Can a Contactless Car Breath Sensor Reliably Estimate Blood Alcohol? What 70 Volunteers Actually Showed

Short answer: A 2026 controlled human study found a strong association between an experimental infrared-and-carbon-dioxide automotive breath sensor and laboratory-measured whole-blood ethanol. In 70 volunteers across 187 study days, researchers assembled 9,163 closely timed breath–blood measurement pairs. Among the 9,090 pairs with positive blood alcohol, the reported Pearson correlation was 0.8875 (95% confidence interval, 0.8831–0.8918). That is evidence of an association across the study range—not proof that a particular sensor reading accurately identifies a named driver’s blood alcohol level, reliably classifies readings close to a legal threshold, or demonstrates actual driving impairment.

The distinction matters because a vehicle-based breath sensor measures ethanol in exhaled air, uses carbon dioxide to correct for dilution, and may display an alcohol value on a blood-equivalent scale. The contemporaneous reference in this study was alcohol measured in collected whole blood by gas chromatography with flame ionization detection. The two measurements are related but are not the same specimen, measurement process, or legal proposition. The original primary study is Pirooz and colleagues, “Development of the DADSS Breath Alcohol Sensor System for Automobiles: Technical Design and Human Participant Testing,” Sensors (2026), doi:10.3390/s26092685.

What did the 2026 automotive breath-sensor study establish?

The researchers demonstrated that successive generations of the Driver Alcohol Detection System for Safety (DADSS) breath sensor generated alcohol measurements that generally rose and fell with directly assayed blood ethanol under controlled laboratory conditions. The strongest published numerical findings concern pooled breath–blood association. The investigators did not supply a complete independent, person-level validation of classification errors at a legally meaningful cutoff, a trial of routine unprompted in-vehicle use, or measured on-road driving performance in this experiment.

Statistical strength at a population level is useful for engineering development. It cannot replace inquiry into how a particular reading was obtained, whether the instrument passed the relevant quality checks, whether a person other than the driver contributed a signal, or whether the reported display quantity is a breath reading, a calibrated equivalent, or directly assayed blood.

How was the sensor tested, and what counts as an independent participant?

Seventy screened adults, 21–55 years old, participated in study sessions conducted from November 2015 to August 2023. The publication reports a mean age of 29 years and 64% men. Repeated attendance was permitted, with at least 48 hours between separate study days, producing 187 study days. Those 187 study days are not 187 independent people, and the thousands of breath–blood measurements are not thousands of independent volunteers.

Participants fasted overnight and received a standardized light breakfast before alcohol administration. Beverage alcohol doses ranged from 0.3 to 0.9 grams of ethanol per kilogram of body weight, with several consumption schedules and beverage types. During experimental sessions participants sat in a laboratory recliner. Investigators collected serial blood samples and had participants exhale toward nearby DADSS sensors; a reference Alco-Sensor FST breath instrument was also used. Blood and DADSS results were paired within three minutes when possible. The average session lasted approximately three hours.

The setup is important to interpretation: supervised, repeated measurements near fixed sensors in a laboratory are different from uninterrupted monitoring of an unidentified driver in a moving, ventilated cabin. Figures in the original publication show sensor devices mounted on stands in front of volunteers. Images of sensors installed in prototype vehicles demonstrate possible placement; they do not themselves establish operating sensitivity, specificity, or false-alarm rates in traffic.

Which measurements were direct blood tests, breath tests, and calculated equivalents?

The researchers collected blood through an indwelling intravenous catheter into preservative-containing gray-top tubes. Their whole-blood analytical protocol used protein precipitation and gas chromatography with flame ionization detection. The publication reports intra-assay and inter-assay coefficients of variation of 1.10% and 1.96%, respectively, for that laboratory method. These figures describe the analytical method; they are not uncertainty figures for converting a passive breath signal into a legal blood result.

By contrast, the DADSS system detected infrared absorption from ethanol and carbon dioxide in exhaled air. Its calculation addresses how breath is diluted after leaving the mouth. The study presents the DADSS sensor results alongside reference blood concentrations in milligrams per deciliter, using a 2100:1 blood–breath conversion convention. Thus, a DADSS value depicted on a blood-style mg/dL scale does not mean blood was collected or chromatographically analyzed by the sensor.

The independent Alco-Sensor FST was another breath instrument, requiring a directed, sustained exhalation into a mouthpiece. It is an additional breath comparator, not a substitute for the study’s directly assayed blood reference. These distinctions are central to forensic breath-alcohol measurement and blood ethanol chromatography.

What numerical results support the blood–breath association?

