGS
5
Setting steps
bugsData( …….)
,fileName=file.path(getwd(),“modelname.txt")
,digits=5)
PK model
show(samplesStats("*"))
samplesHistory("*",mfrow=c(3,1), ask = FALSE)
modelcheck(“modelname.txt”)
modeldata(“dataname.txt”)
modelCompile(numChains=1)
modelGenInits(
modelUpdate(10000)
) samplesSet(c("ka","cl_F"))
modelUpdate(10000)
samplesDensity("*",
mfrow = c(3,
2), ask = FALSE)
samplesAutoC("*",1, mfrow = c(3, 2), ask = FALSE
6
Validation
The ability of the tdm package estimate
PE (Prediction Error, %) =
(Eq.1)
P pr= predicted value
P true= true values
Convergence of MCMC chain (history, density and
autocorrelation plots)
7
tdm
Menu driven UI
16 PK & 1 PD models
most steady-state (ss)
Four data types
single subject & one conc.
single subject & multiple conc.
many subjects & one conc.
many subjects & multiple
different conc.
Convergence plots
Dose adjustment
Menu
Aminoglycoside
Carbamazepine
History plot
Digoxin
Lithium Lithium carbonate
Lithium citrate
Theophylline salt Aminophylline anhydrous
Aminophylline dihydrous
Oxtriphylline
Auto-correlation plot
Theophylline
Phenytoin
Valproate
Vancomycin
Anti-HIV Enfuvirtide
Indinavir
Ritonavir
Density plot
Immunosuppressant Cyclosporine-A
Everolimus
Tacrolimus
Enoxaparin
Imatinib mesylate
8
Warfarin
Comparison Between tdm and JPKD
Prediction error (%) of PK
parameters were similar to those
using nonlinear regression (empirical
Bayesian) obtained from JPKD (Java
PK For Desktop).
9
Convergence
Low PE(%) is not necessarily imply
that Markov chains converge
successfully.
Also, successful convergence of
Markov chains do not necessarily
result in low PE(%).
In setting of tdm, we did not increase
updating for convergence.
10
Limitation of tdm
Currently tdm is only available for Windows
platform computer (BRugs and OpenBUGS are
now only available for Windows) .
ODE equation can not currently be used to define
model in tdm.
11
tdm website: http://pkpd.kmu.edu.tw/tdm
12
Acknowledge
•
Chun-ying Lee (Changhua Christian Hospital, Changhua,
Taiwan): package building and environment setting
•
Uwe Ligges (Fakultät Statistik, Technische Universität
Dortmund, Dortmund, Germany): coding and compiling (by email)
•
Kurt Hornik (Department of Statistics and Mathematics of the
Wirtschaftsuniversität Wien, Austria): coding and compiling (by
e-mail)
•
Kaohsiung Veteran General Hospital and Dr. Cheng DL
Medical Research Foundation, Kaohsiung, Taiwan: sponsoring
this trip
References
Yamaoka K, et al., A nonlinear multiple regression
program, MULTI2 (BAYES), based on Bayesian
algorithm for microcomputers. Journal of PharmacobioDynamic 1985;8: 246-56.
Application of Bayesian Estimation to a Two
compartment Model in PK/PD
OpenBUGS website:
http://mathstat.helsinki.fi/openbugs/Home.html
R website: www.r-project.org
14
Thanks for your attention
15