Using drinking data and pharmacokinetic modeling to calibrate transport model and blind deconvolution based data analysis software for transdermal alcohol biosensors

Abstract

Alcohol researchers/clinicians have two ways to collect subject /patient field data, standard-drink self-report and the breath analyzer, neither of which is passive or accurate because active subject participation is required. Transdermal alcohol sensors have been developed to measure transdermal alcohol concentration (TAC), but they are used primarily as abstinence monitors because converting TAC into more meaningful blood/breath alcohol concentration (BAC/BrAC) is difficult. In this paper, BAC/BrAC is estimated from TAC by first calibrating forward distributed parameter-based convolution models for ethanol transport from the blood through the skin using patient-collected drinking data for a single drinking episode and a nonlinear pharmacokinetic metabolic absorption/elimination model to estimate BAC. TAC and estimated BAC are then used to fit the forward convolution filter. Nonlinear least squares with adjoint-based gradient computation are used to fit both models. Calibration results are compared with those obtained using BAC/BrAC from alcohol challenges and from standard, linear, metabolic absorption, and zero order kinetics-based elimination models, by considering peak BAC, time of peak, and area under the BAC curve. Our models (with population parameters) could be included in a smart phone app that makes it convenient for the subject/patient to enter drinking data for a single episode in the field.

Figures

Figure 1

Two examples of BrAC and corresponding TAC with the TAC exhibiting markedly different attenuation and latency properties.

Figure 1

Two examples of BrAC and corresponding TAC with the TAC exhibiting markedly different attenuation and latency properties.

Figure 2

Schematic diagram of a calibration scheme for data analysis software based on drinking data rather than alcohol challenge data. A nonlinear pharmacokinetic Michaelis –Menten alcohol clearance model is used to estimate BAC from a patient supplied drinking diary. The resulting estimated BAC is then used with TAC data from the biosensor to calibrate the models used to deconvolve BAC from TAC in subsequent drinking episodes for which no drink diary is available.

Figure 3

Contemporaneous TAC and BrAC data for 11 drinking episodes recorded by a single subject engaging in naturalistic drinking in the field.

Figure 4

A comparison of estimated BAC for drinking episode 1obtained using Methods 1 – 9 based on drink diary input. The triangles are the breath analyzer measured BrAC. It is not surprising that Method 7 which uses the Michaelis - Menten alcohol clearance model with parameters fit based on drinking episode 1data is the most accurate. The Michaelis - Menten model with the mean (over all 11drinking episodes) parameters (method 8) also does reasonably well. As one would expect, the Michaelis - Menten models out - perform the 0 - order kinetics based models at low BAC.

Figure 5

Forward convolution kernels calibrated using TAC data from the first drinking episode and either BrAC data obtained via an alcohol challenge or by using estimated BAC based on drinking diary data computed using the Michaelis Menten alcohol clearance model with parameters as in methods 7, 8, or 9.

Figure 6

TAC and BrAC data for the first drinking episode in Figure 3, along with estimated BAC or BrAC deconvolved from the TAC data from the first drinking episode using a forward model that was calibrated using either BrAC data obtained via an alcohol challenge or estimated BAC based on drinking diary data generated by our nonlinear pharmacokinetic model with a Michaelis Menten alcohol clearance term with parameters determined by (1) training using data from drinking episode 1 (solid line) (i.e. method 7), (2) averaging the parameters obtained by training on each of the 11 drinking episodes in Figure 3 (dashed line) (i.e. method 8, or (3) training on all 11 drinking episodes in Figure 3 simultaneously (dashed-dotted line (i.e. method 9).