TL;DR
This paper introduces "Virtual Breathalyzer," a novel system that uses motion sensors from everyday wearables like smartphones and smartwatches to detect alcohol intoxication. By analyzing subtle changes in gait, the system aims to provide a passive, non-invasive, and real-time method for intoxication detection, potentially preventing drunk driving.
What Problem Does This Paper Address?
Driving under the influence of alcohol (DUI) is a major cause of fatal accidents. Current methods for detecting intoxication, such as breathalyzers, are often inconvenient, require specialized equipment, or are not suitable for real-time, personal use. While some research has explored using technology for intoxication detection, many approaches are not passive, require user cooperation, or haven't been validated against established standards like police-grade breathalyzers. This paper addresses the need for a readily available, passive, and accurate method to detect intoxication, specifically focusing on gait analysis using ubiquitous wearable devices.
Methods
The core of the "Virtual Breathalyzer" approach is to formalize intoxication detection as a binary classification problem using supervised machine learning. The system leverages motion sensor data (accelerometer and gyroscope) from smartphones and wrist-worn devices.
The methodology involves:
- Data Collection: Participants walked in a controlled environment, providing gait samples both before and after consuming alcohol.
- Feature Extraction: A comprehensive set of features was extracted from the sensor data, including:
- Spectrum Domain Features: Derived from Fast Fourier Transforms (FFT) to analyze the distribution of signal power across different frequencies.
- Statistical Features: Standard statistical measures (e.g., mean, median, acceleration) to capture signal characteristics.
- Histogram Features: Representing the distribution of sensor values over time.
- Known Gait Features: Established features known to be indicative of gait patterns.
- Difference Calculation: The key innovation is calculating the difference between features extracted from gait samples taken when the individual was sober versus when they were suspected of being intoxicated. This difference highlights the physiological changes caused by alcohol.
- Machine Learning Models: Various classifiers were trained and evaluated, including Naive Bayes, Logistic Regression, Support Vector Machines, Gradient Boosting Machine (GBM), and AdaBoost.
- Leave-One-User-Out Cross-Validation: To ensure robust evaluation, the model was trained on data from 29 subjects and tested on the remaining one, repeated for all subjects.
Datasets
The study utilized a custom dataset collected from 30 participants across three bars. Each participant provided:
- Two gait samples (one before drinking, one after drinking) recorded using a smartphone (Samsung Galaxy S4) and a wrist-worn device (Microsoft Band).
- Two breath samples measured by a professional breathalyzer (Drager Alcotest 5510) to establish Blood Alcohol Content (BrAC) levels.
This dataset was labeled based on common BrAC thresholds (220, 240, and 380 µg/L) to classify subjects as either "sober" or "drunk." The paper does not introduce a publicly available benchmark dataset but rather a dataset specifically for this research.
Key Insights and Ideas
- Gait as a Biomarker for Intoxication: The fundamental idea that subtle changes in walking patterns (swaying, reduced coordination) due to alcohol can be reliably detected by motion sensors is a key insight.
- Difference-Based Detection: Calculating the difference between sober and intoxicated gait features is crucial. This approach accounts for individual baseline gait variations and focuses on the impact of alcohol.
- Ubiquitous Devices Suffice: The study demonstrates that readily available commercial devices (smartphones and smartwatches) are sufficient for this task, eliminating the need for specialized hardware.
- Combined Device Data is Superior: Using data from both a smartphone and a wrist-worn device significantly improves detection accuracy compared to using either device alone.
- Accelerometer is Highly Informative: Accelerometer data alone provides nearly the same performance as combining accelerometer and gyroscope data, suggesting that measuring changes in acceleration is particularly effective for detecting intoxication-related gait alterations.
Key Results
The "Virtual Breathalyzer" system achieved impressive results:
- High Accuracy: An Area Under the Curve (AUC) of 0.97 was achieved for detecting intoxication at a BrAC threshold of 240 µg/L using the Gradient Boosting classifier.
- Effective Classification: The AdaBoost and Gradient Boosting classifiers generally performed best, with AUCs ranging from 0.915 to 0.979 across different BrAC thresholds (220, 240).
- Low False Positive Rate (FPR) with High True Positive Rate (TPR): When prioritizing the detection of all intoxicated subjects (TPR=1.0), the FPR was managed, with the AdaBoost classifier achieving an FPR of 0.04 at a BrAC threshold of 240.
- Device Importance: Combining data from both smartphone and wrist-worn devices yielded the highest AUCs (e.g., 0.952 for Gradient Boosting at 240 BrAC), highlighting the benefit of multimodal sensing.
- Feature Importance: A combination of all feature types (histogram, statistical, known gait features, and FFT) generally outperformed models trained on individual feature sets, suggesting a holistic approach is best.
Limitations and Future Work
Limitations:
- Imbalanced Dataset at High BrAC: At the highest BrAC threshold (380 µg/L), the dataset was highly imbalanced (90% sober), leading to difficulties in accurate intoxication detection for that specific threshold.
- Generalizability: While validated against police breathalyzers, the study was conducted with 30 participants. Further validation with a more diverse population across different medical, geographic, and demographic backgrounds is needed for general deployment.
Future Work:
- Passive & Virtual Methods: Explore additional passive sensing methods (e.g., skin conductivity, reduced movement) and virtual methods to detect intoxication and drug use.
- Aggregated Data Insights: Investigate deriving insights from aggregated, low-resolution data to enable more generalized virtual intoxication detection.
- Data Quality Assessment: Further research is needed to understand the quality of data from commercial wearable sensors and how it compares to dedicated medical sensors.
- Dimensionality Reduction & Feature Analysis: Identify key features that contribute most to performance and explore dimensionality reduction techniques.
- Population Diversity: Analyze the impact of diverse medical, geographic, and demographic variations on the method's performance before widespread deployment.
Summary updated