The plotted scatter spans a substantial concentration range. A highly significant regression may be expected when both measurement systems increase with exposure, even if individual readings exhibit clinically or legally consequential differences. A correlation of 0.8875 does not mean that 88.75% of individual readings are accurate, that 88.75% of impaired drivers are identified, or that an instrument has an 88.75% pass rate. Those are distinct performance statistics that require different study designs.

The publication pooled four DADSS sensor generations after identifying some differences in their correlations. Pooling increases data available for a general relationship but can conceal generation-specific offset, calibration, sampling geometry, and threshold performance. A case involving a specified hardware and software version should not treat a pooled result as an instrument-specific validation for every deployment configuration.

How did performance differ across device generations?

Testing involved generations 3.0, 3.1, 3.2, and 3.3, introduced sequentially. The paired observations were not equally distributed: 4,851 pairs came from version 3.0, 1,372 from 3.1, 2,319 from 3.2, and 621 from 3.3. A pooled result therefore gives substantially more weight to earlier tested devices than to the last generation.

One particularly important engineering qualification appears in the source’s hardware table: generations 3.0 through 3.3 are listed with directed-exhalation operation, whereas generation 4.0 is described as adding passive monitoring. The study methods likewise describe volunteers exhaling toward sensors. The 2026 report is therefore not equivalent to a field trial proving that an entirely unprompted generation-4 device can identify the driver’s alcohol concentration in every passenger compartment. The terms contactless, passive, and unprompted should not be used interchangeably.

How does infrared alcohol detection use carbon dioxide to address breath dilution?

The device measures light absorption in two infrared channels: one sensitive to ethanol and one to carbon dioxide. Breath leaving the airway mixes with ambient air, which lowers the measured concentration of both gases. The engineering idea is to use the exhaled carbon-dioxide signal to estimate dilution and then infer the undiluted respiratory ethanol signal. The reported infrared sampling frequency is five measurements per second.

That approach is physiologically plausible because exhaled carbon dioxide is much more concentrated than typical room-air carbon dioxide. But correcting dilution does not remove every biological or sampling source of variation. Breath ethanol depends on airway exchange, breath temperature, timing after drinking, mouth contamination, the device’s field of view, and the assumptions used to translate breath units into a blood-style display.

The device design adopts a conventional 2100:1 blood–breath partition factor. An assigned factor is an engineering or reporting convention; it is not an independent individual measurement of that person’s actual partition ratio. A defensible case report must identify what the display states and which conversion, if any, was performed before comparing it with a statutory threshold.

Does the reported r = 0.8875 prove accuracy at 0.08% BAC?

No. An overall correlation answers whether the values tend to increase together, not whether the sensor correctly classifies a participant just above or below any specified cutoff. A useful threshold validation would report, under a specified operating configuration, the number of correctly and incorrectly classified persons on each side of the decision limit, ideally with predefined tolerance bands and a separate test sample.

The 2026 publication does not present a complete person-level confusion matrix, sensitivity, specificity, near-threshold misclassification distribution, or independently validated prediction interval for an individual vehicle occupant. The reported blood-assay precision does not supply those missing sensor metrics. Researchers also performed an alignment step to address systematic overestimation of DADSS values. Alignment is not itself proof of calibration transfer to every future sensor or environmental condition.

For litigation, the exact legal question must be specified: whether the issue is detecting some ethanol, screening for a programmed vehicle threshold, establishing a laboratory blood concentration, or explaining a person’s ability to drive. Each claim has a different required validation target. A system may be promising as a screening and prevention technology without becoming conclusive retrospective evidence in a contested individual case.

What did the investigators observe about mouth alcohol and elapsed time?

After a participant drank alcohol, the breath traces sometimes showed a transient elevation attributed by the investigators to residual alcohol in the mouth. Their graphical examples show this effect dissipating in roughly 15–20 minutes before the paired breath time courses more closely followed blood. These time courses are examples under controlled dosing and measurement conditions, not a validated observation period or a universal instruction to disregard every earlier sample.

The distinction becomes consequential for sensor designs that continuously sample the air near an occupant. Mouth alcohol may briefly elevate a breath signal independently of proportional changes in circulating ethanol. A case-specific assessment would examine drinking time, sample timing, repeated signals, any quality flags, and independent specimen data. The study does not show that a generic 15-minute waiting period eliminates all possible interference from products, contaminants, nearby drinkers, or cabin sources.

Did participants’ confidence about driving correspond to measured impairment?

The study also obtained subjective ratings every approximately 20 minutes, including how intoxicated people felt and whether they would drive. The authors reported that some participants remained willing or felt safe to drive when alcohol levels reached the displayed 80 mg/dL threshold. The paper highlights a subset of 17 participants given 0.9 g/kg in a graphical example and discusses a figure near 30% for a driving-confidence response.

