[Paper Review] Virtual Breathalyzer

Ben Nassi, Lior Rokach+1 more|arXiv (Cornell University)|Dec 14, 2016

Anomaly Detection Techniques and Applications24 references4 citations

TL;DR

This paper proposes a virtual breathalyzer system that detects alcohol intoxication using motion data from wearable devices (smartwatches, smartphones, Google Glass, fitness trackers) via supervised machine learning. By analyzing free gait patterns, the system achieves 0.95 AUC and 0.05 FPR at 100% TPR, offering a non-intrusive method to alert users or disable connected cars when drunk driving is detected.

ABSTRACT

Driving under the influence of alcohol is a widespread phenomenon in the US where it is considered a major cause of fatal accidents. In this research we present a novel approach and concept for detecting intoxication from motion differences obtained by the sensors of wearable devices. We formalize the problem of drunkenness detection as a supervised machine learning task, both as a binary classification problem (drunk or sober) and a regression problem (the breath alcohol content level). In order to test our approach, we collected data from 30 different subjects (patrons at three bars) using Google Glass and the LG G-watch, Microsoft Band, and Samsung Galaxy S4. We validated our results against an admissible breathalyzer used by the police. A system based on this concept, successfully detected intoxication and achieved the following results: 0.95 AUC and 0.05 FPR, given a fixed TPR of 1.0. Applications based on our system can be used to analyze the free gait of drinkers when they walk from the car to the bar and vice-versa, in order to alert people, or even a connected car and prevent people from driving under the influence of alcohol.

Motivation & Objective

• To develop a non-invasive, wearable-based method for detecting alcohol intoxication using motion data instead of breathalyzers.
• To formalize drunkenness detection as a supervised machine learning task using gait patterns from multiple wearable sensors.
• To evaluate the effectiveness of different sensor types and feature sets (e.g., histograms, frequency, statistics) in detecting intoxication.
• To validate the system against real police-grade breathalyzer measurements in real-world bar settings.
• To explore the feasibility of integrating this system into connected vehicles to prevent drunk driving.

Proposed method

• Collected gait motion data from 30 subjects using four wearable devices: Google Glass, LG G-Watch, Microsoft Band, and Samsung Galaxy S4.
• Formalized the detection task as both binary classification (drunk vs. sober) and regression (estimating breath alcohol concentration, BrAC).
• Used a police-grade breathalyzer as the ground truth for labeling intoxication levels.
• Extracted four feature sets: histograms of movement distribution, frequency domain features (FFT), known gait features, and statistical features.
• Trained and evaluated multiple classifiers (SVM, RF, GBC, LR), with Gradient Boosting Classifier (GBC) yielding the best performance.
• Conducted ablation studies to assess the contribution of each feature set to overall detection accuracy.

Experimental results

Research questions

• RQ1 Can motion data from common wearable devices reliably detect alcohol intoxication based on gait changes?
• RQ2 Which combination of wearable devices and sensor features yields the highest detection accuracy for intoxication?
• RQ3 How do different feature sets (histogram, frequency, statistics, gait features) contribute to the performance of the detection model?
• RQ4 Can a system based on free gait walking detect intoxication with sufficient accuracy to be used in real-world interventions?
• RQ5 Is the system robust enough to be deployed in connected cars to prevent ignition when the driver is intoxicated?

Key findings

• The system achieved an AUC of 0.95 for detecting intoxication at a BrAC threshold of 240 μg/L using data from four wearable devices.
• Using only a smartphone and smartwatch, the system still achieved an AUC of 0.94, demonstrating strong performance with minimal hardware.
• The histogram of movement distribution was the most important feature set, yielding the highest AUC (0.80) when used alone and the lowest AUC (0.42) when excluded.
• Frequency-based features contributed the least, with an AUC of 0.53 when used alone and a higher AUC (0.97) when excluded, suggesting limited predictive power.
• The system maintained 100% true positive rate (TPR) with only 5% false positive rate (FPR), indicating high sensitivity and low false alarm rate.
• The results suggest that detecting intoxication requires multi-part body movement data, as single-body sensors alone are insufficient for reliable detection.