Detecting Ethanol Intoxication and Impairment Using Wearable Biosensors

Abstract

Reliable objective measures of a person’s intoxication and impairment from alcohol consumption are not readily available to the public. Wearable biosensors have the potential to provide a ubiquitous on-demand tool to deliver this kind of objective assessment in real world settings. This study evaluated the feasibility of assessing ethanol intoxication in N=28 healthy participants in a police academy’s intoxication lab using wrist-worn biosensors to continuously measure heart rate, skin temperature, electrodermal activity, and accelerometry. Participants consumed ad hoc standard alcoholic drinks in a controlled setting and had regular breath alcohol content assessments and underwent standard field sobriety testing. The analysis showed statistically significant changes in each physiologic parameter between the sober and intoxicated periods. An XGBoost model was applied to this data producing machine learning algorithms to identify impairment with an accuracy as high as 0.80. These results demonstrate that it is feasible to assess ethanol intoxication using wrist-worn biosensors.

1. Introduction

Humans have been making and consuming alcohol for thousands of years (Phillips, 2014), and today ethanol is the most widely abused drug worldwide. During that time, society has become familiar with the negative consequences of over-indulgence leading to intoxication and impairment (Yip, 2019). Despite this, research has shown that both social drinkers, as well as those with DUI/DWI (driving under the influence/driving while intoxicated) charges, are routinely unable to accurately gauge their own levels of intoxication (Fillmore & Van Dyke, 2020; Köchling et al., 2021). Approximately 32% of all vehicle-related deaths in the United States involve drivers intoxicated with ethanol (Dunne & Katz, 2015; Jones et al., 2020), and intoxicated individuals are more likely to make decisions while intoxicated that they later regret, with several studies reporting roughly 2/3 of individuals reporting low-risk regrettable behaviors (Dunne & Katz, 2015). Widespread alcohol use has significant impacts to individual health, population health, healthcare costs, and has an overall economic impact, despite a multitude of local and governmental efforts (DiMaggio et al., 2021; U.S. Department of Transportation, National Highway Traffic Safety Administration [NHTSA], 2025; Yip, 2019).

Although ethanol intoxication may be suspected clinically, it remains difficult to objectively and accurately detect without directly testing breath or blood. Law enforcement routinely uses breath alcohol content (BrAC) measured via alcohol meters (commonly referred to as “breathalyzers”) to determine if, and to what degree, an individual is intoxicated with ethanol by directly measuring exhaled ethanol. These results are typically reported in milligrams of alcohol per liter of breathed air. While this testing modality can be performed anywhere with rapid results, and can be easily purchased by individuals online, the devices are not widely used by the general public or readily on hand in most homes or social gatherings. In addition, even if these devices were present, an individual would need to remember to use it while under the influence of alcohol. Physicians and other healthcare professionals commonly rely on blood alcohol concentration (BAC) to determine if, and to what degree, an individual is intoxicated with ethanol. While this test is also accurate and reliable, it can only be performed by specialized personnel who can draw and then analyze blood samples. This makes this testing modality even less practical for the general public than the measurements of BrAC.

To address this problem, there has been a growing effort to create devices that can passively detect impairment to both study and help treat those with alcohol use disorder (AUD) (Davis-Martin et al., 2021; Paprocki et al., 2022). Unlike the episodic detections methods just mentioned, passive detection has the advantage of continuously detecting real-time changes in physiologic data as long as individuals are wearing the monitoring device. Such devices include a variety of smartphone applications, wrist watches, and one shoe based device that have used accelerometry data, transdermal BAC, and heart rate data. Of the wearable devices, measurement of a person’s BAC is typically the primary means used to detect alcohol use and to determine intoxication. A key drawback of the reliance on BAC is that humans can develop a significant degree of tolerance to the effects of alcohol such that BAC measurements may not reflect the degree to which a person is clinically impaired (Martin et al., 2013; Yip, 2019).

