An interpretable and predictive connectivity-based neural signature for chronic cannabis use

Abstract

Background:

Cannabis is one of the most widely used substances in the world, with usage trending upward in recent years. However, though the psychiatric burden associated with maladaptive cannabis use has been well-established, reliable and interpretable biomarkers associated with chronic use remain elusive. In this study, we combine large-scale functional magnetic resonance imaging (fMRI) with machine learning and network analysis and develop an interpretable decoding model that offers both accurate prediction and novel insights into chronic cannabis use.

Methods:

Chronic cannabis users (n=166) and non-using healthy controls (n=124) completed a cue-elicited craving task during fMRI. Linear machine learning methods were used to classify individuals into chronic users and non-users based on whole-brain functional connectivity. Network analysis was used to identify the most predictive regions and communities.

Results:

We obtained high (~80% out-of-sample) accuracy across four different classification models, demonstrating that task-evoked connectivity can successfully differentiate chronic cannabis users from non-users. We also identified key predictive regions implicating motor, sensory, attention and craving-related areas, as well as a core set of brain networks that contribute to successful classification. Importantly, the most predictive networks also strongly correlated with cannabis craving within the chronic user group.

Conclusion:

This novel approach produced a neural signature of chronic cannabis use that is both accurate in terms of out-of-sample prediction and interpretable in terms of predictive networks and their relation to cannabis craving.

INTRODUCTION

Cannabis is one of the most widely used psychoactive substances in the world, with over 192 million people having reported some level of usage in 2020 (1,2). Nearly one in ten individuals reports using cannabis daily, and approximately the same proportion of people develop addictive symptomatology¹. Though cannabis has a number of therapeutic uses (3), the adverse effects and psychiatric burden of heavy, chronic cannabis use are well-established (4–8). As use increases and governments begin to favor decriminalization and legalization, it is more important than ever to understand the neurobiological underpinnings of chronic cannabis use.

Reliable and interpretable biomarkers of chronic cannabis use have the potential to inform clinical assessment and treatment of long-term cannabis use, and provide insight into its impacts on brain functioning. Functional magnetic resonance imaging (fMRI) combined with machine learning is a promising approach towards this goal (9–11). Models trained on neural activity patterns are capable of classifying many forms of psychiatric dysfunction, clinical symptomatology and even underlying mental processes (11–13). Nevertheless, machine learning models can often suffer from poor interpretability. Biomarkers are most useful when they are also understandable: knowing why a biomarker works or fails is crucial for its use in the clinic and for it to contribute to our understanding of neural and psychological dysfunction (14,15).

In this study, we aim to identify an accurate and interpretable neural signature of chronic cannabis use by using 1) classification to decode fMRI functional connectivity (16), 2) network analysis to discover the networks of brain regions most important to this decoding, and 3) generalized linear modeling to analyze the decoding model in relationship to cannabis craving. To this end, we measure whole-brain functional connectivity in a large fMRI dataset (17,18) of individuals with and without chronic cannabis use doing a cue-elicited cannabis craving task. To date, this is the largest fMRI study (n=290) of chronic cannabis use and the first to classify chronic cannabis use.

We show that we can accurately decode chronic cannabis use from whole-brain functional connectivity, with high out-of-sample accuracy (82.7%), area under the curve, precision, and recall. Treating these decoding model weights as a mapping between the brain’s functional connectivity and chronic cannabis use, we then use network analysis on these weights to find the specific brain networks most important for accurate decoding. Importantly, we also find that these predictive networks are also associated with cannabis users’ cannabis cue-elicited craving, a central feature of chronic use (19,20). The results demonstrate the utility of decoding models to produce accurate and interpretable biomarkers.

MATERIALS AND METHODS

Sample characteristics

This sample came from two pre-existing datasets (17,18) [n=125 and n=198 respectively] of participants with and without chronic cannabis use recruited from the community (i.e., not treatment-seeking or inpatient). All participants were screened via urinalysis to confirm cannabis use status and to exclude for the presence of other drugs. Other criteria included: right-handedness, and the absence of psychosis (current and history) and traumatic brain injury. Participants completed the task after a 72-hour abstinence from cannabis use. Participants completed the Marijuana Craving Questionnaire (21), immediately before and after their MRI scan, the Marijuana Withdrawal Checklist (22) immediately before the scan and the Marijuana Problems Survey (23). See the original studies (17,18) for more details.

