Custom PK models and TCI algorithms using the tci package
In tci: Target Controlled Infusion (TCI)

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  echo = TRUE, 
  message = FALSE, 
  warning = FALSE, 
  fig.align="center",
  fig.height= 4, 
  fig.width = 6
)

Introduction

The tci package provides closed-form solutions for 1-, 2-, 3-, and 3-compartment with effect-site PK models based on solutions described and code provided in @Abuhelwa2015. It does not, however, include an ordinary differential equation (ODE) solver as many other R packages do, including mrgsolve, PKPDsim, and RxODE. tci package functions can nonetheless be applied to models from these packages through the creation of user-defined PK functions.

Custom PK models

Custom PK models based on ODEs or analytical solutions can be passed to pkmod objects through the pkfn argument. To illustrate this functionality, we consider the PK of the analgesic remifentanil. Remifentanil is an opioid derivative that is often administered intravenously to induce analgesia alongside propofol. Here, we consider the three-compartment PK model proposed by \citet{Cascone2013}. Remifentanil is infused into a central compartment, representing the blood supply, and then circulated to two peripheral compartments, representing highly-perfused and scarcely-perfused organs and tissues. Remifentanil is then removed from all three-compartments with separate clearance rates. A diagram of the three-compartment model, reproduced from \citet{Cascone2013}, is displayed in figure \ref{fig:Cascone_fig2}. The differential equations describing the remifentanyl model are given in equations \ref{eq:remif}.

\begin{center} \begin{figure} \includegraphics[width=0.8\linewidth]{figures/Cascone_fig2.png} \caption{Reproduction Figure 2 from Cascone et al. (2013)}\label{fig:Cascone_fig2} \end{figure} \end{center}

\begin{equation}\label{eq:remif} \begin{aligned} V_1 \cdot \frac{dC_p}{dt} &= -Cl_1 \cdot C_1 + k_{21}\cdot V_2 \cdot C_2 + k_{31}\cdot C_3 \cdot V_3 + -[(k_{12}+k_{13}+k_{10})\cdot C_1] \cdot V_1 + k_R(t) \ V_2 \cdot \frac{dC_2}{dt} &= k_{12} \cdot C_1 \cdot V_1 - k_{21} \cdot C_2 \cdot V_2 - Cl_2 \cdot C_2 \ V_3 \cdot \frac{dC_3}{dt} &= k_{13} \cdot C_1 \cdot V_1 - k_{31} \cdot C_3 \cdot V_3 - Cl_3 \cdot C_3 \end{aligned} \end{equation}

To implement the ODE system in mrgsolve we initialize a model using mrgsolve::mcode with default parameter values.

library(tci)
library(mrgsolve) # implement ODE equation
library(xtable)   # printing tables
library(ggplot2)  # plotting results
library(reshape2) # melt function

form <- '
$PARAM V1 = 7.88, V2=23.9, V3=13.8, CL1=5, CL2=0.828, CL3=0.0784,
k10 = 0.172, k12=0.373, k21=0.103, k13=0.0367, k31=0.0124
$CMT A1 A2 A3
$ODE
dxdt_A1 = k21*A2 + k31*A3 - (k12+k13+k10)*A1 - CL1/V1*A1;
dxdt_A2 = k12*A1 - k21*A2 - CL2/V2*A2;
dxdt_A3 = k13*A1 - k31*A3 - CL3/V3*A3;
'
mrg_mod_remif <- mcode("remifentanil", form)

For a custom model to be compatible with PK, it must take as arguments 1) a vector of time points, tm, 2) an numeric value describing a constant infusion rate, kR, 3) a vector of PK parameter values, pars, and 4) initial starting concentrations, init. Notably, init should be created with default values, as pkmod will use the initial values to determine the number of compartments in the model.

pk_remif <- function(tm, kR, pars, init = c(0,0,0)){

  # allow lowercase names
  names(pars) <- toupper(names(pars))
  # store volume
  vols <- pars[c("V1","V2","V3")]
  A0 <- init*vols # initial amounts
  names(A0) <- c("A1","A2","A3") # names required by mrgsolve

