Signal processing and machine learning with transdermal alcohol concentration to predict natural environment alcohol...

Area of Science:

• Biomedical Engineering
• Data Science
• Alcohol Research

Background:

• Wrist-worn alcohol biosensors offer continuous, discreet monitoring of transdermal alcohol concentration (TAC) in naturalistic settings.
• Current limitations include a lack of standardized methods for signal processing and alcohol event detection.
• This gap hinders the reliable use of these devices in alcohol research.

Purpose of the Study:

• To develop and validate a software pipeline for processing wrist-worn alcohol biosensor data.
• To enhance the accuracy of detecting alcohol consumption events using signal processing and machine learning.
• To improve the efficiency and reliability of transdermal alcohol monitoring in research.

Main Methods:

• A novel software pipeline was created to process raw data (TAC, temperature, motion) from the BACtrack Skyn biosensor.
• Biologically implausible temperature readings were screened, and TAC artifacts were corrected.
• Machine learning models (random forest, logistic regression) were trained using TAC features to predict alcohol consumption.

Main Results:

• Data artifact correction improved model accuracy by 10% compared to raw data.
• Both random forest and logistic regression models achieved 97% accuracy in predicting alcohol consumption episodes.
• Key predictive features included area under the TAC curve, TAC rise duration, and peak TAC.

Conclusions:

• The developed software pipeline effectively processes wrist-worn alcohol biosensor data.
• Machine learning models demonstrate high accuracy in detecting alcohol consumption events.
• This protocol significantly enhances the efficiency and reliability of TAC sensors for future alcohol research.