For this study, 15 cannabis users were excluded for not endorsing at least one current or lifetime symptom of maladaptive cannabis use, from the Substance Use Disorder modules of the SCID IV (24). Participants without chronic cannabis use reported no such symptoms. 18 participants were excluded for excessive motion (see Additional fMRI denoising in the supplement), leaving a final sample of 290 participants (chronic cannabis users [CC] n=166, healthy controls [HC] n=124 respectively). The two groups did not differ in terms of age but CC were significantly more male and HC had significantly more years of education. See Table 1 and Accounting for differences between groups in the supplement for more details.

Table 1. Participant details by group.

Demographic and substance use details for the cannabis users and non-users (HCs) are provided. For Males the count and % are provided for each sample. For all other variables, means and standard deviations in parentheses are provided. The appropriate statistical procedure is used to test for the presence of group differences. Mann-Whitney U (MWU) tests were used to account for non-normality in the distributions.

	Cannabis Users	Healthy Controls	Group difference
Males	100 / 60%	53 / 43%	X2 P=0.004
Age	30.1 (10.5)	27.4 (7.9)	MWU P=0.142
Years of education	16.2 (2.48)	13.4 (2.82)	MWU P<1e-5
# cigarette smoking days last 90	14.64 (30.47)	0.35 (2.6)	MWU P<1e-8
# drinking days last 90	16.9 (20.38)	7.48 (14.03)	MWU P<1e-5
# cannabis use days last 90	69.02 (15.28)	0.02 (0.17)	MWU P<1e-10

Task design

For the 2009 dataset, the task consisted of two runs each of a pseudo randomized order of 12 tactile/visual stimulus presentations of cannabis cues (e.g., a joint) and neutral cues (e.g., a pencil). Cues were presented for 20s, followed by a 5s period where cannabis craving was self-reported on an 11-point scale, followed by a 20s fixation. The same task was used for the 2016 dataset but with an additional cue type (participant’s chosen fruit) for 3 total cue types and 18 presentations, per run. All participants had 2 runs; the 2009 and 2016 datasets had 281 and 405 TRs per run, respectively. Both versions of the task conform to well-established standards for studying drug cue-induced effects (25).

Although there are differences in the collection sites and scanner parameters between the datasets, they have highly similar task designs with overlapping stimulus sets and the same number of cannabis trials, cue presentation times and craving assessments. We preprocess the datasets the same way and combine them in our classification pipeline, while ensuring a proportional split for the different datasets and the group label in the training and testing sets.

fMRI acquisition, denoising, parcellation & functional connectivity

fMRI data was collected during task performance in a 3T MRI (see fMRI acquisition in the supplement). Fmriprep (26) was used for standard preprocessing steps, including coregistration, normalization and nuisance variable estimation. 18 participants were dropped for high motion, and remaining high motion volumes were despiked using ArtRepair (27). For each participant, an average time series was computed for each region in the Stanford functional atlas (28) and then correlated (Pearson) pairwise with all other regions (see Brain parcellation and Functional connectivity in the supplement). The non-redundant functional correlations (“functional connectivity”) were the inputs to our linear classifiers (Fig. 1A)

Fig. 1. Machine learning pipeline.

(A) Raw voxelwise time series are preprocessed using the fmriprep preprocessing pipeline. Minimally preprocessed files are brain-masked and smoothed with a 4mm FWHM Gaussian kernel. Nuisance/task regression is performed (see for list of regressors used). Clean voxelwise time series is parcellated into 90 functional ROIs using Stanford functional atlas. (B) Parcellated data are divided into 2 sets; the training set is used for training and cross-validation, the testing set is used to evaluate the optimized classification models (shown in the cylinders). The optimization set is further divided into 5 folds for cross-validation. For each split, 4 of the folds are used for training (blue) and the remaining 1 fold is used for testing (green). Four linear classification algorithms are selected for hyperparameter tuning (L1- and L2-penalized logistic regression and linear support vector classification). Regularization strength is selected as an additional hyperparameter. Cross-validated accuracy, AUC, and precision/recall scores are used as performance metrics. (C) The optimized hyperparameter tuned model is re-trained with the full training dataset and evaluated using the testing dataset. (D) Subject-specific connectivity matrices are then weighted by the model weights and used in a network analysis to characterize the brain networks important for the prediction of chronic cannabis. The network properties are also compared to run-by-run cannabis craving in chronic cannabis users.

