Connectome-based prediction of craving in gambling disorder and cocaine use disorder

Abstract

Introduction

Craving, involving intense and urgent desires to engage in specific behaviours, is a feature of addictions. Multiple studies implicate regions of salience/limbic networks and basal ganglia, fronto-parietal, medial frontal regions in craving in addictions. However, prior studies have not identified common neural networks that reliably predict craving across substance and behavioural addictions.

Methods

Functional magnetic resonance imaging during an audiovisual cue-reactivity task and connectome-based predictive modelling (CPM), a data-driven method for generating brain-behavioural models, were used to study individuals with cocaine-use disorder and gambling disorder. Functions of nodes and networks relevant to craving were identified and interpreted based on meta-analytic data.

Results

Craving was predicted by neural connectivity across disorders. The highest degree nodes were mostly located in the prefrontal cortex. Overall, the prediction model included complex networks including motor/sensory, fronto-parietal, and default-mode networks. The decoding revealed high functional associations with components of memory, valence ratings, physiological responses, and finger movement/motor imagery.

Conclusions

Craving could be predicted across substance and behavioural addictions. The model may reflect general neural mechanisms of craving despite specificities of individual disorders. Prefrontal regions associated with working memory and autobiographical memory seem important in predicting craving. For further validation, the model should be tested in diverse samples and contexts.

Keywords:

• Craving
• cue-reactivity
• addictive behaviours
• cocaine
• gambling
• machine learning

Introduction

Cravings, or strong urges to engage in specific behaviours, are a feature of substance-use disorders and disorders due to addictive behaviours and have been directly linked to other aspects of addictions (Tiffany and Wray 2012; Starcke et al. 2018). The relevance of craving during the course of addictions has been demonstrated in various studies showing associations between craving and diminished control over behavioural engagement (Serre et al. 2015; Hawker et al. 2021), treatment outcomes and relapses (Hawker et al. 2021; Paliwal et al. 2008; Vafaie and Kober 2022), and neurobiological measures (Antons et al. 2020; Zilverstand et al. 2018).

In laboratory settings, craving is often induced by visual cues (e.g., images or videos) that are related to the addictive behaviour (Starcke et al. 2018; Carter and Tiffany 1999; Jasinska et al. 2014). This method is based on theories assuming that conditioned learning processes and incentive sensitisation (Berridge and Robinson 2016; Robinson and Berridge 2008) lead to cue-reactivity, a physiological, emotional, and cognitive response to behaviour-related stimuli. This not-always-conscious cue-reactivity response can be accompanied by increases in conscious craving (Robinson and Berridge 2008). In addition, multiple mechanisms have been proposed to be related to craving such as the processing of internal and external triggers, desire thinking (imaginal prefiguration and verbal perseveration), craving experiences (urges and associated deficits), inhibitory control, and impulses to execute the behaviour (Brandtner et al. 2021). Considering the complexity of mechanisms involved in craving, it is not surprising that theories and models of addiction propose that the development of cue-reactivity and craving is related to neural changes across the whole brain including the ventral and dorsal striatum, insula, orbitofrontal cortex, anterior cingulate cortex, dorsolateral prefrontal cortex, and inferior frontal gyrus (Zilverstand et al. 2018; Berridge and Robinson 2016; Koob and Volkow 2010; Brand et al. 2019; Hill-Bowen et al. 2021). Various neuroimaging studies report alterations in neural activity during cue-reactivity and craving for both substance-use disorders (e.g., cocaine-use disorder, CD) and disorders due to addictive behaviours (e.g., gambling disorder, GD) (Antons et al. 2020; Zilverstand et al. 2018), and these neural mechanisms may be similar across addictive behaviours (Hill-Bowen et al. 2021; Irizar et al. 2020; Kober et al. 2016; Yip et al. 2017).

Not only functional brain activity, but also functional connectivity between specific brain regions, has been associated with cue-reactivity and craving (Zilverstand et al. 2018; Limbrick-Oldfield et al. 2017; Koehler et al. 2013; Zhang et al. 2018). Most such studies have used seed-based approaches. By definition, seed-based approaches are limited to a priori selected brain regions. Therefore, it is difficult to understand alterations within whole-brain connectivity patterns. Furthermore, traditional correlation or regression approaches hold the risk of overfitting. Connectome-based predictive modelling (CPM; Finn et al. 2015; Shen et al. 2017) is a machine-learning approach which generates brain-behaviour models from whole-brain functional-connectivity data. This method includes a cross-validation which reduces the problem of overfitting and increases the likelihood of replication in future studies by testing the model in a novel sample. The linear approach allows not only the prediction of behaviour based on connectomes, but also the identification of central brain regions (i.e., nodes) and networks that predict the behaviour and thereby contribute to the understanding of neural mechanisms. CPM has been used to predict future abstinence during treatment among individuals with CD and with opioid use disorder (Lichenstein et al. 2021; Yip et al. 2019), as well as craving in gaming disorder (Zhou et al. 2022).