Wearable biosensors that measure physiologic and behavioral changes, as opposed to the BAC, can likely be used to obtain a more objective measure of an individual’s state of intoxication and impairment due to alcohol consumption. These small, noninvasive devices, similar in design and functionality to commercially available smart wristwatches or fitness tracking devices that are already readily accepted by people for daily use, have already been used to evaluate the physiologic response to other intoxicating substances in the past (Carreiro et al., 2014, 2015, 2016; Chintha et al., 2018). Research grade versions of such devices measure surrogate markers of the sympathetic nervous system (SNS) activity such as electrodermal activity (skin conductance), skin temperature, and heart rate, in addition to blood oxygen content and body motion in three axes. By measuring each of these variables up to 60 times per second, biosensors can create massive data sets that capture rapid fluctuations in physiology. Alcohol use produces changes in SNS output, coordination, and behavior that may be detectable by these wrist-worn biosensors. In one study, a paired wristwatch and smartphone system was used to measure changes in accelerometry and gyroscopic data as participants wore goggles that simulated various levels of BAC (McAfee et al., 2017). Using a machine learning algorithm, the authors of this study were able to correlate physiologic changes with the different levels of BAC goggles reporting up to 89.45% accuracy. While encouraging, this study uses only simulated inebriation and only two of the many biometric features available in modern smart watches. In addition, this result was obtained with a 99% and 1% train-test split. Wearable biosensors, therefore, represent a powerful new method for assessing physiologic changes in participants who imbibe alcohol.

The primary aim of this study is to determine whether it is feasible to use wrist-worn biosensors to identify signs of alcohol intoxication through changes in the biometric data. The secondary aim is to determine whether this data can be used to create machine learning algorithms that can identify the presence of alcohol intoxication and impairment.

2. Physiology of Alcohol Intoxication

The physical manifestations of alcohol intoxication consist of a spectrum of symptoms that become more obvious and severe as an individual’s BAC increases (Mirijello et al., 2023). However, the BAC/BrAC necessary to produce these symptoms is different between individuals due to differences in alcohol tolerance. While someone who is alcohol naïve will experience gross intoxication with obvious signs of impairment at a BrAC of 0.20 g/dl - 0.25 g/dl, someone with significant AUD can develop significant tolerance such that they can hold a normal conversation and appear only minimally impaired at a similar BrAC. So, while you are considered impaired with a BrAC above 0.08 g/dl from a legal perspective in the United States, the degree of intoxication and impairment at a BrAC of 0.08 g/dl will vary from person to person. In addition, while some may require a BrAC higher than 0.08 g/dl to achieve impairment, others with lower tolerance may reach appreciable impairment well below this threshold (Martin et al., 2013). This makes biochemical testing like BAC/BrAC relatively imperfect gauges of impairment and should be supplemented by a physical assessment for signs and symptoms of intoxication.

According to the Diagnostic and Statistical Manual of Mental Disorders 5th edition, the symptoms used to diagnose acute alcohol intoxication include problematic behavioral changes (i.e. aggression or inappropriate sexual activity), psychological changes (i.e. mood lability or impaired judgement), slurred speech, lack of coordination, unsteady gait, nystagmus, and stupor/coma (American Psychiatric Association, 2013). The standard field sobriety tests (SFSTs) administered by law enforcement assesses for these signs of motor dysfunction when making a determination of impairment. In addition to changes in movement and behavior, alcohol intoxication has also been shown to cause an increase in heart rate (Brunner et al., 2017; Mirijello et al., 2023), decrease in body temperature (Morris, 2024; Paprocki, 2022), and increase in perspiration or electrodermal activity (Morris et al., 2024; Paprocki et al., 2022). All of these changes can be continuously detected and measured using wrist-worn biosensors.

3. Methods

3.1. Study Protocol

This observational pilot study was approved by the University at Buffalo Institutional Review Board (IRB). Study participants were recruited from a pool of volunteers at the Erie County Police Academy’s biannual sobriety field test training for cadets, also known as the “wet lab”. These volunteers were recruited by the cadets and instructors from the local community. Participants in the six sessions of the sobriety field test training had already volunteered to consume alcoholic beverages with the goal of reaching a Breath Alcohol Concentration (BrAC) of between 0.1g/dL and 0.2g/dL and complete a series of sobriety assessments administered by cadets. The inclusion criteria for this study were: ≥ 21 years of age, physical ability to wear the biosensor on a wrist, and being sober at the time of enrollment. All participants provided written informed consent prior to participation and prior to ingesting any alcohol.