Linear classification

Four linear classification algorithms [L1- and L2-regularized Logistic Regression (LR) and linear Support Vector Machine (SVM) (29–31)], were used to train models to predict (i.e., classify) whether functional connectivity data was a CC or HC (i.e., the class label). Various regularization strengths (0.001, .05, 1, 5, 10) were tested for all four algorithms, with larger values reflecting stronger (L1 or L2) penalization. See Details about Linear Classification in the supplement.

The full dataset was divided into training and testing sets, using an 80/20 split: the training set (80%) included both runs from 232 participants and the testing set (20%) included the other 58 participants (Fig. 1B and 1C). The train-test split preserved the proportion of overall group labels (CC or HC), and included participants from the two datasets in proportion to their percentage of the overall sample, to avoid overfitting due to group imbalances or dataset-specific noise.

Cross-validated classifier training

5-fold cross-validation on the 80% training set was used to optimize the penalty strength (α) across the 4 candidate algorithms (Fig. 1B). The training set was randomly split into five equally sized subsets (folds), stratified by group label (CC or HC) to ensure the group proportions were the same as in the full dataset. For each split, a model was trained on four of the folds and predicted the group label (CC or HC) on the remaining fifth fold, with each of the five folds acting as the prediction set in turn. Training inputs were the run-level functional connectivity; participant-level predictions were computed from the participants’ average functional connectivity across the two runs.

Evaluating classifier performance

For each of the 4 algorithms, the regularization strength resulting in the highest cross-validated training prediction accuracy was used to fit a new model on all of the training data. The model with the best testing set participant-level prediction accuracy and Area Under the Receiver Operating Characteristic (ROC) Curve (AUC) score (64) (Fig. 1C) was used for further analysis.

Predictive network analysis

The model weights and functional connectivity together reflect the relative importances of each region-to-region functional connectivity for the model’s prediction of whether a participant (or run) is CC or HC. To estimate this, we multiplied each participant’s functional correlations by the model weights to produce weighted connectivity matrices - which we refer to as “predictive connectivity” (Fig. 2). Then, we performed network analysis on these predictive connectivity matrices to examine network properties that helped distinguish CC from HC (Fig. 3).

Fig. 2. Generating subject-specific weighted connectivity matrices.

For each run for a participant, the absolute values of the functional connectivity were element-wise multiplied by the absolute values of the L₁ SVM model weights to produce a weighted functional connectivity matrix (shown for a single, random run). The weighted functional connectivity represents the importance of each model-weighted functional correlation to the resulting prediction for that participant: larger values represent a larger contribution. Run weighted connectivities were then averaged across runs for each participant to generate the participant-level weighted connectivity matrices. These were thresholded to 2% sparsity to restrict to only highly informative connections.

Fig. 3. Network properties workflow.

Network properties are calculated at three different levels to characterize the subject-specific networks. (A) The degree centrality of each node of the network, i.e. a brain ROI, is obtained by calculating a normalized sum of surviving links to other nodes. In principle, this provides a measure of the importance of a region’s connections to other regions for prediction. (B) At the meso-level, community detection algorithms are used to divide the full network into modular sub-networks that are highly connected to each other. These communities correspond to brain patterns that together are highly important for prediction of chronic cannabis use. (C) At the network-level, global efficiency of the network is calculated by determining the inverse average shortest path. For each node, the distance to every other node is calculated and averaged. The process is repeated for every node and averaged across nodes. The inverse of this averaged shortest path length is the efficiency of the network. High efficiency networks exchange information well because they are densely connected, and thus have fairly low average path lengths.

