Cue-induced effects on decision-making distinguish subjects with gambling disorder from healthy controls

While an increased impact of cues on decision-making has been associated with substance dependence, it is yet unclear whether this is also a phenotype of non-substance related addictive disorders, such as gambling disorder (GD). To better understand the basic mechanisms of impaired decision-making in addiction, we investigated whether cue-induced changes in decision-making could distinguish GD from healthy control (HC) subjects. We expected that cue-induced changes in gamble acceptance and specifically in loss aversion would distinguish GD from HC subjects. 30 GD subjects and 30 matched HC subjects completed a mixed gambles task where gambling and other emotional cues were shown in the background. We used machine learning to carve out the importance of cue-dependency of decision-making and of loss aversion for distinguishing GD from HC subjects. Cross-validated classification yielded an area under the receiver operating curve (AUC-ROC) of 68.9% (p=0.002). Applying the classifier to an independent sample yielded an AUC-ROC of 65.0% (p=0.047). As expected, the classifier used cue-induced changes in gamble acceptance to distinguish GD from HC. Especially increased gambling during the presentation of gambling cues characterized GD subjects. However, cue-induced changes in loss aversion were irrelevant for distinguishing GD from HC subjects. To our knowledge, this is the first study to investigate the classificatory power of addiction-relevant behavioral task parameters when distinguishing GD from HC subjects. The results indicate that cue-induced changes in decision-making are a characteristic feature of addictive disorders, independent of a substance of abuse. Remarks To ensure a more convenient reviewing process, we positioned figures and tables at their destined position.


INTRODUCTION
Gambling disorder (GD) is characterized by continued gambling for money despite severe 2 negative consequences 1 . Burdens of GD include financial ruin, loss of social structures, as well 3 as development of psychiatric comorbidities 2 . In line with this clinical picture of impaired 4 decision making, GD subjects have also displayed impaired decision making in laboratory 5 experiments 3,4 . 6 Besides impaired decision making, cue reactivity has been a crucial concept in understanding 7 addictive disorders including GD 5,6 . Through Pavlovian conditioning, any neutral stimulus can 8 become a conditioned stimulus (i.e. a cue) if it has been paired with the effects of the addictive 9 behavior 7 . In addictive disorders, including GD, cues may induce attentional bias, arousal, and 10 craving for the addictive behavior in periods of abstinence 8,9 . Treatment of addictive disorders 11 may focus on identifying and coping with individual cues that induce craving for addictive 12 behavior 10 . If we understood better how cues exert control over instrumental behavior and 13 decision-making, we would be able to improve treatment tools and even public health policy 14 for GD and perhaps other addictive disorders. In the present study we were thus interested in 15 broadening our understanding of the basic mechanisms of impaired decision making in 16 addictions, especially with respect to cue-induced effects on value-based decision making. 17 The effect of cues exhibiting a facilitating or inhibiting influence on instrumental behavior and 18 decision making is known as Pavlovian-to-Instrumental Transfer (PIT) 11 . PIT experiments 19 usually have three phases: a first phase where subjects learn an instrumental behavior to gain 20 rewards or avoid punishments, a second phase where subjects learn about the value of arbitrary 21 stimuli through classical conditioning, and a third phase (the PIT phase), where subjects are 22 supposed to perform the instrumental task, while stimuli from the second phase (changing from 23 trial to trial) are presented in the background. The PIT phase measures the effect of value-charged cues on instrumental behavior despite the fact that the background cues have no 1 objective relation to the instrumental task in the foreground. For instance, certain cues could 2 increase the likelihood of gamble acceptance or the sensitivity to the gain offered in the gamble. 3 In the current study we focus only on the PIT phase. PIT has recently drawn attention in the 4 study of substance use disorders (SUDs) 12 . This is because PIT effects can persist even when 5 the outcome of the instrumental behavior has been devalued 13 , and because increased PIT has 6 been associated with a marker for impulsivity 14 and with decreased model-based behavior 15 . 7 Lastly, PIT effects tend to be stronger in subjects with a substance-use-disorder than in healthy 8 subjects 12,16 , and increased PIT has been associated with the probability of relapse 12 . 9 Increased PIT effects are based on Pavlovian and instrumental conditioning and on their 10 interaction. This highlights how addictive disorders rely on learning mechanisms 17 . GD is an 11 addictive disorder independent of any influence of a neurotropic substance of abuse. The study 12 of PIT in GD may thus further shed light on whether increased PIT in addictive disorders is a 13 result of learning, independent of any substance of abuse, or even a congenital vulnerability 18 . 14 We are aware of three studies that have observed in GD subjects increased cue-induced effects 15 on decision-making and instrumental behavior, comparable to increased PIT effects. In two 16 single-group studies, GD subjects have shown higher delay discounting (preferring immediate 17 rewards over rewards in the future) in response to a casino environment vs. a laboratory 18 environment 19 and to high-craving vs. low-craving gambling cues 20 . In a third study, GD 19 subjects have been more influenced than HC subjects by gambling stimuli in a response 20 inhibition task 21 . To our knowledge, however, there are no studies yet that have investigated the 21 effect of cue reactivity on loss aversion in GD as a possibly relevant PIT effect in GD. 22 Loss aversion (LA) is, besides delay discounting, another facet of value-based decision-making. 23 It is the phenomenon wherein people assign a greater value to potential losses than to an equal 24  Psychology of Humboldt-Universität zu Berlin. They were sitting upright in front of a computer 10 screen using their dominant hand's fingers to indicate choices on a keyboard. Subjects were 11 attached five passive facial electrodes, two above musculus corrugator, two above musculus 12 zygomaticus, and one on the upper forehead. We recorded electrodermal activity (EDA) from 13 the non-dominant hand. Subjects of the validation sample completed the task in an fMRI 14 environment (head-first supine in a 3-Tesla SIEMENS Trio MRI at the BCAN -Berlin Center of Advanced Neuroimaging). Results of the fMRI and peripheral-physiological recordings will 1 be reported elsewhere. 2