In the present study, CPM was used to re-analyze data from an fMRI study in which craving was triggered by videos in individuals with CD and GD (Kober et al. 2016). We aimed to (1) predict subjective craving across addictions based on functional-connectivity data assessed during a video-cue-reactivity task. If craving could be predicted based on functional-connectivity data across CD and GD, we further aimed to (2) identify central nodes and networks involved in the prediction of craving and to decode their functions based on meta-analytic data.

Materials and methods

Participants

Data from a previously published study (Kober et al. 2016) were re-analyzed. The study was approved by the Yale Human Investigation Committee. Because of potentially confounding effects of excessive head motion on patterns of connectivity (Shen et al. 2017), data from 6 participants (2 CD, 4 GD) with excessive motion (>0.25 mm mean frame-to-frame displacement) were excluded. The final sample included 28 individuals with CD (aged 29–51 years, M = 42.71, SD = 6.13; 11 female) and 24 individuals with GD (aged 19–55 years, M = 32.29, SD = 11.18; 7 female). Diagnoses for CD and GD were made based on the Structured Clinical Interview for DSM-IV for cocaine dependence and the Structured Clinical Interview for Pathological Gambling (SCI-PG; Grant et al. 2004). All participants provided written informed consent after receiving a complete description of the study. As craving is conceptualised as an intense and urgent desire to use a substance or behave in a certain way, developing during the course of addiction, we believe that the subjective experience of craving and its neural manifestation would especially be present in the clinical samples and less in a healthy control sample. For the prediction of craving based on neural connectivity, we therefore decided to focus on the CD and GD samples.

Neuroimaging data acquisition and preprocessing

During fMRI, participants saw six videos in a counterbalanced order. In these videos, a male or female actor was shown describing cocaine use, gambling, or a sad experience. After each run, individuals rated their urges to use cocaine and urges to gamble on a 1 (not at all) to 10 (a lot) scale. To predict the urge to engage in the specific behaviour, only those two videos in which individuals were confronted with their problematic behaviour and the corresponding urge measures were used (i.e., CD: cocaine videos, cravings for cocaine; GD: gambling videos, urges for gambling). Within a previous analysis of some of these data, different epochs of the videos were considered (Kober et al. 2016); for the following CPM analysis, the fMRI data acquired during the two complete videos were used to estimate brain connectivity, thus representing the neural state during cue-reactivity and following other approaches of similar data (Potenza 2008). The mean of the two urge ratings that followed the video on the problematic behaviour were used as a measure for craving.

For preprocessing, functional images were motion-corrected (SPM 8) and smoothed (AFNI’s 3dBlurToFWHM, http://afni.nimh.nih.gov), brains were extracted from structural images (optiBET; Lutkenhoff et al. 2014), and functional and structural images were registered to MNI stereotactic space. Linear registrations (BioImage Suite; Joshi et al. 2011) were used to align functional and structural images. See Kober, Lacadie (Kober et al. 2016) for a more detailed description of the task and neuroimaging data acquisition.

Functional connectivity

The BioImage Suite (Joshi et al. 2011) was used for whole-brain functional connectivity analyses, as previously described (Finn et al. 2015; Yip et al. 2019). Preprocessed functional data were used for analyses. Pearson’s correlation coefficients were calculated as measures of functional connectivity between each node-by-node pair and normalised to z-scores using Fisher’s z-transformation. A 268-node brain parcellation (Shen et al. 2013), which includes the cortex, subcortex, and cerebellum, was used for node definition. This resulted in a 268 × 268 symmetric connectivity matrix for each participant.