First, we took the absolute values of the participant-level weighted connectivity matrices (i.e., the correlation magnitudes), which represent the strength of the functional connections between regions. Then we binarized these values at 2% density, to remove spurious correlations and improve the stability and modularity of network features (32–35). These binarized matrices were treated as adjacency matrices, where nodes were the regions and the edges were the importances of the connectivity between each pair of regions to group label prediction. From these matrices, we calculated the degree centrality of each brain region, (Fig. 3A), which is the fraction of regions out of all regions to which it is connected after thresholding. Regions with high “predictive degree centrality” across participants had a high number of predictively valuable correlations with other regions. The regions were ranked by their median degree centrality score.

We then identified network communities with high classification importance (Fig. 3B). First, a group-average weighted connectivity matrix was calculated by taking the mean of the un-thresholded participant-level weighted connectivity matrices. This matrix was then thresholded at 2% density, generating the group-average graph. Then, a hierarchical community-detection algorithm was used to detect clusters of brain regions (“communities”) that were highly interconnected but sparsely connected to other regions (see Details about predictive network analysis in the supplement). The group-average weighted correlation matrix was reordered based on identified community structure to reveal modular clusters. The communities were then ranked by their mean degree centrality score, with those having the highest ranks defined as the most predictively important.

Community predictive importance was corroborated by estimating community derived accuracies using only the weighted connectivity values from those communities, with the following stepwise approach. Starting with the highest average degree centrality ranked community, the non-redundant correlations of the regions within that community to all other regions were used to generate each participant’s signed distance to the decision hyperplane (with the whole-brain intercept). To turn these distances into predictions, the linear classification boundary was varied from the 1st to 100th percentile to find the threshold value that produced the highest training accuracy. In each subsequent step, the correlations from the next highest ranked community were added in, and the accuracy was re-calculated. The smallest number of communities needed to reach a training accuracy of 100% was selected as the minimal community subset and then tested in the held-out data, using the community-specific classification boundary identified in the training set.

Classification confidence and craving analysis

In linear classification, the magnitude of the hyperplane distance (see Details about linear classification in the supplement) can be considered the model’s “confidence” in its prediction: the further it is from the hyperplane, the more likely it belongs to the predicted group (Fig. 4). We tested whether cue-induced cannabis craving was related to the model’s confidence in its classifications of cannabis users (see Avoiding circularity in the craving analysis in the supplement). At the participant-level, we modeled the relationship between average confidences and average cannabis craving ratings with Ordinary Least Squares (OLS). We also modeled the run-by-run relationship with generalized estimating equations (GEE). GEE accommodates correlated observations; we used a Gaussian link function and an independence working correlation structure.

Fig. 4. Relationship between classification model confidence and craving.

The toy example above outlines the workflow for demonstrating a relationship between the chronic use classification model and the momentary craving ratings during the task. (A) The support vector machine classifier discovers the hyperplane for optimal separation between the two classes (here: chronic users vs. non-users). In the example above, the colored points are chronic users and the gray points are non-users. The distance between any point and the hyperplane serves as a proxy for the “confidence” of the model in classifying that point. The sign of this distance corresponds to the prediction of the model, i.e. positive distances are classified as chronic users, and negative signs are classified as non-users. The green point has a high positive confidence, and thus the model is very confident in its prediction for the green point as a chronic user. The yellow point is close to the hyperplane with a positive distance, and is classified as a chronic user with low confidence. Note that chronic-using individuals can be misclassified as non-users, e.g. the red point, which will have a low and negative confidence score. (B) Each individual’s average craving rating is calculated across the duration of the task. A generalized linear model is fit to the confidence scores derived from the SVM classifier and the craving ratings. The plot shown here demonstrates a simple association between craving and confidence, but the full model accounts for repeated task runs within participants as well as possible confounds. Note that the x-axis represents signed confidence scores; the blue decision hyperplane in (a) is represented as a confidence score of 0 in the second plot. Overall, subjects with high craving are more likely to have high confidence scores in the SVM classification.

