EVIDENCE LIBRARY
Does 80% Wearable Accuracy Establish Alcohol-Impaired Driving?
A 28-person pilot classified alcohol-related labels from wrist signals. Its accuracy depended on the label, and neither classification was a direct test of driving.
No. An 80% classification accuracy does not establish that a wearable can determine whether a person was driving while impaired. In a 28-person police-academy pilot, the best reported model distinguished categories defined by a breath-derived alcohol threshold. Performance differed when the target was cadet-assessed sobriety-test failure. No driving outcome was measured.
The key DUI question is what the model was trained to recognize. A concentration threshold, a field-test judgment and impaired driving are different outcomes, even if a paper uses “impairment” for more than one of them.
What the wrist device measured
Kaczor and colleagues observed volunteers participating in police-academy alcohol training sessions. Participants drank their choice of alcoholic beverages at their own pace while wearing a research-grade wrist sensor. Twenty-eight completed the protocol and contributed analyzed data.
The device recorded motion, heart rate, skin temperature and electrodermal activity. It did not directly measure ethanol in blood, breath or sweat. Alcohol exposure was assessed separately with a breath instrument.
The investigators compared an initial sober period with a later period during the training session. Heart rate, skin temperature, electrodermal activity and a movement-variability measure increased significantly. These are physiologic or behavioral signals associated with the session; they are not specific chemical identification of ethanol.
Two labels produced different questions
One model label used the study’s alcohol-threshold rule. The paper reports breath-instrument results using g/dL, a blood-style unit, and applies a threshold of 0.08 g/dL. Those numbers should not be represented as directly measured blood concentrations or as mass per volume of breath. The reported units require clarification before any case-specific numerical conversion.
The second label used a majority judgment from cadet teams administering the study’s four-part sobriety assessment. Twenty-four participants met the alcohol-threshold label, while 19 met the cadet-assessment label. The categories did not identify identical sets of people.
Calling one label a reference standard for a machine-learning analysis means that the model is evaluated against that label. It does not independently prove that the label captures every aspect of impairment.
The headline accuracy omitted important differences
For the alcohol-threshold label, the integrated XGBoost model had reported accuracy of 80%, sensitivity of 78% and specificity of 81%. Sensitivity describes identification of positive labeled segments; specificity describes identification of negative labeled segments.
For the cadet-assessment label, the integrated model had approximately 74% accuracy, 52% sensitivity and 85% specificity. Its sensitivity therefore indicated substantial missed positive labels. A motion-only model had higher accuracy for that target, approximately 78%, with sensitivity around 64% and specificity around 86%.
There was no single “80% accurate impairment detector” that performed identically across both definitions. Nor can overall accuracy be turned into the probability that a particular positive result is correct; that interpretation depends on the population, prevalence and error pattern.
Many segments did not mean many independent participants
The analysis used 840 overlapping five-minute windows from 28 participants. The investigators used subject-wise cross-validation, keeping a participant’s windows together across training and validation folds to reduce leakage between them.
That is a useful safeguard, but the independent human sample remained 28. The study was a single-site feasibility analysis, without an external validation cohort or a test of performance in moving traffic. Overlapping windows and repeated observations should not be described as 840 independently recruited people.
Why the findings are not yet a DUI decision rule
Physical activity and social behavior were not standardized, there was no sober comparison group undergoing the same activities, and some participants used nicotine. Similar step counts across the two analyzed periods do not exclude every activity-related influence on the signals.
The authors also reported that the academy’s training breath instrument was not routinely calibrated. Cadets were inexperienced examiners, although supervised. Both issues affect confidence in the labels against which the models were evaluated.
The study did not validate a consumer smartwatch, establish a legal alcohol rule or show that a model could decide whether an individual was safe to drive. It supports further investigation of passive sensing, with clear definitions and external testing.
The site’s wrist reaction-time pilot analysis concerns a different sensor task. The laboratory sensitivity analysis explains the separate evidence needed to connect a detectable alcohol effect with prediction of driving.
Primary source
Kaczor EE, Golleru M, Carter DC, et al. Detecting Ethanol Intoxication and Impairment Using Wearable Biosensors. Proceedings of the 59th Hawaii International Conference on System Sciences. 2026:3532–3542. https://doi.org/10.24251/HICSS.2026.422. Full text.