Connectome-Based predictive modeling (CPM)

For CPM, a validated custom MATLAB script (Shen et al. 2017) was used. The connectivity matrices from each participant and subjective measures of craving served as input variables for the script. As it has been done in previous works (Yip et al. 2019; Zhou et al. 2022), we used a leave-one-out cross-validation for model training and evaluation. Model-relevant features were selected by correlating with Spearman’s rank correlation each edge in the connectivity matrix with the craving measure over all but one participant (training dataset). Those edges which correlated significantly (p = .05) with the craving measure were selected as features. The modelling distinguished between positive (positive connectivity/correlation coefficient) and negative (negative connectivity/correlation coefficient) networks. For each participant of the training dataset, the sums over the weights of these selected edges were taken as single-subject summary-measures for both positive and negative networks. Then, linear models were fitted for the brain-behaviour relationships. Finally, the linear model was tested with the summary measure based on the selected features of the left-out participant. To determine the level of significance, permutation testing with 1000 iterations was used.

To identify whether results were biased by differences in connectivity strengths between groups, the CPM analysis was repeated while excluding edges whose connectivity significantly differed between groups. For this purpose differences in connectivity between groups were identified using a mass univariate edge-wise approach as implemented in the MATLAB code freely available at https://github.com/YaleMRRC/CPM (Dadashkarimi et al. 2019). Vectors indicating connectivity strengths were created across all participants for each edge based on the previously created 268 × 268 connectivity matrices. A t-test was applied on these vectors to identify differences in connectivity strengths between groups of individuals with CD and GD. A network-based significance statistic was used with 1000 iterations to control the family-wise-error rate (Zalesky et al. 2010). To evaluate direction of effects, mean values were calculated over all correlation coefficients for each group and each identified cluster separately. The edges that showed significant differences between groups in the MANOVA were excluded from the CPM analysis. Results were visualised by using the connectivity viewer of the BioImage Suite web application (https://bioimagesuiteweb.github.io/webapp/connviewer.html).

Identification of important nodes and networks

First, brain regions with high importance within positive and negative networks were identified as those nodes with highest degree (≥15). Based on meta-analytic results of Neurosynth (Yarkoni et al. 2011), functionality of these nodes was identified. The decoding was done by entering the MNI coordinates of the highest degree nodes in the meta-analytic tool and retrieving terms associated with this region. All terms associated with the region as indicated by the Neurosynth meta-analysis (z-score in Neurosynth meta-analysis >0) were selected. In a second step, terms describing functionality (in contrast to anatomy or general terms) were manually identified and represented in word clouds (wordcloud function integrated in MATLAB) where the sizes of the words reflect the z-scores. In case of similar or semantically overlapping terms (e.g., ‘wm’ and ‘working memory’), those overlapping terms were merged into one term with a weight equal to the sum of the merged term’ z-scores.

Second, overlap with ten canonical networks (medial frontal, fronto-parietal, default-mode, motor/sensory, subcortical/basal-ganglia, salience/limbic, visual I, visual II, visual-association, and cerebellum/brainstem; see visual representation of networks in the Online Supplement SF1) was identified (Lichenstein et al. 2021; Yip et al. 2019). To identify the specificity of single canonical networks in predicting craving, the CPM analysis was repeated by excluding single canonical networks. In addition, the predictive strength of single networks was identified by excluding all edges but the ones of the one single network.

Ethics

The study procedures were conducted in accordance with the Declaration of Helsinki. The Yale Human Investigation Committee approved the study. All subjects were informed about the study, and all provided informed consent.

Results

Behavioural results

The whole sample showed a mean craving score after the disorder-specific videos of 4.99 ± 3.26 (min–max: 1–10). After cocaine videos were presented, individuals with CD showed higher craving to use cocaine compared to individuals with GD (M_diff = 2.19, p <.05), while after gambling videos individuals with GD compared to individuals with CD showed higher craving to gamble (M_diff = 3.70, p < .001). As previously reported (Kober et al. 2016), individuals with GD compared to those with CD showed higher mean craving, GD: 6.65 ± 2.73, CD: 3.57 ± 3.03, t(50) = 3.82, p < .001, Cohen’s d = 1.06, CI_d [0.48, 1.65], with both groups showing variances in subjective craving. Again, for the CPM analysis, the cocaine use craving measures were used for individuals with CD and the gambling urge measures were used for individuals with GD.

Predicting craving across groups

The overall model that combines both positive and negative networks significantly predicted craving, r = 0.272, p = .038 (Figure 1(A,B)). Regression coefficients were higher for the negative network (r = 0.376, p = .009) and did not reach significance for the positive network (r = 0.163, p = .143). Overall, 245 edges (0.68% of all possible edges) differed significantly between groups during cue-reactivity (see Online Supplement SF2 for direction of effects and visualisation clusters differing across groups). Overall, only 3.6% (overall: 42 edges, negative network: 30 edges, positive network: 12 edges) of all edges predicting craving across groups also showed significant between-group differences in cue-reactivity networks (see Online Supplement SF3 for detailed description). Results of the CPM analysis remained similar when excluding these edges (r = 0.250, p = .051, negative network: r = 0.343, p = .017, positive network: r = 0.162, p = .142), suggesting that group connectivity differences in cue-reactivity networks did not bias the CPM analysis.