To rule out alternative explanations, control models with other variables were also fitted to predict classification confidence. These variables were evaluated for whether (1) their inclusion as covariates changed the relationship between cannabis cue-induced craving and confidence, or (2), when modeled alone, they had statistically significant effects. These variables were: sex, years of education, days smoking cigarettes in the last 90 days, days drinking alcohol drinking in the last 90 days (which all significantly differed between groups), control cue-induced cannabis craving ratings, pre- and post-scan cannabis craving, cannabis withdrawal, general cannabis problems (see Sample characteristics) and dataset membership

RESULTS

Classifier training performance

We trained decoding models using four different linear machine learning algorithms (L1- and L2-regularized logistic regression [LR] and linear support vector machine [SVM]) to predict chronic cannabis use from whole-brain functional connectivity (Fig. 5). Relatively weak regularization was better in model training (best alphas: L1 LR α=0.001, L2 LR α=0.05, L1 SVM α=0.001, L2 SVM α=0.05), suggesting widespread functional connectivity was important to classifying CC and HC.

Fig. 5. Performance metrics for linear classification algorithms.

Run-level performance in the training set was measured with 5-fold cross-validation of the training set (232 subjects, 2 runs per subject). All four models performed well in cross-validation metrics with the mean receiver operating characteristic (ROC) curve well above chance (red dotted line). (A) The logistic regression algorithm returns class probabilities which can be directly mapped to the ROC. (B) The linear support vector classification algorithm returns only a decision function, corresponding to the signed distances to the hyperplane. These distances are converted to probabilities using Platt’s method in order to map to the ROC.

Classifier testing performance

Next, we evaluated the performance of our four models on the held-out test data (see Table 2). Accuracies for all four models are much higher than the chance levels from simply predicting the dominant class (~55%) and from 1000 training label permutations (~51%, for all models). These models also outperformed models trained on regional activity estimates (see Comparing functional connectivity to activity for classification in the supplement), suggesting that the interactions between brain regions contain more information about chronic cannabis use than activity magnitudes. For the functional connectivity-based models, while all performed similarly (see Classification model similarity in the supplement), the L1 SVM classifier was the best in terms of both participant-level accuracy (82.7%) and AUC (0.892), with both above chance for both datasets (accuracy/AUC - 2016 dataset: [n=34, 88.2% / 0.867], 2009 dataset: [n=24, 75% / 0.738]). Thus, this model was used for subsequent interpretation.

Table 2. Out-of-sample performance metrics.

Out-of-sample (OOS) performance metrics are shown for run-level and participant-level. After hyperparameter optimization, each model was trained on the full training set (232 participants). Precision and recall metrics are displayed for the poorer performing class, as the lower limit. There are 58 participants in the testing set. The L1 SVM model has the best OSS performance across the four metrics.

	L1 LR		L2 LR		L1 SVM		L2 SVM
	Run	Pt	Run	Pt	Run	Pt	Run	Pt
Accuracy	75.0%	74.1%	79.3%	81.0%	78.4%	82.7%	77.5%	81.0%
AUC	0.812	0.853	0.851	0.879	0.87	0.892	0.834	0.852
Precision	0.717	0.72	0.733	0.759	0.721	0.75	0.771	0.8
Recall	0.765	0.69	0.750	0.781	0.734	0.750	0.692	0.731

Predictive connectivity

Our next goal was to infer the specific brain connectivity implicated in classifying chronic cannabis use. We defined a brain region’s predictive degree centrality as the number of predictively valuable connections a region has with other regions. The top twenty regions of highest predictive degree centrality (along with the lowest two for comparison) are shown in Fig. 6. This ranking was validated with an alternative method and a comparison to a meta-analytic map (see Mean weighted connectivity validates degree centrality rankings and Comparison of network analysis to craving meta-analytic map in the supplement, respectively). Regions from numerous resting state networks are represented in the top twenty regions, indicating widespread connectivity is important for distinguishing individuals with chronic cannabis use from controls.

Fig. 6. Subject-level degree centrality.

Degree centrality represents the normalized number of weighted connections for each brain region that survive thresholding. In other words, it provides a measure of the level of distributed connectivity displayed by a brain region. In the plot above, degree centrality is calculated for each region independently, for each subject. The top twenty regions of highest mean degree centrality are shown, in addition to the lowest two for comparison. Regions identified as having high degree centrality across participants include middle frontal gyrus, bilateral ACC, mid/superior temporal cortex, and bilateral medial PFC.