Those statements concern subjective self-assessment, not observed driving, crash risk, or a judicial finding of impairment. The participants were seated for sensor validation, not assigned to drive. Nor is the reported confidence result a validated individual diagnostic test. Its appropriate use is to illustrate the potential gap between what someone reports and a biological measurement, while maintaining the distinction between willingness to drive, functional performance, and an event-specific legal inference.

What remains untested for real vehicles and contested cases?

The original study supports laboratory feasibility but leaves several practical questions open. The paired observations came from screened healthy adults in supervised sessions, with sensor placements and directed breathing procedures controlled by research personnel. Production vehicles can vary in cabin geometry, air conditioning, passenger presence, sensor placement, driver posture, and environmental alcohol. The paper’s illustrated vehicle installations do not provide full field-performance tables for those conditions.

Equally, it does not establish the identity of the alcohol source in a mixed-occupant environment; the analytical system detects volatile ethanol, not the biological identity of its source. CO2-based dilution correction is not a person-identification method. Sensor response also should not be equated with a criminal evidential breath test that has a separately approved design, quality-assurance program, and jurisdiction-specific evidentiary status.

Selection matters: study volunteers were medically screened and were generally young adults, while real drivers span broader ages and health conditions. Repeated readings from the same participants create statistical dependence. The paper reports pooled association, not independent-person accuracy across varied real-world environmental strata or the prospective rate of vehicle false lockouts. Nothing in the report demonstrates that a particular operator, driver, manufacturer, or state implemented a specified threshold in a production automobile.

How should an expert evaluate a vehicle breath-sensor result?

Begin with the complete original source record: hardware generation, firmware/software version, installation geometry, sampling prompts, instrument response logs, time stamps, calibration or verification history, and actual displayed units. Determine whether the measurement represents raw exhaled alcohol, a CO2-corrected breath estimate, or a calculated equivalent on the blood concentration scale. The source’s regression fit is a population relationship, not a substituted venous laboratory report.

Next separate specimen validity from inferential validity. A contemporaneous whole-blood GC result, if properly collected and tested, answers a different question from an optical signal in the cabin. The strength of a matched sensor–blood association cannot establish an individual’s blood alcohol at an unmeasured earlier moment without additional pharmacokinetic assumptions. The distinction is consistent with the separate analysis of ethanol pharmacokinetics and model limitations.

Finally, determine whether the expert opinion concerns technical screening, a legal threshold, functional impairment, or why a particular incident occurred. A careful opinion stays within the validation population and measurement endpoints and identifies any missing error-rate, environmental, source-attribution, and person-specific evidence. This avoids both dismissing a promising sensor solely because it measures breath and exaggerating its evidentiary meaning because it correlates strongly with blood.

Frequently asked questions

Is DADSS the same as a police evidential breathalyzer?

No. The DADSS generations studied used infrared ethanol and carbon-dioxide signals to support contactless automotive sensing. An evidential breath instrument’s statutory approval, calibration requirements, sampling protocol, and admissibility depend on separate specifications and rules.

Were the 9,163 paired measurements obtained from 9,163 drivers?

No. They were repeated observations from 70 volunteers over 187 study days. Treating each paired reading as a distinct independent human subject would overstate the independent sample size.

Was actual blood alcohol directly measured?

Yes, in the research laboratory. Investigators collected whole blood and quantified ethanol by gas chromatography with flame ionization detection. The DADSS device itself measured breath; it did not analyze a blood specimen.

Can the published correlation establish an individual’s BAC?

Not by itself. Correlation describes covariation across the analyzed observations. It does not supply a person-specific prediction interval, an individual partition ratio, or an error-rate analysis at a disputed threshold.

Does a 2100:1 conversion factor prove every person has that ratio?

No. It is a reporting convention built into the engineering approach. Real breath–blood relationships can differ because of physiological and sampling factors.

Did the trial validate truly unprompted in-car monitoring?

No comprehensive real-world unprompted validation is reported here. Human testing involved generations 3.0–3.3 and supervised exhalation toward sensors; the paper’s hardware table identifies passive monitoring with the later planned generation 4.0.

Did the study prove that a detected driver was impaired or caused a collision?

No. Researchers measured alcohol and collected subjective ratings. They did not conduct an on-road driving trial, reconstruct an actual crash, or establish individual functional impairment or causation.

Original primary scientific source

Pirooz K, Allen T, Spicer R, et al. Development of the DADSS Breath Alcohol Sensor System for Automobiles: Technical Design and Human Participant Testing. Sensors. 2026;26(9):2685. https://doi.org/10.3390/s26092685. This analysis addresses the published study’s own methods, cohort, repeated paired measurements, and limitations, not unpublished proprietary test results.