EVIDENCE LIBRARY
Can a Wrist Reaction-Time Sensor Prove Alcohol Impairment? A 19-Person Pilot Study
A 19-person alcohol study linked wrist reaction times to breath readings. See why the sensor does not yet establish BAC, driving impairment, or causation.
Short answer: A wrist-worn sensor can measure the timing of a person’s response to a light, and a small controlled alcohol study found slower responses during higher breath-alcohol readings. That is not the same as proving a person’s blood-alcohol concentration, measuring all aspects of driving impairment, or establishing why a particular crash happened. In Blackwell and colleagues’ 2026 pilot experiment, alcohol-related response-time changes were detectable at the group level, but the primary pooled regression explained only 8.4% of observed variability. The study did not validate the wearable as a stand-alone forensic impairment test or as an ignition interlock.
The original study is Blackwell et al., “On-body measure of reaction time correlates with intoxication level,” PLOS ONE (2026), DOI 10.1371/journal.pone.0323858. The distinction between a measurable laboratory association and a defensible conclusion about an individual is the central forensic issue.
What did the wrist device actually measure?
The researchers evaluated a Pison Technology wrist wearable with surface electromyography (EMG) electrodes and an inertial sensor. Participants watched for a brief visual cue and opened a hand as quickly as possible. The EMG signal identified the beginning of neuromuscular activation. The motion sensor identified the beginning of movement. Neither sensor measured ethanol molecules.
The researchers distinguished premotor time (PMT)—the interval between visual cue and detected muscle activation—from reaction time (RT), the interval between cue and movement onset. The difference between those times represents an estimated electromechanical delay in this testing arrangement. A proprietary algorithm helped detect the signal onsets. These are timed responses to a specific light-and-hand-opening task; they are not field sobriety tests, road-driving observations, or chemical assays.
A participant’s result may be influenced by cue visibility, fatigue, attention, motivation, motor function, equipment contact, prior training, and the way the baseline was established. Such influences matter when someone tries to translate one device reading into a claim of alcohol-caused incapacity.
How was the alcohol experiment designed?
The paper describes 19 recruited adults aged 21–36: 14 assigned to consume alcohol and five nondrinking controls. Its participant section separately states that 18 completed testing, and its within-person threshold analysis reports 13 alcohol-exposed participants with relevant paired data. These figures concern different stages of study participation; they should not be collapsed into one apparently uniform sample size.
The alcohol group was dosed to approach a targeted breathalyzer-displayed BAC equivalent of approximately 0.10–0.12%. The protocol involved measurements before drinking and repeated blocks of testing after dosing during a roughly 90-minute session. At successive intervals, investigators used a BACTrack S80 breath instrument and the wrist sensor. Additional alcohol could be administered if the breath-instrument target was not reached at the designated time.
The controls did not drink, and the study was not a blinded alcohol-versus-placebo crossover. Several participants had different device-familiarization histories; some baseline recordings were obtained on a different day, including after the laboratory intervention. Those distinctions can affect a response-time comparison independently of ethanol.
Which measurements were direct, and which were inferred?
| Reported item | How it was obtained | What it cannot establish alone |
|---|---|---|
| Breathalyzer-displayed alcohol level | A BACTrack S80 breath test repeated during the laboratory session | Not a separately assayed venous whole-blood ethanol concentration or a complete forensic laboratory validation of this specific study’s testing |
| Premotor time | Interval from visual cue to wrist surface-EMG onset, with software detection | Not a direct ethanol measurement, diagnosis of intoxication, or comprehensive cognitive capacity assessment |
| Reaction time | Interval from cue to motion onset detected by the wearable’s inertial sensor | Not the same as steering, braking, hazard perception, or actual road performance |
| Change in reaction time (“Delta RT”) | Derived difference relative to baseline or comparison condition in the authors’ analysis | Not an independently validated individual alcohol estimator; baseline quality matters |
| Above-versus-below 0.08% category | Research grouping of breath-instrument alcohol readings around the paper’s threshold | Not an independent determination of the controlling legal offense, measured whole-blood BAC, or functional impairment in a case |
For additional analytical context, the site’s breath alcohol measurement guide addresses what a breath instrument actually samples. Likewise, interpreting a specimen measurement involves different questions from interpreting a drug’s pharmacologic effects.
What numerical changes did the authors report?