  # pass parameters as list
  pars <- sapply(pars, as.list)

  # update parameters and initial values (as amounts)
  mrg_mod_remif <- update(mrg_mod_remif, param = pars, init = A0)

  # dosing regimen - mrgsolve function in terms of amount infused
  event <- ev(amt =  kR*max(tm), time = 0, tinf = max(tm))

  # simulate responses (skip tm=0 unless specified)
  dat <- mrgsim_q(x = mrg_mod_remif, # pk model
                  data = event, # dosing event
                  stime = tm) # evaluation times

  # skip tm=0 unless specified in tm
  dat <- dat@data[-1,]

  # return concentrations with compartments in rows and times in columns
  cons <- t(dat[,c("A1","A2","A3")]) / vols

  rownames(cons) <- colnames(cons) <- NULL
  return(cons)
}

We can now evaluate the remifentanil PK model as we would any of the internal PK functions. The optimized parameter values identified by \citet{Cascone2013}, which we will use as an example, are reproduced in Table \ref{tab:Cascone-tab1}.

tab_cascone <- data.frame(Parameter1 = c("$V_1$","$V_2$","$V_3$","$CL_1$",
                                         "$CL_2$","$CL_3$"),
                          Optimized_val1 = c("7.88 L","23.9 L","13.8 L",
                                             "2.08 L/min","0.828 L/min","0.0784 L/min"),
                          Parameter2 = c("$k_{10}$","$k_{12}$","$k_{21}$","$k_{13}$","$k_{31}$",""),
                          Optimized_val2 = c("0.172/min","0.373/min","0.103/min","0.0367/min","0.0124/min",""))

names(tab_cascone) <- c("Parameter", "Optimized value", "Parameter", "Optimized value")
tab_cascone_df <- xtable(tab_cascone,
                         caption = "Reproduction of Table 1 from Cascone et al. (2013): Values and dimensions of the three-compartmental model parameters.",
                         label = "tab:Cascone-tab1")
print(tab_cascone_df,
      include.rownames = FALSE,
      caption.placement = "top",
      timestamp = NULL,
      booktabs = TRUE,
      hline.after = c(-1,0,6),
      sanitize.text.function = function(x){x},
      include.colnames=TRUE)

dose_remi <- inf_manual(inf_tms = 0, inf_rate = 60, duration = 20)
pars_remif <- c(V1 = 7.88, V2=23.9, V3=13.8, CL1=5, CL2=0.828, CL3=0.0784,
                k10 = 0.172, k12=0.373, k21=0.103, k13=0.0367, k31=0.0124)
mod_remif <- pkmod(pkfn = pk_remif, pars_pk = pars_remif)
p1 <- predict(mod_remif, inf = dose_remi, tms = 0:80)
ggplot(melt(data.frame(time = 0:80, p1), id = "time"), 
       aes(x = time, y = value, color = variable)) +
  geom_line()

Custom TCI algorithms

The tci package implements the Jacobs and Shafer algorithms plasma- and effect-site targeting algorithms, respectively [@Jacobs1990;@Shafer1992]. These algorithms aim to reach the target concentrations as quickly as possible without overshooting the target. There may, however, be situations in which the speed of target attainment is not the only goal. In these cases, a user may wish to specify a different TCI algorithm.

An example of this is the algorithm proposed by @VanPoucke2004 that limits the maximum percentage overshoot of the target in the central compartment. The motivation for this is that there may exist cases in which excessive plasma concentrations are associated with toxicity to the patient. Here, we construct a similar algorithm that limits the absolute, rather than the percentage, overshoot in the central compartment.

In this example algorithm the user specifies a permissible amount of overshoot in the central compartment, \texttt{lim_amt}, beyond the nominal target. At each step, the example TCI algorithm defines a maximum plasma concentration to equal the target effect-site concentration plus the permissible overshoot: \texttt{Cp_max\ =\ Cet\ +\ lim_amt}. It then calculates the infusion required to reach or maintain \texttt{Cp_max} over the subsequent ten seconds, \texttt{pinf}. It then calculates the maximum effect-site concentration if \texttt{pinf} is given, \texttt{Ce_max}. If \texttt{Ce_max} is less than the target concentration, \texttt{pinf} can be administered without overshoot. If \texttt{Ce_max} is greater than the target concentration, then targeting the effect-site directly will result in a maximum plasma concentration less than \texttt{Cp_max} and the effect-site targeting algorithm is applied.