Finally, at the community scale, we used community detection algorithms to discover modular sub-networks within the weighted connectivity matrices (see for details). Fig. 7 shows the thresholded group-average weighted connectivity matrix reorganized by the discovered community structure. Each community was then ranked by its predictive degree centrality score. The top six communities ranked by within-community degree centrality are visualized in Fig. 7 and community membership of brain regions is displayed in Table 3.

Fig. 7. Predictive communities in group-average weighted connectivity matrix.

The Girvan-Newman community detection algorithm was applied to the group-average weighted connectivity matrix. Girvan-Newman segregates communities within a group by iteratively removing edges with the highest betweenness centrality until a target modularity score is reached. Each disconnected set of nodes is then characterized as a community. The group-averaged thresholded weighted connectivity is sorted by community assignment. Each colored square represents one of the top six communities by average degree centrality within community. The color-corresponding communities are projected onto the brain and colored by resting-state network assignment as determined by the Stanford functional parcellation. See Table 3 for a full description of community membership.

Table 3. Community membership.

For each community identified, the brain regions within that community are listed along with their mean degree centrality (DC) across participants. Communities are ordered by highest average DC within-community.

Community 1	Mean DC	Community 5	Mean DC
R inferior parietal/angular gyrus	0.0474	L ventral precuneus	0.0166
L inferior parietal/angular gyrus	0.0663	L fusiform gyrus	0.0287
R inferior frontal gyrus	0.0145	R superior frontal gyrus	0.0220
Bilateral calcarine cortex	0.0175	R precentral/fronto-opercular	0.0305
R precuneus	0.0241	L mid-occipital cortex	0.0104
R inferior temporal cortex	0.0624	R ventral precuneus	0.0418
L mid occipital cortex	0.0271	L cerebellum	0.0110
R middle frontal gyrus	0.0707	Community 6	Mean DC
R inferior cerebellum	0.0310	R superior/inferior parietal cortex	0.0067
R middle frontal/dlPFC	0.0243	L superior temporal gyrus	0.0437
Bilateral medial posterior precuneus	0.0426	Bilateral medial precuneus	0.0166
R lateral angular gyrus	0.0436	Community 7	Mean DC
R mid occipital cortex	0.0301	L inferior frontal (triangular)	0.0059
Community 2	Mean DC	Bilateral middle cingulate	0.0130
L superior frontal gyrus	0.0155	L inferior frontal gyrus	0.0122
Bilateral ACC	0.0574	L middle frontal/dlPFC	0.0258
R cerebellum	0.0087	L inferior temporal gyrus	0.0175
L inferior temporal gyrus	0.0397	L parahippocampal gyrus	0.0113
Community 3	Mean DC	R parahippocampal gyrus	0.0160
R superior temporal gyrus	0.0624	Bilateral mPFC	0.0488
L medial angular gyrus	0.0247	Community 8	Mean DC
R frontal gyrus	0.0162	L inferior parietal cortex	0.0151
L lateral angular gyrus	0.0306	Community 9	Mean DC
R posterior cingulate	0.0269	R VPN of thalamus	0.0129
R mid-temporal cortex	0.0498	L VPN of thalamus	0.0310
Community 4	Mean DC	L LGN of thalamus	0.0170
L middle thalamus	0.0209	R middle thalamus	0.0084
L anterior cerebellum	0.0129	L supramarginal cortex	0.0052
R anterior insula	0.0259	Community 10	Mean DC
L pre/post-central gyri	0.0393	Bilateral mid-posterior cingulate	0.0172
R pre/post-central gyri	0.0430	Bilateral posterior cingulate	0.012
L superior temporal/auditory	0.0232	Community 11	Mean DC
L prefrontal cortex	0.0138	L mid-posterior temporal cortex	0.0122
R superior temporal/auditory	0.0518	L anterior insula	0.0222
Bilateral anterior thalamus	0.0285	L middle frontal gyrus	0.0177
Bilateral SMA	0.0177	L crus cerebellum	0.0082
Cerebellar vermis	0.0213	R crus cerebellum	0.0093

To confirm that this ranking reflected the predictive importance of the communities, we performed a stepwise prediction analysis to determine the minimum number of communities necessary to reach 100% training accuracy. The first eleven communities (65/90 total regions) were sufficient to reach this target. These same communities were then tested in the testing data, with the same threshold: testing set participant-level accuracy was 84.5%, exceeding our full model testing accuracy of 82.7%.