3.2. Hardware

The wearable biosensor worn was the Empatica Embrace Plus (Empatica, Boston, MA), an FDA-approved research grade biosensor that continuously detects and records heart rate, skin temperature, electrodermal activity (EDA), and accelerometry in three axes. Data collected with the sensor was stored in the internal memory and continuously uploaded to the Empatica Care Lab Application, which was installed on a set of linked iPads (Apple, Cupertino, CA), via encrypted Bluetooth connection. The Empatica Care Lab application uploaded that data to a HIPAA compliant cloud-based server that researchers could access on the Empatica Care website with a ~ 7-minute delay. All other data collected during the study period was stored on the University at Buffalo REDCap.

3.3. Data Collection

Upon arrival at the academy, volunteers for the sobriety field test training were approached by research staff who explained the study in detail and obtained written consent for those who agreed to participate. The participants were assigned a unique study ID number, asked to complete a demographic survey, and outfitted with an Empatica biosensor on their non-dominant wrist to wear throughout the 6-hour training session. After the biosensor was placed, study participants underwent ~ 20-minutes of police academy orientation, providing baseline biometric data, before engaging in the standard sobriety field test training activities under the direction of the police academy instructors, with no interventions from research staff. Academy instructors administered an initial BrAC test with a Preliminary Breath Test device, the Alco-Sensor FST (Intoximeters Inc, St. Louis, MO) to demonstrate all participants started the exercise sober with a BrAC of 0.0g/dL. After the first BrAC test, instructors began serving participants alcoholic beverages, and participants were offered a variety of commercially available products, including beer, hard seltzer, hard cider, wine, liquor with mixers, and liquor shots. Subjects were not limited to a single type of alcoholic beverage and were allowed to drink as much as they wanted at their own pace. Instructors would monitor the changes in BrAC every 1–2 hours and would encourage participants to increase or decrease the rate of drink consumption to ensure participants achieved the targeted BrAC concentration range of 0.1 g/dL - 0.2 g/dL.

Participants were allowed to converse, play games, and move around as desired. Research staff recorded the times when participants consumed each alcoholic beverage, the type of alcohol in each beverage, the time and duration that they engaged in physical activity, and the room temperature and humidity throughout the session. One hour before the end of the session, instructors stopped serving alcohol and the cadets began assessing the sobriety of each participant in 4–5 rotating teams of 3–4 cadets each. The details of the sobriety assessment procedures were intentionally withheld from the participants to better simulate a real-world encounter with an intoxicated subject for the cadets. The sobriety assessment consisted of 4 tests: an 18-step straight line walk and turn (W&T), a one leg stand (OLS), a horizontal gaze nystagmus (HGN), and a vertical gaze nystagmus (VGN). During each test, academy cadets assigned participants points for errors made that would suggest impairment. A participant would “fail” a test if they scored two points or more. If a participant failed two or more of these tests, then they were considered impaired by that team of cadets. If a majority of teams scored a participant as impaired, then they were labeled as impaired for our study. At the end of the standard field sobriety tests (SFST), each study participant had a final BrAC assessment and the Empatica wearable devices were returned to end the session.

3.4. Data Analysis

3.4.1. Signal processing and filtering

An overview of the signal processing and modeling steps is provided in Figure 1. Raw accelerometer readings were smoothed with a fourth-order Butterworth low-pass filter at a 10 Hz cutoff (sampling rate = 64 Hz), effectively attenuating vibration and electronic noise without distorting meaningful human motion. EDA was decomposed into tonic (< 0.05 Hz) and phasic (0.05–1 Hz) components using separate fourth-order Butterworth low-pass filters. Heart rate (HR), and skin-temperature streams retained their default units.

Figure 1.

End-to-End Signal-Processing and Machine-Learning Pipeline for Impairment Detection using Wearable Sensor Data

3.4.2. Segmentation, Sliding windows and labeling

Forty-minute physiological signal blocks, comprising the first 20-minutes of data collection for each session when participants are sober and the final 20-minutes immediately after the BrAC measurement taken at the beginning of the SFSTs were segmented from the raw recordings. Each 40-minute block was then partitioned with a 5-minute sliding window that advanced in 1-minute steps, yielding 30 valid windows per block. Once gaps and edge effects were excluded, there remained a total of 840 candidate windows across 28 participants. Each window carried two impairment labels depending on the BrAC and SFST results. The BrAC label was set as impaired when the breath-alcohol reading at the end of the block was at least 0.08 g/dL; otherwise, it was not impaired. The SFST label was based on the outcome of the standardized field sobriety test performed right after data collection, again marked as impaired or not impaired.

