How a THC Brain-Imaging Classifier Defined Impairment

A brain-imaging classifier can only be evaluated in relation to the outcome used to label its training data. In a THC study using functional near-infrared spectroscopy, “impairment” was defined through agreement between clinical ratings and an algorithm incorporating heart rate and self-reported intoxication. It was not established by an on-road driving test.

The reference definition came first

Gilman and colleagues studied regular cannabis users in a double-blind crossover experiment involving individualized oral dronabinol and placebo. A total of 169 participants received study drug and had at least one post-dose scan.

The investigators measured prefrontal oxygenated-hemoglobin responses during a working-memory task. To construct the classification reference, they required concordance between two clinical raters and a separate physiological and subjective algorithm. Participants with discordant determinations did not supply the same clearly labeled training examples.

Classification and drug detection are different objectives

The investigators trained machine-learning models using temporal and connectivity features of the imaging signal. Their analyses distinguished participants operationally labeled impaired from comparison observations and separately examined scans labeled “not clearly impaired” after THC.

Importantly, receiving THC was not automatically treated as equivalent to impairment. Participants given similar doses could receive different reference labels. That is a useful conceptual distinction, but it does not turn the chosen label into an independently proven measure of every safety-sensitive activity.

What external validation still has to answer

A classifier’s accuracy depends on the population, the reference definition and the conditions of testing. Predictive value also depends on how common the target condition is in the population where the device is used. Performance in this controlled experiment should not be assumed for an unselected roadside population, other drugs or relevant medical conditions.

The journal issued a correction to Figure 2; the corrected original should accompany interpretation of the study.

This research asks an individual-classification question. Our article on prefrontal connectivity after THC concerns a related but distinct imaging approach. Neither kind of signal should be described as a validated legal conclusion about a person’s driving without evidence connecting the test to that intended use.

Primary research

Identification of ∆9-tetrahydrocannabinol (THC) impairment using functional brain imaging. DOI: 10.1038/s41386-021-01259-0.