The paper grouped repeated reaction-time observations by its displayed alcohol categories: a control group, low (at or below 0.05%), medium (above 0.05% through 0.08%), and high (above 0.08%). Its reported mean Delta RT values, in the study’s timing scale, differed across groups. Do not confuse them with the total time required to perceive, decide, and brake in an approaching roadway hazard.
| Comparison or model | Study-reported result | Scientific interpretation |
|---|---|---|
| Control, low, medium, high categories | Mean Delta RT: 72.6, 67.5, 104.4, and 112.7, respectively | Higher categories were associated with larger study-defined response-time changes; the low category was not larger than control |
| Medium versus high category | Pairwise difference was not significant, p = 0.13 | The study did not establish that every additional increase in breath level produced a reliably separable reaction-time category |
| Threshold-group pooled contrast | Below-limit mean 73.5 versus at/above-limit mean 112.4; difference 38.88 study timing units | A statistical contrast among repeated observations, not a validated cutoff for classifying one new driver |
| Pooled simple regression | R² = 0.084; F(1, 3547) = 323.3, p < 0.001 | Reported alcohol-level variable explained about 8.4% of variation in Delta RT; most variation was unexplained by that model |
| Within-person comparison | Delta RT 73.21 ms below versus an increase of 39.16 ms at/above the study threshold; seven of 13 participants had individually significant differences | A limited within-person observation; not universal subject-level classification accuracy |
The authors also reported an association between premotor time and reaction time in nondrinking controls (r approximately 0.87). This is an association between two timing measures from the same task; it does not show that the device chemically identifies alcohol or differentiates ethanol from other causes of delay.
Does a very small p-value establish that this wearable identifies impaired drivers?
No. Statistical significance and accuracy for an individual decision are different properties. The publication reports many more recorded response observations than human participants: one major analysis has denominator degrees of freedom in the thousands, whereas only 19 people were recruited. Repeated measurements from a person are correlated. Treating each response as if it were wholly independent of the other responses by that person can make a p-value look more precise than a subject-level generalization warrants.
The authors themselves acknowledge that their analyses did not fully account for the repeated nature of the measurements or potential covariates, and recommend mixed-effects modeling in subsequent studies. Their discussion also notes small and uneven comparison groups, differences in participants’ experience with the device, inconsistent baseline environments, and possible false-positive findings from multiple testing. These limitations are substantive, not merely stylistic qualifications.
For a proposed forensic classifier, one would want a prospectively fixed decision rule, participant-level uncertainty estimates, sensitivity and specificity at predefined thresholds, false-positive and false-negative counts, external validation in new participants, repeatability, and evidence that the threshold performs under real-world conditions. The publication does not provide a validated courtroom or ignition-interlock classification rule meeting that description.
How meaningful is the reported R² of 0.084?
The authors’ pooled linear model associated breathalyzer-displayed alcohol level with the study-defined reaction-time change, but its reported R² of 0.084 means that only 8.4% of the observed variation in that response outcome was accounted for by the modeled alcohol-level variable in that analysis. It is not a probability of being impaired, a BAC measurement error, or an “8.4% accurate” test.
Direction of prediction matters. The simple linear regression reported in the article used the breathalyzer-displayed alcohol value to predict the reaction-time change. A consumer device attempting to estimate an unknown alcohol level from a measured reaction time would be solving the reverse problem. The significance and R² of the published forward regression are not, by themselves, a validation of an accurate inverse BAC estimator.
Neither the small R² nor the statistically significant group contrasts resolve whether a new person’s delayed response reflects alcohol, sleep loss, task unfamiliarity, a medical or neurologic condition, distraction, expectancy, or another influence. A person can have a slower-than-usual response without any ethanol exposure, and a comparatively normal result on this particular task would not exclude impairment in another domain.
Could the device distinguish alcohol effects from fatigue or other drugs?
Not on the evidence generated in this experiment. The subjects were evaluated principally after drinking alcohol versus nondrinking controls. The research did not include randomized controlled arms isolating sleep deprivation, sedating medications, cannabis, neurologic disease, trauma, or simultaneous drug exposure. A wrist timing sensor detects a response, not the pharmacologic identity of its cause.
The investigators suggest that reaction-time wearables might eventually inform fitness-to-drive decisions or become an ignition-interlock component, but those are proposed applications. The paper did not prospectively demonstrate that an installed interlock reliably prevented unsafe driving without erroneously excluding unimpaired people. The study’s limited laboratory task must not be described as real-world driver-impairment validation.
Is the breathalyzer reading the same as measured blood BAC?
No. The paper frequently uses the expression “BAC” for values obtained through the BACTrack S80 breathalyzer. Its described experiment did not measure contemporaneous venous whole-blood ethanol by a forensic laboratory method. A displayed blood-equivalent value supplied by a breath instrument must be identified as such. It should not silently be relabeled as a direct analysis of whole blood.