3.4.3. Feature Extraction

Time and frequency domain statistics were extracted for each axis (mean, standard deviation (SD), min, max, range, median, skew, kurtosis, root mean square (RMS), entropy, zero crossing rate, jerk metrics, dominant frequency, total power). In addition, features were derived for the vector magnitude, heart rate (mean, SD), electrodermal activity (tonic, phasic, filtered-phasic, peak count, mean normalized conductance), skin temperature (mean, SD), and demographics (BMI, sex), producing a total of 66 features per window (Accelerometer = 16 × 3 = 48, Accelerometer Magnitude = 3, EDA = 9, HR = 2, Temperature = 2, Demographics = 2). For each of the physiological and motion features that could be directly output from the sensor, a one-way repeated-measures analysis of variance (ANOVA) was conducted to compare the first and final twenty-minute segments.

3.4.4. Machine learning classification, training, validation, and feature evaluation

A single binary outcome (Impaired vs. Not Impaired) was modeled using four types of classifiers: Random Forest, SVM, XGBoost, and Gradient Boosting. After cleaning the data and removing incomplete windows, 840 5-minute segments remained. Each segment received two impairment labels: one based on breath-alcohol concentration (impaired if BrAC ≥ 0.08g/dL, otherwise not impaired) and one based on the majority cadet judgment of the SFST (impaired or not impaired). Machine learning classifiers were then trained on each labeling scheme using the same feature sets. Each classifier was embedded in a pipeline with z-score scaling and assessed via subject-wise five-fold GroupKFold cross-validation to avoid data leakage between windows from the same participant. Model performance was quantified using sensitivity, specificity, overall accuracy, impaired class F1 score, and the area under the ROC curve (AUC). Tree-based models additionally furnished feature-importance rankings and learning-curve diagnostics were generated to monitor over-fitting. Across all models, XGBoost performed best and obtained more balanced results for sensitivity and specificity, so all results reported below are for the XGBoost model performance.

Three feature configurations were evaluated: 1) Accelerometer-only (axis-level time- and frequency-domain metrics plus vector magnitude, 51 features), 2) Physiological (heart rate mean/SD, EDA tonic/phasic measures including peak counts and normalized conductance, skin-temperature statistics, BMI, sex, 15 features), and 3) Integrated feature set (both sets combined, 66 features). Accelerometer-derived features alone have reliably distinguished drug-induced physiological changes in prior work (Carreiro et al., 2016; Chintha et al., 2018), while adding cardiovascular and electrodermal signals has been shown to boost detection accuracy (Carreiro et al., 2020). The comparison across feature sets therefore quantifies both the standalone predictive power of locomotion and the incremental gains from integrating autonomic, thermal, and electrodermal information.

4. Results

4.1. Sample demographics

Twenty-nine participants were enrolled in the study. One participant was removed by academy instructors due to excessive emesis from intoxication and did not complete the full protocol, so was excluded from the analysis. Data from the remaining 28 (96.6%) participants were analyzed. Participants’ ages ranged from 21 to 64 years (mean 30.2 ± 10.6), with 43 percent being female. Participants self-reported their height and weight, resulting in a mean BMI of about 26 kg/m². Five participants reported daily prescription medications for depression and high cholesterol, two used occasional reflux treatments, and one reported only daily vitamin use. In the 24-hours before testing, roughly one-third used nicotine products and three reported recent cannabis or off-session alcohol use. Alcohol-use histories showed a median age of first alcohol use of 18 years, with half the cohort drinking at least weekly. Most participants typically consumed two to four standard drinks per occasion; beer was the preferred beverage (Table 1).

Table 1.

