Killian and colleagues paired smartphone accelerometer readings with ankle-monitor readings in a field dataset from 13 participants. They divided the motion data into short windows and trained a model to classify those windows using a transdermal alcohol threshold.
The researchers shifted the transdermal readings by 45 minutes before aligning them with phone motion. Their best classifier achieved 77.5% accuracy under the reported evaluation. That figure describes agreement with labels produced by a specific threshold and alignment. The system did not directly measure blood alcohol or clinical impairment.
Start by identifying the reference endpoint and the unit counted as correct. Then look for validation on people and circumstances that were absent from training. An accuracy percentage for windows within a small field dataset leaves open how the classifier would perform for a new person, a different phone location, or a different drinking pattern. A model intended to trigger a wellness prompt also answers a different question from an opinion about one person at one time.
Killian et al., Knowledge Discovery in Healthcare Data workshop proceedings, 2019