Figure 1. Positive and negative networks predicting craving and model performance. Part A presents positive (red) and negative (turquoise) networks resulting from CPM analysis that predicted craving. Larger spheres represent nodes with higher degrees. Part B shows the association between actual and predicted craving values based on the positive (red) networks, negative (turquoise), and both networks together (gray). Part C shows all nodes predicting craving with a degree of 15 or higher and their corresponding edges separated for positive and negative networks based on macroscale brain networks as implemented in the Shen 268 atlas (Shen et al. 2013). Brain regions are presented in approximate anatomical order (see legend for description of regions). Longer-range connections represent connections with regions more distant than shorter-range connections. The left part of the circle figures represents the right hemisphere and the right part the left hemisphere.

Important nodes and networks predicting craving

Overall, the networks predicting craving were complex and included 1159 edges (positive: 543 edges, negative: 616 edges) which encompassed 3.23% of all possible edges. Nodes and edges predicting craving are summarised in Figure 1(C). The highest degree nodes connected with more than 15 other nodes and were mostly located in the prefrontal cortex. In the positive network, the right dorsolateral prefrontal cortex (Brodmann area 46, MNI: 48.29, 35.68, 15.15) and dorsal posterior cingulate cortex (Brodmann area 31, MNI: −6.5, −53.94, 37.44) showed the highest degrees. In the negative network, the right dorsolateral prefrontal cortex (Brodmann area 46, MNI: 48.29, 35.68, 15.15), right inferior frontal gyrus part opercularis (Brodmann area 44, MNI: 55.35, 9.62, 22.22) and triangularis (Brodmann area 45, MNI: 36.98, 20.81, 5.89), left orbitofrontal cortex (Brodmann area 11, MNI: −5.42, 29.14, −10.12), left anterior prefrontal cortex (Brodmann area 10, MNI: −6.93, 48.31, −5.71), left frontal eye field (Brodmann area 8, MNI: −39.35, 17.2, 46.7), and left and right fusiform gyrus (Brodmann area 37, left MNI: −46.67, −39.97, −24.3; right MNI: 46.47, −59.85, −14.76;) were the highest degree nodes. Results of the Neurosynth functional decoding analysis are presented in Figure 2(A).

Figure 2. Identification of important nodes and networks. Part A presents the results of the decoding analysis with Neurosynth. Results from important nodes of the positive network are presented in red, results from the negative network in blue. The size of the words corresponds to the z-scores in Neurosynth. Part B shows the overlap of the positive (red) and negative (blue) networks with canonical networks. DLPFC: dorsolateral prefrontal cortex; IFG: inferior frontal gyrus; OFC: orbitofrontal cortex; PFC: prefrontal cortex.

Overlap of positive and negative networks with canonical network definitions are presented in Figure 2(B). When iteratively excluding each single canonical network from the CPM analysis within the computational lesion prediction, results remained nearly the same. As in the primary CPM analysis, especially when using the negative network mask, craving was associated with edge strength, although single networks were excluded, indicating that any single network was particularly specific for the prediction of craving (see Online Supplement ST1). Nonetheless, when predicting craving based on single networks, the default-mode, fronto-parietal, motor/sensory, salience/limbic, visual I, and visual-association networks could significantly predict craving when using the negative network mask. None of the networks alone could predict craving when using the positive network mask.

Discussion

In the current study, we (1) demonstrated that the level of subjectively perceived craving could be predicted across addictive disorders based on whole-brain connectivity measures. Of note, increased segregation (negative connectivity, asynchronous activity) between networks was predictive of craving. Accordingly, (2) the most important nodes and specific networks predictive of craving could be identified. These nodes were located within the prefrontal cortex, cingulate cortex and fusiform gyrus. Their functionality is associated with working memory, autobiographical memory, valence ratings, physiological responses, and finger movements/motor imagery. The craving networks were shown to be complex and include regions involved in complex cognitions, as has been reported in previous studies and theories (Zilverstand et al. 2018; Berridge and Robinson 2016; Koob and Volkow 2010; Brand et al. 2019; Hill-Bowen et al. 2021).