Participant characteristics (N=28)

Characteristic	Count (%)
Sample size	28
Age (mean ± SD, range)	30 ± 11 y (21–64)
Sex	Female: 12 (43%)
	Male: 16 (57%)
Race/Ethnicity	African American: 1 (3.6%)
	Caucasian (Hispanic/Latinx): 3(10.7%)
	Caucasian (non-Hispanic): 24 (85.7%)
BMI (mean ± SD)	26 ± 5 kg/m2
Age at first alcohol use (median, range)	18 y (14–21)
Drinking frequency	≥ once weekly: 14 (50%)
	Monthly or less: 14 (50%)
Typical drinks per occasion	2–4 drinks: 20 (71%)
	5–6 drinks: 5 (18%)
	> 8 drinks: 2 (7%)
Preferred beverage	Beer: 16 (57%)
	Liquor: 7 (25%)
	Wine: 3 (11%)
	Mixed/other: 2 (7%)

4.2. Participant impairment

Table 2 summarizes the BrAC achieved by participants measured just before the SFSTs began, with five views of the data. The “All” column gives the distribution for the entire cohort of 28 participants. The impaired and not impaired by BrAC columns apply the statutory cut-off of 0.08 g/dL, dividing the cohort into impaired and not-impaired counts and reporting their aggregate BrAC statistics. The impaired and not impaired SFST columns make the same division using the majority-cadet field-sobriety judgment. Each entry provides the number of participants in each category together with the group mean, median, minimum, maximum, and standard deviation values.

Table 2.

Impairment Data Summary (N = 28)

	All	Impaired by BrAC	Not Impaired by BrAC	Impaired by SFST	Not Impaired by SFST
# of Subjects (n)	28	24	4	19	9
Mean BrAC (g/dL)	0.12	0.13	0.07	0.13	0.10
Median BrAC (g/dL)	0.13	0.13	0.07	0.13	0.09
Min BrAC (g/dL)	0.06	0.08	0.06	0.06	0.07
Max BrAC (g/dL)	0.20	0.20	0.07	0.20	0.15
BrAC SD (g/dL)	0.04	0.03	0.01	0.03	0.03

On the BrAC-based rule, 24 of 28 subjects (86 %) met or exceeded the 0.08 g/dL threshold and were judged impaired. Under the SFST rule, only 19 participants (68 %) failed the sobriety assessments and were judged impaired. While the mean BrAC for the impaired by BrAC and impaired by SFST groups were almost the same, the mean BrAC for the not impaired by BrAC group was notably lower, reflecting occasional discordance between the biochemical measurements and cadet assessment. To further evaluate this discordance, the accuracy, sensitivity, and specificity for the SFST’s ability to detect a BrAC of ≥ 0.08 g/dl was calculated using the confusion matrix in Figure 2. The accuracy was found to be 0.75, the sensitivity 0.75, and the specificity 0.75.

Figure 2.

Confusion matrix evaluating the effectiveness of the SFST to detect a BrAC 0.08 g/dl

4.3. Changes in physiological and motion features

Table 3 reports a Repeated Measures ANOVA analysis comparing the changes in the recorded biometric features between the initial 20-minute sober period and the final 20-minute intoxicated period. The statistical comparisons showed that the mean heart rate, EDA, skin temperature and accelerometer variability all rose significantly between the first and final 20-minute segments across the full cohort (all p < 0.05). Subjects participated in a variety of games and non-uniform activities that were a potential source of confounding for this data set. Step counts were calculated from the accelerometry data and used as a surrogate marker for this non-uniform physical activity. The difference in step counts between the periods used for this analysis was not statistically significant.

Table 3.

Comparison of Biometric Features between the Sober and Intoxicated Periods (N = 28)

Feature	Start Mean (SD)	End Mean (SD)	p-value
Heart rate (bpm)	82.54 (10.47)	103.24 (18.50)	<0.001*
EDA (μS)	0.34 (0.58)	1.44 (2.23)	0.0071*
Skin temperature (°C)	29.93 (1.78)	31.59 (1.60)	<0.001*
Accelerometer STD (g)	0.044 (0.015)	0.057 (0.021)	0.0064*
Step counts	10.64 (5.23)	10.97 (4.62)	0.80

^*p < 0.05

4.4. Window counts and class balance

Data-processing yielded 840 analyzable 5-minute windows (30 per participant). When the BrAC criterion was applied, 360 windows were classified as impaired and 480 as not-impaired (positive-class prevalence = 43 %). Applying the SFST criterion to the same windows produced 285 impaired and 555 not-impaired segments (positive-class prevalence = 34 %). These class balances form the baseline for the performance metrics reported in the following subsections.

