Smartphone Sensors Show Promise for Detecting Subjective Cannabis Intoxication
A smartphone may be able to recognize patterns associated with feeling high—but a 2021 feasibility study found that the technology detects self-reported subjective intoxication, not objective impairment or cannabis use itself.
In the study, researchers led by Sang Won Bae at Stevens Institute of Technology monitored 57 young adults in Pittsburgh who reported using cannabis at least twice a week. Participants provided information about their cannabis use and rated how intoxicated they felt on a scale from 0 to 10 while their phones continuously collected sensor data. The researchers published their findings in Drug and Alcohol Dependence.
The team analyzed 102 smartphone features, including movement detected by the accelerometer, travel and location patterns from GPS, and other measures such as activity and environmental conditions. The researchers then used a machine-learning model called a Light Gradient Boosting Machine to classify reports as no intoxication, low intoxication, or moderate-to-intensive intoxication.
Using smartphone sensor data alone, the model achieved 67% accuracy on the study’s holdout test data. When researchers added the day of the week and time of day, accuracy rose to 90%. Time information alone produced 60% accuracy, suggesting that participants’ routines accounted for part of the model’s performance. The most informative sensor features included travel patterns and movement.
Those figures should be interpreted carefully. The study did not compare smartphone predictions with a laboratory measure of THC, a standardized psychomotor test, or an independently observed level of impairment. Instead, participants reported when they used cannabis and how high they felt. The model therefore identified patterns associated with subjective cannabis intoxication in this particular group—not a universally validated measure of whether someone was too impaired to drive or perform another task safely.
The researchers described the work as a proof of concept. A future system might use changes in phone-derived behavior to offer a just-in-time prompt, such as suggesting that a person avoid driving or arrange a ride. But such an application would require testing with larger and more diverse populations, independent measures of impairment, and data collected across different phones, settings, cannabis products, and patterns of use.
The study also had important limitations. Its sample included only 57 young adults, who reported 451 cannabis-use episodes. The labels used to train the model were based on participants’ self-reports, which can be affected by memory, expectations, tolerance, and differences in how people interpret an intoxication scale. The results may not generalize to older adults, occasional users, people who do not use cannabis, or individuals using cannabis alongside alcohol or other substances.
Detecting current cannabis impairment remains difficult. As the National Highway Traffic Safety Administration explains, there is no chemical equivalent to an alcohol breath test that reliably quantifies cannabis-related impairment. THC and its metabolites can remain detectable after the impairing effects have diminished, and the relationship between a test result and an individual’s level of impairment is inconsistent. The Centers for Disease Control and Prevention likewise notes that it is difficult to connect a person’s THC concentration with their driving impairment.
Smartphone sensing could eventually complement behavioral or clinical assessments, but the Stevens study was an early feasibility test rather than a ready-to-use impairment detector. Its main contribution was showing that ordinary phone sensors, combined with contextual information such as time and day, may contain signals related to how intoxicated some cannabis users feel in everyday life.
