Click for code
= function(minute,dose,m,k,beta) {
exp_beta *dose*k*beta*(1-exp(-k*minute))^(beta-1)*exp(-k*minute)
m }
Dieter Menne
13C breath test data are evaluated in clinical practice and in research as an indirect measure of gastric emptying. The normative literature on the underlying concepts is given in the documentation of exp_beta()
, t50_maes_ghoos()
and t50_bluck_coward()
.
Here, I present a subjective viewpoint on the pharmakokinetics of breath test time series. After a look at the Maes/Ghoos method (Ghoos et al. (1993), Maes et al. (1998)) as it is currently used, alternatives are discussed in the next section.
Within the standard framework implemented in this package, data are fitted to an exponential beta function
\[{PDR} = m*{dose}*k*\beta(1-e^{-k*minute})^{(\beta - 1)}*e^{-k*minute}\]
The figure below shows examples of emptying curves. The so-called lag time of the Maes/Ghoos method is at the time of the maximum. Calling this “lag” is a highly confusing definition; technically, lag is the time until a first noticeable rise occurs in the curve, best illustrated by the blue curves with beta = 8
, which has an estimated lag of about 15 minutes. If one looks more closely, even this definition is misleading, because the curves with higher beta are not zero, but behave like a polynomial curve with a power of beta - 1
at the origin. A real lag would be best illustrated by a curve shifted to the right by t_lag
and pdr values of 0 for t < t_lag
, indicating a real physical transport of the meal from the stomach to the small intestine.
In practice, when 13C curves are fitted, beta
is typically in the range of 1.5 and 2.5; values of beta
near 1 often lead to instabilities and are protected by priors in the Bayesian breathteststan::stan_fit()
implemented in package breathteststan
.
The half-emptying as defined by Maes et al. (1998) is the time where the area under curve (AUC) is half the AUC extrapolated to infinity. In pharmakokinetics terms, it is the time where half of the bioavailable 13C has been metabolized. This time is determined both by gastric emptying and by the pharmacokinetics of 13C bound in octanoate/acetate.
The AUC extrapolated to infinity, which is the denominator in the definition of t50
, is a vulnerable variable. Even if we accept that the exponential beta function is a reasonable model for breath test time series, extrapolating individual curves from data only slightly longer than t50
results in ambiguous area estimates, often fails to converge, and even more often gives wildly romantic estimates of t50
.
To illustrate how brittle the AUC to infinity is, look back at a different function that had been used to fit 13C time series. Ghoos et al. (1993) used the following Gamma function to fit the data:
\[pdr = a*t^b*e^{-ct}\] As the figure below shows, this function gives perfectly valid fits of breath test time series. However, the area under curve extrapolated to infinity is infinite, so it cannot be used to compute a half-emptying time from the AUC.
Therefore, I recommend using one of the population methods, as implemented in nlme_fit()
and stan_fit()
to evaluated gastric emptying data in studies.
These methods provide “borrowing strength” to keep the flock of lambs safely together and protect outliers from the big bad wolf.
The half-emptying time determined with the 13C breath test method correlates to some extend with gastric emptying in within-subject measurements, but it is not more than a surrogate for gastric emptying times measured by MRI (with secretion) or scintigraphy (meal-only). For data set usz_13c_d
in this package, MRI emptying data are available and can be compared with breath test data. The type of meal strongly biases the estimates because of the the dependency on lipophilic/lipophobic layering; comparing different meal types is not reliable even within-subject, and less so between-subject.
Many attempts have been made to extract more information from the emptying curve, and to correct for the overall too high values; the Bluck-Coward method implemented in this package is one of them, but the results are are even less consistent with those from MRI than those from the Maes/Ghoos method. Sanaka et al. (2004) mention deconvolution and use an approximate correction for the pharmcological weighting function for the Wagner-Nelson method.
