THC Brain-Imaging Accuracy: Why a 90% Figure Does Not Mean Every Impaired Person Was Detected
An experimental brain-imaging classifier achieved 90% accuracy against a clinical definition of THC impairment, but its recall was 56%. Those figures answer different questions: high overall accuracy did not mean the method detected every clinically impaired scan.
A clinical reference, rather than a driving verdict
The randomized crossover study enrolled 183 adults who used cannabis regularly. Participants received individually selected oral dronabinol doses, ranging from 5 to 80 mg, and placebo on separate visits. Resting-state imaging measured prefrontal oxygenated and deoxygenated hemoglobin signals; it did not chemically measure THC in blood.
The reference classification combined clinician judgments with an algorithm using heart rate and reported intoxication. This was a defined research standard, not an independently established gold standard for unsafe driving. Plasma THC and metabolite concentrations were not measured.
Read accuracy alongside recall and precision
For resting-state scans, the reported accuracy was 0.90, recall 0.56 and precision 0.60. Recall describes the proportion of reference-positive scans identified. Precision describes the proportion of predicted positives that agreed with the reference. A dataset containing many more nonimpaired than impaired scans can produce high accuracy while still missing a substantial fraction of positive cases.
The false-positive rate was 0.05. That is not interchangeable with saying that only 5% of positive classifications were wrong. These quantities use different denominators.
What the validation did well—and what remains untested
All scans from each participant stayed within the same cross-validation fold, reducing the risk that a person’s scans appeared in both training and testing. Nevertheless, internal cross-validation is not external roadside validation.
Resting-state and working-memory-task models did not differ significantly in performance. The imaging classifier improved several metrics compared with expanded field tests, but recall favored the field tests and the F1-score difference was not significant.
Questions for evaluating a proposed device
Request the reference definition, confusion matrix, decision threshold and validation population. Ask whether independent testing includes inhaled cannabis, sleep deprivation, alcohol and combined exposures. Those distinctions determine whether a promising laboratory result applies to the person and circumstances being evaluated. The study supports further validation; it does not establish a universal brain-scan verdict on individual driving impairment.
Related: Interpreting THC concentration and impairment.
Primary research
Detection of Δ9-Tetrahydrocannabinol Impairment Using Resting-State Functional Near-Infrared Spectroscopy: A Randomized Clinical Trial. DOI: 10.1001/jamanetworkopen.2025.56647.