4.5. Model-performance comparison

As summarised in Table 4 (BrAC labels) and Table 5 (SFST labels), across both impairment definitions, a common pattern emerged: accelerometer features alone delivered consistent discrimination, physiological features alone lagged, and combining the two offered the most balanced performance. With BrAC as ground truth, the integrated model edged out the accelerometer baseline on every metric, whereas the physiological-only model showed a notable drop in accuracy and sensitivity. The findings were more pronounced for the SFST task, where motion cues again carried most of the signal, physiological cues missed many impaired cases, and the multimodal model recovered much of that lost sensitivity while preserving high specificity. These results confirm that wrist-motion statistics are highly informative for alcohol impairment, but that adding heart-rate, electrodermal, temperature, and demographic information provides a small yet consistent boost, particularly when impairment is judged by field sobriety rather than breath alcohol.

Table 4.

XGBoost Performance for BrAC-Based Impairment Detection Across Feature Configurations

Combination	Accuracy	Sensitivity	Specificity
ACC	0.79	0.77	0.81
Non-ACC	0.67	0.60	0.72
ACC + non-ACC	0.80	0.78	0.81

Table 5.

XGBoost Performance for SFST-Based Impairment Detection Across Feature Configurations

Combination	Accuracy	Sensitivity	Specificity
ACC	0.78	0.642	0.856
Non-ACC	0.67	0.354	0.836
ACC + non-ACC	0.74	0.519	0.847

4.6. BrAC Impairment Classification Performance

Using only axis-level and vector-magnitude descriptors (51 features), the model achieved 79% accuracy, 77% sensitivity, 81% specificity, and an AUC of 0.85. Training accuracy was 88%, validation accuracy 79%, and the F₁ score for the impaired class was 0.76. Inspection of feature-importance rankings highlighted RMS and spectral-power metrics on the X and Z axes as top predictors. Restricting the features to heart rate, EDA, temperature, BMI, and sex (15 features), resulted in a drop in performance to 67% accuracy with 60% sensitivity and 72% specificity. The impaired class F₁ was 0.61, with training and validation accuracies at 88% and 67%, respectively. Mean skin temperature, mean EDA, tonic EDA, and heart-rate variability emerged as the strongest non-locomotor indicators when feature importance was inspected.

Fusing all 66 features yielded the best balance for the model performance, with 80% accuracy, 78% sensitivity, and 81% specificity Figure 3. Training accuracy was 91%, while validation accuracy was 79%. This integrated model’s top predictors included accelerometry with the RMS for the X- and Z- axes and standard deviation of the heart rate, EDA peak count and variability, as seen in Figure 4.

Figure 3.

Confusion matrix of the performance for the model including all features

Figure 4.

Top 10 features importance bar plot for integrated feature set analysis

4.7. SFST Impairment Classification Performance

When the SFST was used as the outcome measure for classification, using only axis-level and vector magnitude descriptors resulted in an accuracy of 78%. X-axis RMS and Z-axis median were the dominant features, followed by Z-axis mean, minimum, and range. Restricting the feature set to those not from the accelerometer led to a substantial drop in performance, most notable by the 35.4% sensitivity and 67% accuracy. EDA mean, tonic EDA, standard deviation of heart rate, and BMI emerged as the strongest non-locomotor features.

With the combined feature set, classification of the SFST outcome had 74% accuracy, with a high 85% specificity, but only 52% sensitivity (Figure 5). The integrated model’s top predictors spanned accelerometry RMS on the X-axis and median on the Z-axis, together with tonic EDA and EDA variability (Figure 6).

Figure 5.

Confusion matrix of the performance for the model including all features

Figure 6.

Top 10 feature importance barplot for Integrated Feature Set Analysis

5. Discussion

5.1. Feasibility

The results of this study demonstrate that wearable biosensors can identify signs of alcohol intoxication. When comparing the biometric data of the participants when sober (BrAC of 0.0g/dL) to the biometric data of the participants at the end of the session (mean BrAC of about 0.12g/dL, Table 2) there were statistically significant changes in mean heart rate, mean EDA, mean skin temperature, and mean standard deviation of acceleration. From this biometric data, successful machine learning models were generated to identify impairment due to alcohol intoxication using two different definitions to classify impairment: a BrAC measurement of ≥ 0.08g/dL or a determination of impairment via the SFST. While both models performed best using an XGBoost algorithm, the BrAC model performed best with a combination of accelerometer and physiologic features while the SFST model performed best using accelerometer features only. The overall best performing model used the BrAC impairment classifier and achieved an accuracy of 0.80, a sensitivity of 0.78, a specificity of 0.81, and area under the curve of 0.84. This model also had better sensitivity, specificity, and accuracy for detecting a BrAC of ≥ 0.08 g/dl than the SFSTs performed by the cadets in this study.