There are also separate issues of instrument calibration, sampling conditions, mouth alcohol, body temperature, timing relative to absorption, and the relation between breath and blood. A research report showing that two variables vary together does not validate a universal conversion for a different instrument and a different person. The site’s pharmacokinetics and forensic interpretation guide provides the broader framework for distinguishing measured concentrations from pharmacokinetic inferences.
Does the research measure actual driving or crash causation?
No road driving, steering emergency, hazardous intersection, divided-attention traffic task, or real collision was an outcome in this pilot. Participants responded to a small light by opening a hand in a laboratory setting. That task can supply data about one component of psychomotor performance, but it does not capture visual scanning, hazard prediction, lane maintenance, judgment, or performance across the full circumstances of a driving incident.
Alcohol can affect several of those functions, but proving a general pharmacologic possibility is not the same thing as identifying the mechanism of a specific collision. In a case, scene reconstruction, witness observations, contemporaneous specimen records, controlled analytical testing, the driver’s actions, medical factors, and the applicable burden of proof must be considered separately. Neither a slow sensor response nor a statistically significant group mean establishes that alcohol caused a collision.
What would an expert examine before relying on a wrist-reaction result?
A competent examination would seek the device make and software version, raw EMG and accelerometer streams, clock synchronization, onset-detection algorithm, predeclared exclusions, baseline date and environment, wear position and contact, practice trials, test timing, operating conditions, instrument calibration and service records, study protocol, subject-level data, and the model used to interpret the response.
The expert should determine whether the proposed inference is limited to the measured latency on this task, extends to an association with a breath-instrument display in a small experimental group, or purports to be a reliable individual impairment determination. The evidentiary foundation and uncertainty requirements become progressively more demanding across those three propositions.
Consistency of testing matters as much as apparent precision. A sensor providing response times with many decimal places is not necessarily accurate as an estimator of BAC or safe driving. Methods, bias, sampling, data dependency, and external validity are all relevant to admissibility and weight.
What conclusion is scientifically supportable?
Blackwell et al. provide preliminary evidence that a wrist-worn EMG-plus-motion system can detect laboratory response-timing changes associated with increasing breathalyzer-displayed alcohol levels. Its larger response-time changes occurred in the medium and high reported alcohol categories, and several individuals changed when they crossed the study threshold. However, the sample was small, measurements were repeated within people, and the principal pooled model explained only 8.4% of observed timing variance. The researchers expressly identify the need for larger, more rigorous validation.
The defensible conclusion is “promising laboratory association requiring validation,” not “validated test that proves a driver’s BAC, alcohol impairment, or crash causation.” For a broader directory of specimen, analytical and interpretive methods, see the forensic toxicology evidence library.
Frequently asked questions
Can a smartwatch tell exactly what someone’s BAC was?
No. A wrist EMG-and-motion reaction-time sensor measures response timing. The study compared those measurements with breathalyzer-displayed alcohol values; it did not validate a technique for determining an individual’s exact blood ethanol concentration.
Did the study show slower responses after drinking alcohol?
Yes, for selected reported group and within-person comparisons. Those findings are preliminary associations from a small experiment, with statistical and baseline limitations that must accompany any interpretation.
Were there 3,500 independent drivers in the analysis?
No. The research recruited 19 people. The thousands in some statistical degrees of freedom refer to repeated observations, not thousands of independent participants.
Does an R² of 0.084 mean the sensor was 8.4% accurate?
No. It means the study’s pooled regression explained approximately 8.4% of variation in the modeled response-time outcome. Classification accuracy, specificity, sensitivity and individual prediction errors are different measures.
Did crossing the paper’s 0.08% threshold establish a statutory DUI violation?
No. The threshold was used to organize research breath-instrument readings. Whether a particular statute is satisfied depends on jurisdiction, admissible specimens and evidence, timing, and other legally relevant facts.
Can a normal result prove someone was fit to drive?
No. A normal performance on one hand-opening reaction-time task does not rule out deficits in divided attention, hazard recognition, motor coordination, judgment, or other safety-critical driving abilities.
Is this wearable validated as a forensic field test or ignition interlock?
Not by this pilot. Prospective subject-level accuracy, error rates, independent replication, real-world conditions, and fit-for-purpose validation would be required before claiming reliable individual classification for those applications.
Primary scientific source: Blackwell MH, Valk HL, Eaton J, et al. (2026). On-body measure of reaction time correlates with intoxication level. PLOS ONE 21(4):e0323858. https://doi.org/10.1371/journal.pone.0323858.