Relationship between classification confidence and craving

If this model is meaningful mechanistically, we would also expect to see a relationship between model parameters and processes central to chronic cannabis use, such as craving. As expected, the cannabis group reported much higher average cannabis craving, after cannabis (CC=4.82, HC=0.10) and neutral cues (CC=2.61, HC=0.01). We used these craving ratings to test a hypothesis about the relationship of the hyperplane distances (i.e., the model’s confidence in its prediction of CC or HC) to cannabis craving. Our prediction was that cannabis cue-elicited craving ratings (in CC participants only) would positively correlate with hyperplane distances, for the full model (whole-brain) and for the subset of communities that produce the best classification (the top 11). In other words, we predicted that the higher the cannabis cue-related craving, the more confident the model would be that the participant is a CC (confidence in CC classification indicated by the signed distance from the hyperplane). This prediction was supported: at the participant level, there was significant effect of mean cannabis cue-induced craving on the mean confidence scores in both the whole-brain (OLS; t=2.216, P=0.028, 95% CI=[0.325 – 5.623], adj. R²=0.023) and the 11 top communities (t=2.483, P=0.014, 95% CI=[0.658–5.771], adj. R²=0.030). The effect was very similar on the run-by-run level (GEE; whole-brain: z=2.192, P=0.028, 95% CI=[0.304–5.443]; 11 communities: z=2.455, P=0.014, 95% CI=[0.629–5.609]).

To test for the specificity of the craving and confidence relationship, we compared it to alternative models that included covariates for demographic and lifestyle variables that differed between the groups (sex, years of education, cigarette and alcohol use over the previous 90 days), as well a covariate for the dataset. Including these covariates did not substantially change the craving and confidence relationship (i.e., the craving predictor was significant at P<0.05 in all models). Further, none of these variables reached statistical significance when modeled on their own.

This confidence effect was also specific to craving during cannabis cues: it was not related to cannabis craving during neutral cues, or to general cannabis craving before or after the MRI scan (21). The effect also did not relate to other measures of cannabis use or dysfunction, including withdrawal symptoms prior to the scan (22), general problems with cannabis problems (23), current or lifetime symptoms of maladaptive cannabis use or the number of days of cannabis use over the previous 3 months (24). Cigarette and alcohol use over the previous 3 months were also unrelated to cannabis classification confidence, both in cannabis users alone and in both groups when a group label regressor was included.

DISCUSSION

Functional connectivity predicts chronic cannabis use

In this study, we developed a novel modeling approach that balances accurate prediction and model interpretability in chronic cannabis use. In the largest fMRI sample of people with long-term cannabis use (CC) and healthy controls (HC) to date, we classified chronic use from functional connectivity during a cue-elicited craving task with high out-of-sample accuracy (82.7% participant-level, 78.4% run-level). This accuracy compares favorably to previous fMRI studies using functional connectivity to classify drug use, in both nicotine smoking (36–38) and cocaine use disorder (39–41), even though most studies did not test out-of-sample and had much smaller sample sizes (both of which can inflate prediction performance). Furthermore, this is one of the first fMRI studies (42), and the largest to date, to classify chronic cannabis use. Compared to other large studies on cannabis use (e.g., using ENIGMA, ABCD & UK Biobank datasets), our study identifies functional rather than anatomical differences (43,44), focuses on evoked neural responses rather than resting state fluctuations (45) and investigates the phenomenon of cannabis craving.

Distributed patterns of predictive connectivity

Our functional connectivity-based models outperformed models trained on regional activity estimates. This suggests there is more information about chronic cannabis use in the interactions between regions than in their isolated activities, corroborating a recent review of decoding studies in psychiatric research (16). Given this, our next goal was to use network analysis to discover brain network patterns that differentiated the groups, starting with the individual regions that were most critical to successful prediction.

