Smartphones as Breathalyzers: AI-Based Detection of Alcohol Impairment from Smartphone Motion Sensors

Document Type

Presentation

Date of Original Version

3-27-2026

Abstract

Alcohol impairment significantly degrades motor coordination, balance, and postural control, resulting in measurable alterations in human gait dynamics. These impairments contribute substantially to roadway accidents and injury-related fatalities worldwide with one person dying every 34 minutes in the US each year. Conventional detection approaches, such as breathalyzer tests, require active participation and are typically administered only after impairment is suspected. In contrast, modern smartphones are equipped with inertial measurement sensors capable of continuously capturing motion signals during routine activities. These sensors provide an opportunity to analyze gait patterns and investigate whether behavioral indicators of intoxication can be detected directly from smartphone motion data. This research investigates the feasibility of detecting alcohol impairment using machine learning models trained on motion signals collected from smartphone sensors. Specifically, the study analyzes multivariate time-series data recorded from tri-axial accelerometers and gyroscopes embedded in smartphones during controlled walking tasks. These sensors capture linear acceleration and angular velocity signals along three spatial axes, producing high-frequency measurements that characterize the dynamics of human locomotion. The raw sensor signals are segmented into fixed-length temporal windows representing short intervals of gait activity. Each segment encodes temporal patterns associated with stride regularity, stability, and movement variability that may be affected by alcohol-induced motor impairment. The primary objective of this work is to develop computational models capable of identifying intoxication-related signatures in these motion signals. Time-series representations derived from accelerometer and gyroscope measurements are analyzed to extract discriminative features that capture subtle variations in gait behavior. Machine learning models are then trained to differentiate between sober and alcohol-influenced walking patterns using these representations. Particular attention is given to addressing challenges common to wearable and mobile sensing applications, including sensor noise, inter-subject variability in gait patterns, and limited availability of labeled intoxication data. To improve model robustness and generalization, ongoing work explores representation learning strategies designed to leverage large volumes of unlabeled motion data collected during natural walking activities. These approaches aim to learn latent embeddings that capture fundamental characteristics of human gait dynamics while remaining sensitive to behavioral perturbations associated with intoxication. Such representations may enable improved performance when labeled training data is limited, while also facilitating transfer across subjects and sensing environments.

This document is currently not available here.

Share

COinS