There are many publication that tried an ad-hoc scintigraphic correction (Keller et al. (2009), t50_maes_ghoos_scintigraphy()
) or one based on pharmacolokinetics (Bluck and Coward (2006), t50_bluck_coward()
). I do not know of any publication which has used rigorous statistical tests such as statistical cross-validation and validation with different meal types to show that some method gives “better” results than the default Maes/Ghoos method.
In the following, I argue that there is nothing to gain from playing with “better” methods to fit the breath test time series. Independent per-subject information must be available to separate acetate pharmacokinetics, based on concentration gradient diffusion, from gastric empyting, which is physical transport.
Assume we do not mix the labelled acetate with the meal, but hide it in acid-resistant coating as it is used for PPI tablets. No 13C will be recorded as long as the tablet is in the stomach, but it will quickly release 13C labelled acetate in the duodenum or small intestine, mimicing a one-time dose. As described by standard pharmacokinetics, this results in a PDR response following a first-order compartmental model. Forgetting about decorative normalization constants, the response is the difference between two exponential functions:
\[{foc} = e^{-k_1 * {minute}} - e^{-k_2 * {minute}}\] The normalization constant includes clearance and dose and only scales the curve. The correct normalization constants can be found in standard texts on pharmacokinetics (Gabrielsson and Weiner (2006)) or in as a base R function stats::SSfol()
in Pinheiro and Bates (2000).
The left panel in the above plot shows three examples of 13C fed by an enteric tables, remaining in the stomach for 0, 30 and 90 minutes and then quickly being released in the smaller intestine. Note that we have a real lag caused by physical transport here: before the tablet is in the smaller bowel, there is no response at all. This is different from the polynomial-like pseudo-lag with the beta-exponential function.
The right panel in the above plot shows the summary PDR curve when the three tablets are fed together and by chance are released after 0, 30 and 90 minutes. This mimics the case of a stomach releasing the meal in three chunks of the same size and composition, well known from pharmacokinetics for multiple doses of a drug.
The real flow out of the stomach is not pulse-like, but more like that of a continuous infusion with changing flow. When we assume that the stomach empties with a power-exponential function as in the left panel (see gastempt::powexp()
), the flow is a wide peak as shown in the right panel below. The PDR response technically is the convolution of the flow with the pulse response shown as red curve (t_lag = 0) in the above plot.
The correct half-emptying time t50
in this example was 82 minutes, as determined from the left panel; t50
determined by fitting with the Maes/Ghoos version in the right panel was 109 minutes.
The bias from fitting the PDR curve naively is caused by the convolution of mechanical transport in gastric emptying, and pharmacological drug kinetics. The latter is a nuisance parameter and should be eliminated - or is it a parameter of interest in its own?
I propose a revised procedure to separate pharmacology from transport. Patient first receive 30 mg acetate in an enteric table together with water. DOB data are recorded for 60 minutes, and after that the normal procedure of meal with 100 mg acetate/octanoate. The resulting PDR will look as follows:
The first narrow peak up to 1 hour is the pharmacological response; by fitting it, the kinetic time constants that broaden the wide peak can be determined, and can be used to remove the bias of the transport component visible in the main peak by mathematical deconvolution.
Can anyone supply a data set for a pilot?
A popular alternative when curve fitting does not work is the Wagner-Nelson method ((Sanaka?) (2004)), which uses a nonparametric approach to the initial slope. However, it uses the assumption that the final slope is the same for all subjects (k=0.01/min, 0.65/h), which strongly affects the estimate of half-emptying time, so there is little justification for using the Wagner-Nelson method.
With some of the functions in this package or with breathteststan::stan_fit()
in the sister package, all curves can be fitted. I have not seen a single example where the Bayesian method fails when multiple records are analyzed with a mix-in, so there is no excuse for using the Wagner-Nelson method any longer.
If you must, you can use function ComputeAndSaveWNFit in my legacy package D13CBreath
. The function is somewhat awkward to use because it has been written for an application with a tightly integrated database, but feel free to steal the code and run.