While this study demonstrates the feasibility, and potential superiority, of using wrist-worn sensors for the detection of alcohol intoxication and impairment, the Empatica devices employed in this study are mainly used for research purposes and not readily accessible by the general public. However, the biometric features measured in this work can also be measured on more common smart watch brand devices such as the Apple Watch© or Garmin©. Such devices have been shown to reliably capture physiological data even in uncontrolled, real-world settings (Wolf et al., 2024). If the findings of this study can be replicated using such ubiquitous and widely accepted technology, wrist-worn smart devices could be used as the platform for a readily available machine learning algorithm-based system that could detect alcohol intoxication and impairment in real-world settings.

5.2. Changes in physiology

Our subject population exhibited statistically significant c hanges i n a ll f our b iometric features between the sober and intoxicated periods. There was a mean heart rate increase of 20.7 bpm, a mean increase in EDA of 1.1 μS, and a mean change in aggregate accelerometer measurements of 0.013 g. All of these changes were expected based on the results of prior work (Brunner et al., 2017; McAfee et al., 2017; Mirijello et al., 2023; Morris et al., 2024; Paprocki et al., 2022). The average skin temperature also increased by a mean of 1.66 °C, which makes physiologic sense, but was not found in prior literature (Morris et al., 2024). This scoping review by Morris et al investigates the effects of alcohol intoxication on the human body’s response to heat stress. This review found several articles documenting decreased core temperature and one study showing increased skin blood flow that could facilitate radiant dispersal of heat, but none found significant changes in skin temperature. The reason for this discrepancy is unclear and is confounded by the fact that these studies exposed their subjects to elevated ambient temperatures and this work did not. However, further work will need to be done to confirm this finding through replication.

5.3. Algorithm selection and feature importance

The XGBoost algorithm is most appropriate for this type of signal, since it boosts the high signal motion cues and down-weights the noisier physiological inputs, delivering balanced performance without added complexity. As seen from the feature importance plots, the X and Z-axis RMS (and related spectral powers) were the dominant predictors, while heart rate, EDA, and skin temperature metrics were used in fewer models. This pattern confirms that intoxication is mainly expressed in movement dynamics, echoing prior smartphone and wearable sensor studies ((McAfee et al., 2017; Nassi et al., 2022; Suffoletto et al., 2020)), that likewise found gait sway or axis-level RMS to be the strongest predictors of impairment.

5.4. BrAC vs SFST - Intoxication vs impairment

As discussed, the combination of features that produced the best performing predictive algorithm were different between the BrAC model and the SFST model (Table 4, Table 5). The BrAC model performs best when a combination of both accelerometry and physiologic features were used, whereas the SFST model performs best when only accelerometry features were used. Just prior to the SFSTs, participants had achieved a mean BrAC of 0.12g/dL (Table 2), which is above our impairment cut off of a BrAC of ≥ 0.08g/dL. Since there were statistically significant changes in accelerometry, HR, EDA, and body temperature, from the 20-minute sober baseline to the 20-minute period at the beginning of the SFST, it would make sense that the BrAC model would perform best when both the accelerometry and physiologic features were used (Table 3). The SFST assesses changes in movement that can be readily appreciated on a visual examination performed by law enforcement officers. Since this assessment is geared towards identifying changes in motor movement, judgement, and reaction time that are associated with impairment from alcohol use, it would make sense that a model based on accelerometry features would perform well for this classifier.

Given there were significant changes in both accelerometry and physiologic data features associated with an elevated BrAC, it is unclear why the SFST model performs worse when the physiologic data features are included. It is possible that this finding indicates that changes in EDA, HR, and body temperature caused by an increasing BrAC are not as strongly correlated with the changes in motor movements as expected. If this is borne out in further experimental work, this finding could represent the difference between identifying the physiology of an intoxicated state with the BrAC model vs. identifying the physiology of an impaired state due to intoxication with the SFST model.

