Data overview
A total of 479 contrast images from 101 participants and 5 independent cohorts were used for training and testing the pattern to predict drug and food craving (two drug using cohorts, two of their matched controls, and another sample of drug users with no matched controls). All participants viewed images of drugs and food under two instructions conditions: a craving instruction and an instruction to use a cognitive strategy to reduce craving (see methods; [74-76]. Contrast images were computed for the onset of the visual drug and food cues (see Figure 1a), separately for each level of craving (1-5 Likert scale) for every participant.
fMRI results
Description of the Neurophysiological Craving Signature (NCS). Parallel to previous studies on fMRI-based prediction of pain and emotion [36, 39, 77, 78], LASSO-PCR and study-stratified 10-fold cross-validation was used to predict the level of craving based on fMRI contrast images. Figure 2 presents a thresholded display of the resulting weight map based on bootstrapping. While the unthresholded map is used for prediction, the thresholded map illustrates the brain areas that most robustly contribute positive or negative weights to the predictive pattern. Areas with positive weights included ventromedial prefrontal cortex, dorsal anterior cingulate cortex, subgenual cingulate/ventral striatum, retrosplenial cortex, parietal and temporal areas, cerebellum, and amygdala. Negative weights were found in visual areas, lateral prefrontal, parietal and somatomotor areas, among others (see Table 1 for a list of FDR-corrected coordinates). Of note, many areas, including somatomotor cortex, parietal, temporal, and bilateral insula included clusters of both positive and negative weights.
Predictive performance of the NCS. The trained pattern resulted in a cross-validated prediction-outcome correlation of r = 0.50 (S.D. +-0.49) within-person and r = 0.33 across all data points, with a mean absolute error of 1.30 points on the 1-5 Likert scale. A multi-level general linear model (GLM) confirmed a strong relationship between out-of-sample predicted and actual level of craving with a large effect size (beta = 0.35, STE = 0.04, t(100) = 8.43, p < 0.00001, Cohen’s d = 0.84). The strength of the predictive performance varied across datasets, but was significant in all 5 cohorts, with effect sizes (Cohen’s d) ranging from 0.59-1.38 (see Table 2).
Classification of high versus low craving. We next assessed the accuracy of this pattern to differentiate between high versus low levels of craving. Forced-choice binary classification of highest versus lowest levels of craving per participant was achieved with a cross-validated accuracy of 78% +-4.1%STE, binomial test p < 0.00001 (sensitivity = 78%, specificity = 78%, area under the curve [AUC] = 0.89). Even across subjects (single-interval classification), this pattern separated brain responses to the highest versus lowest individual levels of craving with 72% cross-validated accuracy (+-3.3%STE, binomial test p < 0.00001, sensitivity = 65%, specificity = 78%, AUC = 0.75). While this level of predictive accuracy does not provide perfect separation of high versus low craving, it is remarkable, since all stimuli were drugs or highly palatable food items, thus differences in classification performance were not driven by external stimulus characteristics but by the personal history and internal motivational states of the participants.
Differentiating drug users from non-users. We next tested whether individual craving-pattern responses to drug and food cues could be used to predict whether a participant was a drug user or non-user (see Figure 4a for group averages, Figure 4b for individual effects, and Figure 4c for ROC plots). While pattern expression in brain responses to food cues were not significantly differentiating drug users from non-users (59% accuracy +-4.9%STE, P = 0.92, AUC = 0.41), NCS pattern responses to drug cues significantly classified drug users from non-users, with 75% accuracy (+-4.3%STE, P = 0.000616, sensitivity = 86%, specificity = 60%, AUC = 0.76). When testing a pattern based on the drug>food contrast, the response in the NCS separated drug users from non-users with 77% accuracy (+- 4.2% STE, P = 0. 000113, sensitivity = 86%, specificity = 64%, AUC = 0.83, see Figure 4c).
