Identification of ∆9-tetrahydrocannabinol (THC) impairment using functional brain imaging

Abstract

The primary cannabinoid in cannabis, Δ9-tetrahydrocannabinol (THC), causes intoxication and impaired function, with implications for traffic, workplace, and other situational safety risks. There are currently no evidence-based methods to detect cannabis-impaired driving, and current field sobriety tests with gold-standard, drug recognition evaluations are resource-intensive and may be prone to bias. This study evaluated the capability of a simple, portable imaging method to accurately detect individuals with THC impairment. In this double-blind, randomized, cross-over study, 169 cannabis users, aged 18-55 years, underwent functional near-infrared spectroscopy (fNIRS) before and after receiving oral THC and placebo, at study visits one week apart. Impairment was defined by convergent classification by consensus clinical ratings and an algorithm based on post-dose tachycardia and self-rated "high." Our primary outcome, prefrontal cortex (PFC) oxygenated hemoglobin concentration (HbO), was increased after THC only in participants operationalized as impaired, independent of THC dose. ML models using fNIRS time course features and connectivity matrices identified impairment with 76.4% accuracy, 69.8% positive predictive value (PPV), and 10% false-positive rate using convergent classification as ground truth, which exceeded Drug Recognition Evaluator-conducted expanded field sobriety examination (67.8% accuracy, 35.4% PPV, and 35.4% false-positive rate). These findings demonstrate that PFC response activation patterns and connectivity produce a neural signature of impairment, and that PFC signal, measured with fNIRS, can be used as a sole input to ML models to objectively determine impairment from THC intoxication at the individual level. Future work is warranted to determine the specificity of this classifier to acute THC impairment.ClinicalTrials.gov Identifier: NCT03655717.

Conflict of interest statement

JMG and AEE reported having a patent pending (WO 2018/027151) to use fNIRS to measure intoxication. BK received funds from a NIH STTR grant to Highlight-I to conduct the ML analyses. MAH reported scientific advising/consulting for PinneyAssociates, Cannabix, the Canadian Nuclear Safety Commission, Nextage Therapeutics, Suncor Energy, and Dynacare Laboratories. All other authors declare they have no competing interests.

© 2022. The Author(s).

Figures

Fig. 1. Design of probe and machine learning models.

A The continuous wave near-infrared spectroscopy (NIRS) machine was used to measure changes in oxygenated hemoglobin (HbO) and deoxygenated hemoglobin (deoxy-Hb). The NIRS probe comprised of eight sources (red) and seven detectors (yellow) placed over the prefrontal brain region (forehead) of each participant. The mid-column of the probe was placed over Fpz, with the lowest probes located along the F5-Fp1-Fpz-Fp2-F6 line, in accordance with the International 10–20 Placement System. The distance between pairs of source and detector probes ranged from 2.5 - 3 cm. The midpoint of the source-detector distance was defined as channel (Ch) location, labeled numerically (1–20) in the above schematic. The channels were grouped into regions of interest, as illustrated in the schematic. We defined five regions on interest (ROIs) based on channel location. These ROIs are middle prefrontal cortex (MPFC, channels 7, 8, 9, 10, 11, 12, 13, 14); right dorsolateral prefrontal cortex (RDLPFC, channels 15, 17, 18); right ventrolateral prefrontal cortex (RVLPFC, channels 16, 19, 20); left dorsolateral prefrontal cortex (LDLPFC, channels 1, 2, 5); and left ventrolateral prefrontal cortex (LVLPFC, channels 3, 4, 6). B We extracted 95 numerical values (19 from each of the 5 ROIs): mean HbO values for time segments 0–5, 5–10, 10–15, 15–20, 20–25, 25–30, 30–35, and 35–40 s, slope (5–15 s), skewness (0–15 s), kurtosis (0–15 s), area under the curve (0–15, 15–40 s), time to first and second extremum of HbO, magnitude of first and second extremum of HbO, and average and standard deviation of HbO after the first extremum. We used these 95 predictive features to train XGBoost to predict impairment. C We computed pairwise correlations between the HbO concentration change values of all possible channel pairs. We computed dynamic connectivity matrices with sliding windows of 300 time points, and a skip of 100 time points, so that each window corresponded to a small segment of the scan. We then trained an RNN model architecture with the sliding-window correlation matrix feature vector as the input. In this RNN, the core component consisted of fully connected layers that mapped the input to a latent representation, which in turn fed to a hidden state with recurrent connections. A probabilistic prediction was computed at every time point by applying a fully connected layer to the hidden state. We used tanh for all nonlinearities and implemented a fully connected neural network with 5 hidden layers and 128, 64, 32, and 16 nodes. The output of this model was a fully connected neural net with a single hidden layer of 64 dimensions and consisted of recurrent connections that captured temporal dynamics.

Fig. 2. Psychological and physiological measures of participants by impairment and drug status.

A Participants were assessed for impairment via an algorithmic approach and a clinical approach into “Impaired” and “Not Clearly Impaired” groupings, which were then used to build and test a classifier for impairment using fNIRS data. B Time course of A self-reported high (0–100 scale to answer the question: “Are you high right now?”, 0 being “Not at all” and 100 being “Extremely”) and B heart rate (beats per minute) were averaged over (1) individuals identified as impaired post-THC by consensus ratings (red triangles), (2) individuals identified as not clearly impaired post-THC by concordant ratings (black triangles), and (3) all individuals post-placebo (filled circles). Impairment status post-THC was determined based on concordant ratings between two rating methods, algorithmic and clinical experts. Error bars are 95% confidence intervals for the mean generated via bootstrap. Gray bars show time windows for the pre-dose (Scan 1) and post-dose (Scans 2 and 3) fNIRS imaging.

Fig. 3. HbO response at pre- and post-dose by consensus rating (columns 1 and 2), placebo (column 3) and eFSR rating (columns 4 and 5).

The time course of mean HbO changes (µM) in each ROI are plotted in those who received THC and were impaired (Col 1), received THC and were unimpaired (Col 2), received placebo THC (Col 3), received an impaired rating from an eFST (Col 4), and in those who received a not-impaired rating from an eFST (Col 5). Pre-dose HbO response is shown in blue, and peak dose (Scan 2 or 3, whichever had the higher intoxication rating) is shown in red. Blue and red lines represent the mean and the standard error of the mean in each group. Yellow lines represent single timepoints in which the group differences between pre and peak HbO were significant (using a paired t-test with Benjamini-Hochberg FDR correction and p < 0.05). For statistical purposes, only matched pre and post-dose scans were included, which resulted in a sample size of 77 (out of 80) concordant impaired subjects and 55 (out of 57) concordant not clearly impaired subjects. MPFC middle prefrontal cortex, RDLPFC right dorsolateral prefrontal, RVLPFC right ventrolateral prefrontal cortex, LDLPFC left dorsolateral prefrontal cortex, and LVLPFC left ventrolateral prefrontal cortex.

Fig. 4. Machine learning metrics.

A TS, time series, CONN, connectivity, eFST, extended field sobriety test. B Receiver operative characteristic (ROC) curve for combined fNIRS models.