Using CD and GD as exemplary types of substance and behavioural addictions, respectively, neural networks of craving that generalise across disorders could be identified. The results remained similar when excluding edges differing between groups from the prediction model, suggesting robustness of results across disorders. As expected, craving was associated with complex whole-brain positive and negative networks, and particularly the latter. Specifically, the overlap with canonical networks (Figure 2(B)) indicate that a high segregation between the motor/sensory network and the default-mode network as well as between the motor-sensory network and the salience/limbic network was associated with craving. In contrast to these results, a recent study by Zhou, Wang (Zhou et al. 2022) using the same method to predict craving in individuals with gaming disorder, another disorder due to addictive behaviours, found higher integration, especially within the default-mode network and between fronto-parietal and subcortical networks as well as a motor-sensory network, to predict craving. As CPM is a data-driven method, these contrasting results may be explained by specific characteristics of craving within one disorder or other factors such as tasks used or participants’ characteristics. The connectivity model in the current study, which included data from individuals with CD and GD, may generalise across addictions and increase understanding of general mechanisms of craving without specificities of single disorders, while the model by Zhou, Wang (Zhou et al. 2022) may be more specific for the prediction of craving in individuals with gaming disorder, although these and other possibilities warrant additional examination.

Interpreting the nodes identified as being most important with regard to their associated functionality, components of memory (working memory and autobiographical memory) seem to be highly relevant for the processing of craving. The relevance of learning and memory in addictions has been discussed previously (e.g., Goodman and Packard 2016). In the context of cue-reactivity and craving, it has been argued that information related to the addictive behaviour and behaviour-related cues are strongly encoded due to conditioning processes and thereby contribute to cue-reactivity processes in addiction (Ekhtiari et al. 2016). In addition, memories may contribute to the saliency processing of cues and associated feelings of craving. Accordingly, during cue-reactivity, working memory could be important for the integration of cues into consolidated memories, and the retrieval of information may activate a wide associative network, especially for individuals with stronger cravings.

Within craving theories, conditioning and cognitive theories have been described (Ekhtiari et al. 2016; Drummond 2001). While the interpretations in the previous paragraph are more based on conditioning theories, the current results could also be interpreted with regard to cognitive theories. Although the videos used in the current study did not specifically refer to autobiographical memories, self-directed cognitions may be an integral part of cravings. Autobiographical scripts have been effectively used to induce craving and psychophysiological responses (e.g., systolic blood pressure, heart rate) in individuals with opioid, alcohol, and cocaine use disorders (Weinstein and Cox 2006; Potenza et al. 2012). In some theoretical models, interactions between automatic craving responses and more active and conscious processes of desire thinking (mental imagery and verbal perseverance) are considered (Brandtner et al. 2021; Kavanagh et al. 2005; Spada et al. 2015). In particular, desire thinking may require individuals to activate brain regions involved in the processing of autobiographical memories. For example, behaviour-related expectancies have been conceptualised as information in long-term memory including autobiographical information about prior engagement in specific behaviours (Goldman 1999). While working memory may be used to actively retrieve information from autobiographical (episodic) memory, the content of autobiographical memories may serve to form expectancies about behavioural engagement. This active desire thinking based on memories may then increase craving responses (Brandtner et al. 2021). Related to mental imagery, ‘motor imagery’ was one term identified in the decoding analysis (left inferior frontal gyrus pars triangularis).

Some theories of craving also include a physiological component of craving (Drummond 2001; Flaudias et al. 2019; Geerlings and Lesch 1999; Weinstein et al. 1998). The central role of the right orbitofrontal cortex that is associated with the terms ‘skin conductance’ and ‘heart rate’ may represent this physiological component. The involvement of working and autobiographical memories may be used to interpret such physiological responses and in cognitive labelling (Drummond 2001). In addition, the orbitofrontal cortex has been associated with processes of valuation in a meta-analysis of cue-reactivity (Hill-Bowen et al. 2021).

