Estimation of Blood Alcohol Concentration From Smartphone Gait Data Using Neural Networks

Area of Science:

• Biomedical Engineering
• Machine Learning
• Transportation Safety

Background:

• Alcohol impairment significantly affects driving ability, leading to numerous accidents.
• Gait analysis offers a reliable, passive method for assessing intoxication levels.
• Current detection methods often require active participation or invasive procedures.

Purpose of the Study:

• To develop and validate a novel system for detecting driver intoxication using smartphone-based gait analysis.
• To predict Blood Alcohol Concentration (BAC) accurately by analyzing sensor data from smartphones.
• To leverage deep learning models for real-time, non-invasive intoxication detection.

Main Methods:

• Utilized a large, controlled alcohol study dataset.
• Employed Bi-directional Long Short-Term Memory (Bi-LSTM) and Convolutional Neural Network (CNN) architectures.
• Applied advanced pre-processing techniques and model-specific extensions for BAC estimation from smartphone accelerometer and gyroscope data.

Main Results:

• Achieved a Root Mean Square Error (RMSE) of 0.0167 with Bi-LSTM and 0.0168 with CNN.
• Outperformed state-of-the-art models including Bayesian Regularized Multilayer Perceptrons and Random Forest.
• Demonstrated superior generalizability with lower variance across data folds.

Conclusions:

• Smartphone gait analysis using deep learning is a highly effective method for estimating BAC.
• The proposed Bi-LSTM and CNN models offer a promising, non-invasive approach to proactive drunk driving prevention.
• Automated feature learning from raw sensor data eliminates the need for time-consuming manual feature engineering.