EVIDENCE LIBRARY
How Well Did a Personal Breathalyzer Agree With a Police-Grade Device?
A 337-person study found strong correlation between two breath devices, but individual agreement and threshold classification require separate analysis. Its table also exposes mislabeled diagnostic percentages.
A personal breathalyzer tracked a police-grade breath device closely at the group level, but that did not make their individual readings interchangeable. A 337-person field study found a small average overreading by the personal device alongside much wider limits of agreement. Its classification table also requires care: the percentages labeled sensitivity and specificity in the text use predictive-value denominators.
Two breath devices, not a blood reference
Riordan and colleagues tested young adults outside three concerts. Participants were 17–28 years old, with a mean age of 18.4 years; 96% were students. Each provided paired readings from a BACtrack Mobile BT-M5 and a LifeLoc FC10, in randomized order. Both instruments used fuel-cell sensors.
The results were expressed as blood-equivalent alcohol concentrations in g/dL. Neither instrument directly measured blood. Thus, agreement with the LifeLoc assessed agreement between two breath-based estimates, rather than accuracy against a simultaneous laboratory blood result.
This distinction matters when a case contains an actual blood measurement. The site’s blood–breath comparison in 100 volunteers addresses a separate source of variation.
Correlation did not establish close individual agreement
The police-grade device’s mean result was 0.077 g/dL; the personal device’s mean was 0.085 g/dL. The paired mean difference, personal minus reference, was approximately +0.008 g/dL, with a reported 95% confidence interval of +0.0062 to +0.0096 g/dL. The correlation was high, r = 0.955.
Correlation describes how readings vary together. It does not tell us that two instruments give the same result for each person. A device can preserve the ranking of participants while differing meaningfully from its comparator.
The authors’ percentage-difference analysis reported a mean bias of 6.03% and 95% limits of agreement from −73.71% to +85.77%. Those percentage limits should not be read as fixed g/dL error bars or as a person’s probability of having a particular BAC. The differences were not normally distributed, and the paper identified a small concentration-related pattern.
What the threshold table actually shows
For a research comparison at 0.05 g/dL, the published table gives these paired classifications:
| Personal-device classification | Reference positive | Reference negative |
|---|---|---|
| Positive | 219 | 16 |
| Negative | 4 | 98 |
Using those counts, overall agreement is 317/337, or 94.1%. Sensitivity relative to the reference is 219/223, or 98.2%; specificity is 98/114, or 86.0%. These are calculations from the published table, not corrected results obtained from participant-level records.
The text instead calls 93.2% sensitivity and 96.1% specificity. Those numbers correspond to positive predictive value, 219/235, and negative predictive value, 98/102. The text also reports an agreement count of 313, inconsistent with the table’s 317. These discrepancies should be resolved from original records before treating the paper as a precise device-validation benchmark.
The table includes four personal-device negatives when the comparator was positive. An average tendency to overread therefore does not guarantee that an individual reading cannot underclassify someone. The research threshold is not a statement of current law or a clearance to drive.
Field conditions and study limits
The final analysis excluded many people, including those reporting drinking within ten minutes before interview and those without paired readings. The interview added time, but it did not ensure that every participant had the manufacturer’s full waiting interval before testing. Recent drinking, burping and residual mouth alcohol remain relevant field-testing concerns.
The site’s three-case analysis of mouth-alcohol detection explains why a breath instrument’s safeguards and sampling conditions need their own evaluation.
This was a predominantly student sample using one personal-device model available at the time. It did not test all consumer instruments, establish present-day approval for evidentiary use, or measure driving performance. A high correlation in this setting cannot establish a particular driver’s functional ability or reconstruct an earlier alcohol concentration.
Primary source
Riordan BC and colleagues. The accuracy and promise of personal breathalysers for research: Steps toward a cost-effective reliable measure of alcohol intoxication? Digital Health. 2017. https://doi.org/10.1177/2055207617746752.