The required arguments for a custom algorithm are 1) a single numeric value specifying the target concentration, Ct, 2) a pkmod object created by pkmod(), 3) a single numeric value specifying the infusion duration, dtm, and 4) additional arguments that are passed to update.pkmod at the beginning of the algorithm. Argument (4) is used to update starting concentrations when the algorithm is iteratively applied in inf_tci.

tci_plasma_lim <- function(Ct, pkmod, dtm = 1/6, maxrt = 1200,
                          lim_amt = 0.5, ecmpt = NULL, tmax_search = 20, 
                          cetol = 0.05, cptol = 0.1, ...){

  pkmod <- update(pkmod,...)

  # if effect-site concentration is close to target,
  # switch to plasma targeting
  if(with(pkmod,(Ct - init[ecmpt]) / Ct <  cetol &
     (Ct - init[pcmpt])/Ct <= cptol))
    return(tci_plasma(Ct, pkmod = pkmod, dtm = dtm, maxrt = maxrt))

  # maximum tolerable plasma concentration
  Cp_max <- Ct + lim_amt

  # infusion required to reach Cp_max
  pinf <- tci_plasma(Ct = Cp_max, pkmod = pkmod, dtm = dtm, maxrt = maxrt)

  # Administer dtm-minute infusion
  unit_inf <- inf_manual(inf_tms = 0, inf_rate = pinf, duration = dtm)

  # Calculate maximum effect-site concentration
  CeP <- function(tm) predict(pkmod, inf = unit_inf, tms = tm)[,pkmod$ecmpt]
  Ce_max <- optimize(CeP, c(0,20), maximum = TRUE)$objective

  # if max Ce < Ct administer infusion to reach maximum target
  if(Ce_max <= Ct + cetol*Ct) 
    infrt <- pinf
  else
    infrt <- tci_effect_only(Ct, pkmod, dtm, maxrt = maxrt)

  return(infrt)
}

We can now apply the algorithm directly to a pkmod object to calculate a single infusion rate.

mod3ecpt <- pkmod(pars_pk = c(cl = 10, q2 = 2, q3 =20, v = 15, v2 = 30, v3 = 50, ke0 = 1.2))
tci_plasma_lim(Ct = 2, pkmod = mod3ecpt, lim_amt = 0.25)

More usefully, however, we can pass the algorithm to inf_tci through the custom_alg argument and use it to calculate infusion rates required to reach a series of targets.

```r"}

tci target concentrations

tci_targets <- cbind(value = c(1,2,2.5,2.5), time = c(0,3,7,10))

calculate infusion schedule using plasma-limiting algorithm

plim_inf <- inf_tci(target_vals = c(1,2,2.5,2.5), target_tms = c(0,3,7,10), pkmod = mod3ecpt, custom_alg = tci_plasma_lim, lim_amt = 0.25) head(plim_inf)

For comparison, we calculate the infusion schedule associated with direct effect-site targeting.

```r
# effect-site targeting
eff_inf <- inf_tci(target_vals = c(1,2,2.5,2.5),
                    target_tms = c(0,3,7,10), 
                   pkmod = mod3ecpt, 
                   type = "effect")

We now can use the infusion schedule in predict.pkmod or plot.pkmod methods

# predict responses
tms_pred <- seq(0,10,0.1)
plim_pred <- predict(mod3ecpt, plim_inf, tms_pred)
eff_pred <- predict(mod3ecpt, eff_inf, tms_pred)

# plot results
dat <- data.frame(time = tms_pred, 
                  `plasma (custom)` = plim_pred[,"c1"], 
                  `effect (custom)` = plim_pred[,"c4"],
                  `plasma (effect)` = eff_pred[,"c1"], 
                  `effect (effect)` = eff_pred[,"c4"],
                  check.names = FALSE)
datm <- melt(dat, id = "time")
datm$algorithm <- ifelse(datm$variable %in% c("plasma (custom)","effect (custom)"),
                         "Plasma-limiting", "Effect-site")
ggplot(datm, aes(x = time, y = value, color = variable, linetype = algorithm)) + 
  geom_line() + 
  xlab("Minutes") + 
  ylab("Concentration (mg/L)") +
  ggtitle(label = "Plasma-limiting effect-site TCI algorithm")