Advancing Intoxication Detection: A Smartwatch-Based Approach

Basic Information

• Paper ID: 2510.09916
• Title: Advancing Intoxication Detection: A Smartwatch-Based Approach
• Authors: Manuel E. Segura, Pere Vergés, Richard Ky, Ramesh Arangott, Angela Kristine Garcia, Thang Dinh Trong, Makoto Hyodo, Alexandru Nicolau, Tony Givargis, Sergio Gago-Masague
• Institutions: University of California Irvine, Asahi Group Holdings Ltd.
• Classification: cs.LG (Machine Learning)
• Publication Date: October 10, 2024 (arXiv preprint)
• Paper Link: https://arxiv.org/abs/2510.09916

Abstract

Excessive alcohol consumption poses severe health risks with serious consequences for individuals and communities. To promote healthier drinking habits, this research introduces a groundbreaking smartwatch mobile application approach for real-time intervention in intoxication warnings. The research team created a dataset containing TAC (transdermal alcohol concentration), accelerometer, gyroscope, and heart rate data collected over a three-week period. This is the first study combining accelerometer, gyroscope, and heart rate smartwatch data for long-term monitoring to classify intoxication levels. Using smartwatch data, the study conducted a comprehensive evaluation of multiple state-of-the-art classifiers, including Transformer, bidirectional LSTM, GRU, 1D CNN, and hyperdimensional computing (HDC). The HDC model achieved the optimal balance between accuracy and efficiency, demonstrating its practical utility in smartwatch applications.

Research Background and Motivation

Problem Significance

1. Public Health Challenge: Alcohol consumption leads to over 200 diseases, injuries, and health conditions. In 2023, alcohol abuse accounted for 6% of global deaths, with nearly 14% of deaths occurring in the 20-39 age group.
2. Limitations of Traditional Detection Methods: Conventional methods such as blood, urine, saliva testing, and breathalyzers are invasive, limiting their applicability in daily scenarios.
3. Need for Behavioral Intervention: There is a need for non-invasive solutions capable of promoting behavioral change and preventing excessive drinking or dangerous activities such as drunk driving.

Research Motivation

• Develop just-in-time adaptive intervention (JITAI) systems that provide personalized real-time intervention when users' wrist TAC exceeds 35µg/L (equivalent to BAC 0.05%)
• Leverage multi-sensor smartwatch data (accelerometer, gyroscope, heart rate) for long-term real-world monitoring
• Provide deeper understanding of users' physiological and behavioral states compared to traditional methods

Core Contributions

1. Novel Dataset: Created the first long-term intoxication detection dataset combining accelerometer, gyroscope, and heart rate smartwatch data, containing three weeks of data from 30 participants.
2. Comprehensive Multi-Model Evaluation: Conducted thorough evaluation of seven state-of-the-art machine learning models, including SVM, LightGBM, bi-LSTM, GRU, Transformer, 1D-CNN, and HDC.
3. Mobile Application Development: Developed a complete smartwatch-smartphone application system supporting local inference and real-time intervention.
4. Accuracy-Efficiency Balance: The HDC model achieved optimal balance between accuracy and efficiency, suitable for deployment on resource-constrained mobile devices.
5. Real-World Validation: Conducted actual device performance testing on Samsung Galaxy S20 using PyTorch Executorch.

Methodology Details

Task Definition

Input: Smartwatch sensor data (accelerometer x,y,z axes, gyroscope x,y,z axes, heart rate) Output: Binary classification result (intoxicated/sober), with threshold at TAC 35µg/L Constraints: Model must run in real-time on resource-constrained mobile devices

Data Collection and Preprocessing

1. Device Configuration:
   • Apple Watch Series 8: Collects accelerometer, gyroscope, heart rate, and geolocation data
   • BACtrack Skyn wristband: Measures TAC values every 30 minutes
   • BACtrack C8 breathalyzer: Measures BAC values
2. Data Preprocessing:
   • Original sampling frequency: 50Hz
   • Low-pass filtering applied after frequency domain analysis, downsampled to 40Hz
   • Data segmented into 20-second windows
   • 14 valid participants identified (TAC exceeding 35µg/L)

Model Architectures

1. Support Vector Machine (SVM)

• Uses radial basis function kernel
• Includes neural network feature extraction: flatten layer + 0.1 dropout + linear layer (128 dimensions) + ReLU activation

2. LightGBM

• Learning rate: 0.1
• Leaf nodes: 32, maximum depth: 4, number of estimators: 5
• Regularization parameters α and λ: 0.5

3. Bidirectional LSTM (bi-LSTM)

• 4 layers, 128 hidden units per layer
• Bidirectional architecture, dropout rate: 0.1
• Attention mechanism + fully connected layer

4. Gated Recurrent Unit (GRU)

• Single layer, 64 hidden units, dropout rate: 0.1
• Attention mechanism + fully connected layer + sigmoid activation

5. Transformer

• 2-layer encoder, embedding dimension: 128
• Positional encoding + dropout for overfitting prevention

6. One-Dimensional CNN (1D-CNN)

• 3 convolutional layers, 32 filters per layer
• Kernel sizes of 3, 5, and 7 respectively

7. Hyperdimensional Computing (HDC)

• Uses key-value encoding and RefineHD algorithm
• Hypervector dimension: 3000

Technical Innovations

1. Multi-Sensor Fusion: First to combine accelerometer, gyroscope, and heart rate data for intoxication detection
2. Long-Term Monitoring: Continuous three-week monitoring compared to short-term data in previous studies
3. Edge Computing Optimization: Specifically optimized for computational and power consumption constraints of mobile devices
4. Practical Validation: Verified model deployment feasibility on real Android devices