The high importance of prefrontal regions within the CPM-derived model may further indicate how closely potential cognitive and control-related processes are associated with subjective craving. The central role of these nodes (dorsolateral prefrontal cortex, inferior frontal gyrus) in the negative network may indicate that these regions are more segregated from other regions and may explain the poorer inhibitory control that has been associated with increased craving in addictions (Bernard et al. 2021). In addition, these nodes might be highly relevant in the CPM-derived model because of the higher need to regulate craving responses. Consistent with this interpretation, non-invasive stimulation of the prefrontal cortex, in particular the dorsolateral prefrontal cortex, has been used to boost inhibitory control and to reduce craving (Perrotta and Perri 2022; Brainswitch Study Group, 2022). However, results seem inconsistent, variably showing reductions in craving and consumption, only in consumption without effecting craving or only in craving. Accordingly, the involvement of the prefrontal cortex in the processing of craving may be particularly complex. For example, cognitive structures may be used to direct cognitive resources towards desires to engage in addictive behaviours (desire thinking; Brandtner et al. 2021). This hypothesis may be supported by the involvement of the posterior cingulate cortex in predicting craving (here identified as one of the highest-degree nodes) that has been previously related to situations in which individuals are caught up in experiences (Brewer et al. 2013). The posterior cingulate cortex has also been identified as a region that is specifically linked to drug-cue-reactivity and not to cue-reactivity related to natural rewards (Hill-Bowen et al. 2021). Thus, it has been proposed that the posterior cingulate cortex is relevant in reward-guided decision-making and ruminations about substance use. Consistent with this interpretation, the reduced connectivity between the motor/sensory-hub-network and the default-mode network may amplify situations of being caught up in experiences of craving due to reduced cognitive inhibition (c.f. Zhang and Volkow, 2019). Overall, it should be considered that machine learning is a data-driven method with a focus on the prediction over explanation (Yarkoni and Westfall 2017). Although it can be used to interpret results and contribute to theories (Yip et al. 2019), these interpretations should be systematically tested in future studies.

The current results should be seen as preliminary and should be validated in future studies. CPM may be understood as a methodology that complements traditional methods by providing the opportunity to both predict behaviour and to generate information that contributes to the development of theoretical models (Yip et al. 2019). One limitation of the current study was the relatively small sample size. Future studies should use larger samples and employ more robust repeated random splits methodology for cross-validation (Varoquaux et al. 2017). It would be informative to test to which degree the identified craving network could predict relapse or abstinence. The generalisation and validation of the craving networks could be limited by different diagnostic criteria applied to the sample described in the current sample versus those that could be considered in future studies (e.g., those based on DSM IV vs. DSM-5 vs. ICD-11 criteria). It should be investigated whether the current craving network could also predict craving measured in the natural environment and how craving and its neural networks relate to actual behaviours. Although the current sample included males and females with two types of addictions, it should be encouraged to test the model in more diverse and larger samples. In addition, the craving network should be investigated in healthy participants and should be compared to other motivational or affective states than craving. Such studies could investigate the specificity of the network. Within the current study, craving has been measured with one item visual analogue scales. In accordance with theories suggesting multiple types of craving (e.g., reward, relief, obsessive craving; Brandtner et al. 2021; Flaudias et al. 2019; Verheul et al. 1999), more comprehensive and validated questionnaires have been generated and used to measure craving (e.g., Paliwal et al. 2008; May et al., 2014; Ashrafioun and Rosenberg 2012). Although these more complex measures might be less sensitive to short-term changes in craving, future studies should focus on the neural networks of specific types of craving. When validated and replicated, the neural model of craving may be important in identifying networks that could be targeted via neuromodulation (e.g., with repetitive transcranial magnetic stimulation) to reduce cravings by increasing cognitive control and awareness of craving responses (Wu et al. 2020, 2021). This would constitute a next step in translating the current results into clinical and public health advances.

In conclusion, functional neuroimaging data during cue-reactivity in combination with CPM is a promising method to predict addiction-relevant measures such as craving across substance and behavioural addictions. The identified neural model may reflect general mechanisms of craving across individual disorders. Prefrontal regions associated with working memory and autobiographical memory seem important in predicting craving. For additional validation and generalisability, the model should be tested further in diverse samples and contexts.

Disclosure statement

Dr. Potenza has consulted for and advised Game Day Data, the Addiction Policy Forum, AXA, BariaTek, Idorsia and Opiant Therapeutics; been involved in a patent application involving Novartis and Yale University; received research support from the Mohegan Sun Casino and the Connecticut Council on Problem Gambling; and consulted for law offices, the federal public defender’s office and gambling entities on issues related to impulse-control and addictive disorders. The other authors report no financial relationships with commercial interests.

Additional information

Funding

NIDA [grants K01DA039299, R01 DA019039, R01DA039136] provided support for data collection as well as Drs. Yip and Potenza and Ms. Lacadie. Dr. Antons received a scholarship from the Gustav A. Lienert Foundation for the research project. Drs. Antons and Brand work on this article was conducted in the context of the Research Unit ACSID, FOR2974, funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) – [411232260]. This work was also supported by National Institute on Drug Abuse.