Given slight but significant differences in years of education between users and non-users, we tested whether differences in NCS response could be explained by years of education. Individual differences in years of education correlated negatively with NCS response to drug cues (r = -0.18, p = 0.079) and to the contrast drug>food cues (r = -0.29, p = 0.004). However, these associations did not mediate the relationship between drug-use status and NCS responses (p = 0.35) and were not significant when controlling for drug use (p = 0.32). Further, NCS responses to food cues did not correlate with years of education (r = 0.05, p = 0.612). These results speak against the interpretation that NCS differences between drug users and non-users are driven by sociodemographic differences such as differences in education, and rather more likely reflect the fact that drug use (especially illegal drug use) reduces educational achievements [79].
Drug and food craving are predicted by shared brain patterns. An important debate concerns the question whether drug and food craving are based on similar brain processes [67, 68, 70, 73]. If drug and food craving are driven by shared brain processes, then drug craving should be predictable based on a pattern that is trained to predict food craving, and food craving should be predictable based on a pattern that is trained to predict drug craving – at least in drug users. Conversely, if drug and food craving are based on dissociable brain processes, then better predictive accuracy will be gained by training drug- and food-specific (compared to craving-general) brain patterns.
We therefore repeated the procedures described above and tested whether training on drug and food data separately would improve prediction of craving, and whether food craving could be predicted based on a pattern trained on drug data only, and vice versa (Figure 5). Food craving was predicted similarly well by the overall pattern (75% out-of-sample accuracy) as by a craving pattern trained on food cues only (78%). Food craving was also significantly predicted by a pattern trained on drug cues only, but with somewhat lower accuracy across both drug-using and non-using participants (62%). For the prediction of drug craving, the results indicated no substantial improvements for training only on modality-specific (drug) cue trials (69% out-of-sample prediction accuracy) compared to all cues (66%). Drug craving was also significantly predicted by a pattern that was trained to predict food craving (63% out-of-sample predictive accuracy) to training only on food trials. Thus, we did not find evidence for a double dissociation between drug and food craving, but rather significant cross-prediction of drug and food craving. Most importantly, the NCS performed as well as the two cue-specific patterns. Together, this supports the hypothesis of shared representations between drug and food craving – and across drug types.
Modulation by cognitive regulation strategies. Finally, we used a general linear model to assess how craving ratings and responses of the NCS were modulated by the cognitive regulation of craving (ROC) task that was employed in all five studies. Across all participants, craving ratings were influenced by cue type (drug vs. food, F(1,396) = 72.6, p < 0.001) and regulation instruction (NOW vs LATER, F(1, 396) = 92.4, p < 0.001). Drug users reported greater overall craving (F(1, 396) = 56.4, p < 0.001) and this effect interacted with both regulation instruction (F(1, 396) = 5.8, p = 0.017) and cue type (F(1, 396) = 151.0, p < 0.001), such that drug users craved drugs more than non-users, and that they showed slightly higher regulation effects than non-users. This was due to the fact that non-users had very low craving for drug cues, irrespective of regulation condition (see Figure 3d). In line with this ‘floor effect’, there was also a significant interaction between drug-use, cue type, and regulation condition (F(1,396) = 20.3, p < 0.001).
Similar to craving ratings, responses of the NCS were influenced by cue type (drug vs. food, F(1, 396) = 71.9, p < 0.001) and by regulation instruction (NOW vs LATER, F(1, 396) = 43.2, p < 0.001), suggesting that cognitive regulation strategies modify NCS responses. Drug users versus non-users had marginally greater NCS responses overall (F(1, 396) = 3.4, p = 0.067). Drug users’ versus non-users’ signature response differed with respect to cue type (F(1, 396) = 56.9, p < 0.001), such that drug users had higher NCS responses to drug cues than non-users (t(99) = 4.40, p < 0.001), whereas NCS responses to food cues did not significantly differ (t(99) = -1.07). Further, drug users’ versus non-users’ signature response differed with respect to regulation condition (F(1, 396) = 6.9, p = 0.009), such that users showed a larger regulation effect that non-users (t(99) = 2.63, p = 0.0099, see Figure 2e), which was likely driven by more room to down-regulate craving in users compared to non-users.