Experimental Setup

Dataset

• Participants: 30 participants (ages 21-55), with 14 valid participants after screening
• Monitoring Period: 3 weeks of continuous monitoring
• Data Volume: Multiple "sessions" per participant, each session containing continuous sensor data exceeding 1 minute

Experimental Design

1. User Clustering: K-means clustering to categorize users into high, medium, and low TAC level categories
2. Data Partitioning:
   • Test set: 1 user per category (users 16, 25, 26)
   • Training validation set: Remaining 11 users with 3-fold cross-validation

Evaluation Metrics

• ROC-AUC: Area under the receiver operating characteristic curve
• PR-AUC: Area under the precision-recall curve
• Accuracy: Overall accuracy, sober accuracy, intoxicated accuracy
• F1 Score: Harmonic mean of precision and recall

Hardware Performance Testing

• Device: Samsung Galaxy S20 (Snapdragon 865, 8GB RAM)
• Framework: PyTorch Executorch
• Testing: 100 inference iterations, measuring time, memory, and power consumption

Experimental Results

Main Results

Hardware Performance Results

Key Findings

1. Optimal Balanced Performance: HDC and 1D-CNN models demonstrated the best balanced performance, accurately identifying both sober and intoxicated states.
2. Efficiency Advantages: The 1D-CNN model maintains high accuracy while featuring an extremely small model size (0.035MB) and fast inference speed.
3. Deployment Feasibility: All tested models complete inference within the 20-second window, meeting real-time application requirements.

Related Work

Traditional Detection Methods

• Invasive Methods: Blood, urine, saliva testing, gas chromatography
• Breath Testing: Measuring ethanol volume ratio in alveolar air
• Transdermal Sensors: Using electrochemical or enzymatic principles to measure alcohol through skin

Machine Learning Approaches

• Physiological Signals: PPG and ECG measurements, facial temperature, bioimpedance
• Motion Sensors: Motion readings from smartphones and wearable devices
• Traditional ML: Applications of random forests, SVM, CNN and other methods on accelerometer data

Advantages of This Research

Compared to existing studies, this paper is the first to:

1. Combine multiple smartwatch sensors for long-term monitoring
2. Validate model performance in real-world environments
3. Consider computational and power consumption constraints of mobile devices

Conclusions and Discussion

Main Conclusions

1. Technical Feasibility: Smartwatch sensor data can effectively be used for intoxication detection, with HDC and 1D-CNN models achieving practical-level accuracy.
2. Practical Validation: Models can run in real-time on commercial smartphones, meeting JITAI application requirements.
3. Balanced Optimization: The HDC model achieved optimal balance between accuracy and efficiency.

Limitations

1. Sample Size Constraints: Only 14 participants with valid data, potentially affecting model generalization ability.
2. Individual Differences: Significant variations exist in alcohol tolerance and physiological responses among different users.
3. Environmental Factors: Various interference factors in real-world environments may affect detection accuracy.
4. Model Compatibility: Some models (bi-LSTM, GRU) could not be successfully exported to the Executorch framework.

Future Directions

1. Dataset Expansion: Increase diversity of user populations and extend monitoring periods.
2. Personalized Models: Develop personalized models considering user profiles and demographic variables.
3. Optimization Implementation: Improve efficiency using C language and vectorized operations.
4. Privacy Protection: Address data privacy and security concerns.
5. Cross-Domain Applications: Extend to other behavioral health challenges.

In-Depth Evaluation

Strengths

1. Strong Innovation: First application of multi-sensor smartwatch data for long-term intoxication monitoring with significant academic and practical value.
2. Rigorous Experimental Design: IRB-approved research design with strict participant screening and ethical review.
3. Comprehensive Evaluation: Systematic comparison of seven different types of machine learning models.
4. Practical Orientation: Focuses not only on accuracy but also on model deployment feasibility on mobile devices.
5. Complete System: Developed an end-to-end solution from data collection to application deployment.

Limitations

1. Data Scale Constraints: Limited number of valid participants (14), potentially affecting statistical significance and generalization ability of results.
2. Insufficient Feature Engineering: Relatively simple feature engineering on sensor data, potentially underutilizing data potential.
3. Lack of Comparative Benchmarks: Missing direct comparison with existing commercial intoxication detection methods.
4. Long-Term Stability Unverified: Model performance stability during long-term use has not been evaluated.

Impact

1. Academic Contribution: Provides new research directions for smartwatch applications in health monitoring.
2. Practical Value: Potential to develop into actual consumer products with positive public health impact.
3. Technology Advancement: Promotes development of edge AI and mobile health monitoring technologies.

Applicable Scenarios

1. Personal Health Management: Helps users monitor and control drinking habits.
2. Public Safety: Prevents dangerous behaviors such as drunk driving.
3. Medical Assistance: Supports alcohol dependence treatment and rehabilitation.
4. Enterprise Applications: Workplace safety monitoring.

References

The paper cites 38 relevant references covering important research outcomes in multiple fields including alcohol detection technology, machine learning methods, and mobile health applications.

Overall Assessment: This is a research paper with significant practical value that makes pioneering contributions to the field of smartwatch-based intoxication detection. Although there is room for improvement in data scale and model optimization, its complete system design and actual deployment validation establish a solid foundation for subsequent research in this field.