DUI Detection From Gait Using a Multichannel 1DCNN-Attention-BiLSTM Framework.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Transportation Safety

Background:

• Alcohol intoxication significantly impairs functions crucial for driving, contributing to a substantial portion of traffic fatalities.
• Current methods for detecting Driving Under Influence (DUI) are invasive and not suitable for continuous monitoring.
• Existing machine learning and deep learning approaches for passive impairment detection face challenges like feature engineering and data variability.

Purpose of the Study:

• To propose a novel deep learning framework for non-invasive alcohol intoxication detection using smartphone accelerometer data.
• To address limitations of prior methods, including inter- and intra-subject variability and class imbalance.
• To develop a practical and continuous monitoring solution for road safety.

Main Methods:

• A subject-level pre-processing pipeline including filtering, segmentation, and oversampling was used to handle data variability and class imbalance.
• A Multichannel Hybrid 1D-CNN-Attention-BiLSTM (MC-Hybrid) model was developed, integrating 1D-CNNs for feature extraction, self-attention for pattern weighting, and BiLSTM for temporal analysis.
• The model's performance was rigorously evaluated against various machine learning and deep learning baselines, with investigations into window sizes and attention mechanisms.

Main Results:

• The proposed MC-Hybrid model achieved 93% accuracy and an F1-score of 0.8653, significantly outperforming state-of-the-art and baseline methods.
• The self-attention mechanism contributed a 2% performance improvement compared to other attention types, highlighting its effectiveness in identifying critical gait patterns.
• The framework demonstrated robustness in handling gait variability and class imbalance inherent in smartphone sensor data.

Conclusions:

• The novel MC-Hybrid deep learning framework offers a highly accurate and practical non-invasive method for detecting alcohol impairment using smartphone accelerometers.
• This approach has the potential to significantly enhance road safety by enabling continuous monitoring for Driving Under Influence (DUI).
• The study underscores the efficacy of deep learning, particularly attention mechanisms, in analyzing complex sensor data for real-world safety applications.