We found high predictive degree centrality in several sensory and motor related regions, including right inferior temporal cortex and left inferior temporal gyrus, both areas along the ventral visual pathway. Given that the visual and tactile demands of the task were the same across groups, these regions likely reflect more than the passive reception of sensory information and output of motor commands. For example, these regions may facilitate the recognition of drug cues and retrieval of behavioral associations, such as the initiation of drug seeking/use behaviors (46). Regions related to attention and its control also ranked highly on this measure, likely reflecting differential recruitment of attention during cue processing between the groups. These included the right middle frontal gyrus (the area with the highest predictive degree centrality) an important site of convergence for the dorsal and ventral attention networks (47) and the bilateral ACC, an area that features dense cannabinoid receptors (48) and responds abnormally during drug cue exposure and craving generally (49,50) and in cannabis users specifically (51,52). High predictive degree centrality was also detected in regions associated with cue-reactivity and craving, including several areas in the precuneus, which may help process drug cue salience and relevance to the self (53) and in the bilateral medial PFC, which has extensive and recurrent dopaminergic connections with the ventral tegmental area and may direct drug-seeking behavior (54).

We also discovered sets of brain regions (i.e., communities) that were important to successful classification of CC and HC. The eleven communities with the highest average degree centrality produced excellent testing set prediction accuracy, greater than the accuracy using the whole-brain connectivity. These communities do not obviously map onto canonical networks (e.g., salience, default mode, frontoparietal networks), suggesting that they may reflect task-specific network organization that differs from resting-state networks.

Cue-induced craving correlates with classification confidence

Our final goal was to study the effect of cue-induced cannabis craving, a core feature of cannabis addiction (19,20). However, non-craving-related differences between CC and HC, such as cigarette or alcohol use, may enable high classifier accuracy without any craving-related neural connectivity being engaged. To explicitly test the hypothesis that craving-related functional connectivity was used by our model to classify successfully, we related model classification confidence scores to the cannabis users’ craving ratings in response to drug cues. The classification confidence scores from both the whole-brain and top eleven communities models were correlated with craving during cannabis cues in chronic users but not with neutral cue elicited craving, baseline withdrawal or craving on the day of the scan, general cannabis use or symptoms, cigarette or alcohol use, or demographic variables. These results demonstrate that the model learned latent information about cue-elicited craving despite only being trained to classify chronic use. In other words, the neural signature of chronic cannabis use contains information about craving, suggesting drug-related decision-making and craving are deeply intertwined and may be instantiated in similar circuitry—a strong hypothesis that should be tested explicitly in future experiments.

Limitations and next steps

There are several limitations to this study. First, the classification accuracies are not high enough for direct clinical use, suggesting the need for larger samples. This study also precludes inferences about the specificity of the effects of cannabis use: future work should compare cannabis users to chronic users of other drugs and non-drug using individuals with other psychiatric dysfunction, in order to establish cannabis-specific neural signatures. We also don’t predict anything about future cannabis use, which is an important clinical goal and requires large-scale longitudinal studies.

Conclusions

In conclusion, this study is a first step towards building accurate and interpretable decoding models that have both scientific and clinical significance. The models performed well in out-of-sample data, indicating their generalizability. Furthermore, we interpreted the best-performing model to corroborate prior findings, implicating regions related to sensory and motor processes (right inferior temporal cortex and left inferior temporal gyrus), attention (right middle frontal gyrus and bilateral ACC) and cue-reactivity and craving (precuneus and bilateral mPFC). We also find potentially novel network-level properties in the context of cannabis use and craving. Finally, we discover a relationship between the predictive signature associated with chronic cannabis use and subjective cannabis cue-induced craving. Future work can build on this approach of using joint predictive-explanatory models to constrain neurobiological inferences.

DATA AND CODE AVAILABILITY

All the code related to analyses in this study is publicly available at https://github.com/kulkarnik/mj_classifier. The deidentified, parcellated data used for classification are available in the same repository.