Uncertainty Quantification in Estimating Blood Alcohol Concentration From Transdermal Alcohol Level With...

Area of Science:

• Biomedical Engineering
• Machine Learning
• Pharmacokinetics

Background:

• Transdermal alcohol sensors offer a non-invasive method for monitoring alcohol consumption.
• Estimating blood alcohol concentration (BAC) from transdermal signals is challenging due to complex physiological transport dynamics.
• Existing methods often lack robust uncertainty quantification.

Purpose of the Study:

• To develop a physics-informed neural network (PINN) framework for accurate blood alcohol signal estimation from transdermal alcohol signals.
• To incorporate uncertainty quantification into the estimation process.
• To validate the proposed method using real-world drinking episode data.

Main Methods:

• Utilized a generative adversarial network (GAN) with a residual-augmented loss function to model transdermal alcohol transport and estimate parameter distributions.
• Developed a secondary PINN for deconvolution to derive the blood alcohol signal from the transdermal signal.
• Employed a posterior latent variable to refine uncertainty estimates (error bands).

Main Results:

• The PINN-based approach successfully estimated blood alcohol signals from transdermal data.
• The method provided conservative error bands for uncertainty quantification.
• Sharpening of error bands was achieved using a posterior latent variable.
• Demonstrated advantages and limitations through application to an extensive dataset.

Conclusions:

• Physics-informed neural networks offer a powerful tool for estimating blood alcohol signals from transdermal data.
• The developed method provides reliable BAC estimation with crucial uncertainty quantification.
• This approach shows promise for improved non-invasive alcohol monitoring applications.