5.5. Future Work

While our results are promising, our observational study design was limited by its lack of standardized participant movements and behaviors. While attempting to address this confounder by analyzing data from time periods when participants were exhibiting relatively uniform behavior, as demonstrated by the pedometry data (Table 3), we were unable to evaluate specific participant behaviors. For comparison, another study used only accelerometry and gyroscopic data from a paired wristwatch and smartphone system to evaluate gait patterns in a group of college students using BAC goggles that simulated alcohol intoxication over multiple walk tests with different degrees of “inebriation” (McAfee et al., 2017). By analyzing this specific behavior, they made digital phenotypes for sober and intoxicated walking that allowed them to create a machine learning algorithm with ~ 80% accuracy for models having similar parameters as those in the current study. Biometric data collected using smartwatch/smartphone systems could be used to create similar sober/intoxicated digital phenotypes for a variety of common behaviors (standing, sitting, typing, speaking, etc.). In future work, participants could be asked to perform similar stereotypic behaviors repeatedly throughout the experiment to develop multiple digital phenotypes. Machine learning algorithms developed from such training data could become sophisticated enough to reliably detect alcohol intoxication and impairment in complex real-world situations where people act unpredictably.

Once developed and validated on popular wrist-worn sensor product brands like the Apple Watch© or Garmin©, these algorithms could function as a passive detection system for a smartphone-based application designed to help those experiencing episodic acute alcohol intoxication or more chronic alcohol use disorder. Once intoxication/impairment are detected, the application could trigger personalized “just in time” interventions with the goal of harm reduction. Such interventions could include booking a ride home through a ride-sharing application, calling an emergency contact, calling a sobriety coach, or evening locking social media and gambling applications. This detection system also has potential to contribute to other biometric data streams used in population dashboards by state health departments as part of the CDC Data Modernization Initiative to mobilize treatment resources in areas of high alcohol use (Hanafi et al., 2025). However, if this sensitive personal data were to be collected and stored for use in a central database, particular care would need to be taken to ensure individuals are asked for consent to use the data, and to ensure privacy and de-identification. O therwise, the potential harm from surreptitious collection and disclosure of an individual’s alcohol use history could lead to significant distrust in the system and limited adoption (Domaradzka et al., 2024).

6. Limitations

There are several limitations to this study due to the observational nature of the study design, which introduced several uncontrolled variables. This study was performed at a single site with mostly young healthy Caucasian participants, which may limit the generalizability of the results. The data collection performed by the academy instructors was not standardized, leading to a different number of BrAC assessments and collection times which complicated the statistical analysis. The academy instructors do not routinely calibrate the breathalyzer machine they use for their training exercises, which could introduce error into the BrAC measurements used in this study. While the SFST assessments were performed by cadets who had received training on how to do the assessment, the cadets worked in groups to provide consensus scores, and all groups were actively being observed by senior instructors to make corrections, the results could still be biased due to cadet inexperience with working with intoxicated subjects. This study not only lacked behavioral controls to account for the random nature of human behavior, but also for group dynamics. During several study sessions the participants would play drinking games like beer pong, and this physical activity would cause appreciable changes in the biometrics of all the participants that played. This led to confounding as to what degree those changes were from doing to physical activity and how much were due to changes in BrAC. This could have been solved by a sober control group, which this study also did not have. Finally, several participants were found to be using nicotine products at several points during the study sessions despite academy instructor and study staff instructions not to. A subgroup analysis was performed comparing the biometrics of the nicotine users to the non-nicotine users, and there was no statistically significant difference. However, this still adds another possible confounding factor to the analysis.

7. Conclusion

Wearable biosensors are able to detect physiologic changes in participants intoxicated with alcohol. In a population that achieved a mean BrAC of 0.12g/dL, there were statistically significant changes in heart rate, body temperature, electrodermal activity, and aggregate accelerometry data compared to their own sober baselines. Using a BrAC level of ≥ 0.08g/dL and the results of a SFST to determine impairment, we were able to create several machine learning algorithms. The best performing BrAC model had an accuracy of 0.80 and used a combination of accelerometry and physiologic data features, and the best performing SFST model had an accuracy of 0.78 and used only accelerometry features. With more sophisticated data sets, it may be possible to develop algorithms with better accuracy. If this is able to be accomplished, wearable biosensors have the potential to become part of a ubiquitous passive detection system that could identify alcohol intoxication and impairment.