3
We were inspired by established tasks to measure general LA and LA under the influence of 4 affective cues 27,35 . Subjects were each given 20€ for wagering. On every trial, subjects saw a 5 cue that they were instructed to memorize for a paid recognition task after the actual experiment. 6 After 4s (jittered), a mixed gamble, involving a possible gain and a possible loss, with 7 probability P = 0.5 each, was superimposed on the cue. Subjects had to choose how willing they 8 were to accept the gamble (Fig. 1A) on a 4-point Likert-scale to ensure task engagement 35 . 9 Subjects of an independent validation sample completed the task in an fMRI scanner and had 10 an additional wait period to decide on the gamble (Fig. 1B). Gambles were created by randomly 11 drawing with replacement from a matrix with possible gambles consisting of 12 levels of gains 12 (14, 16, …, 36) and 12 levels of losses (-7, -8, …, -18). This matrix is apt to elicit LA in healthy 13 subjects 23,35 . Outcomes of the gambles were never presented during the task but subjects were 14 informed that after the experiment five of their gamble decisions with ratings of "somewhat 15 yes" or "yes" would be randomly chosen and played for real money. As affective cues, four sets 16 of images were assembled: 1) 67 gambling images, showing a variety of gambling scenes, and 17 paraphernalia (gambling cues) 2) 31 images representing negative consequences of gambling 18 cue categories, we randomly drew images from each pool until we had presented 45 images of 23 each category and each image at least once. Hence, we ran 202 trials in each subject. Gambles were matched on average across cue categories according to expected value, variance, gamble 1 simplicity, as well as mean and variance of gain and loss, respectively. Gamble simplicity is 2 defined as Euclidean distance from diagonal of gamble matrix (ed) 35 . HC showed on average 3 1.00 missed trial, GD 1.05 (no significant group difference, F = 0.022, p = 0.882). In fMRI 4 validation study, HC: 3.13, GD: 4.10, (no significant group difference, F = 0.557, p = 0.457).  and loss was counterbalanced (left/right). Gain was indicated by a '+' sign and loss by a '-' sign. In the behavioral 10 sample, subjects had 4s to make a choice between four levels of acceptance (yes, somewhat yes, somewhat no, 11 no; here translated from German version that used "ja, eher ja, eher nein, nein"). In the fMRI sample, subjects 12 had to wait 4s (jittered) before the response options were shown. Direction of options (from left to right or vice 13 versa) was random. Directly after decision, the ITI started. If subjects failed to make a decision within 4s, ITI 14 started and trial was counted as missing. ca.: circa, RT: reaction time Subjective cue ratings 1 After the task, subjects rated all cues using the Self-Assessment Manikin (SAM) assessment 36 2 (reporting on valence: happy vs. unhappy, arousal: energized vs. sleepy, dominance: in control 3 vs. being controlled) and additional visual analogue scales: 1) "How strongly does this image 4 trigger craving for gambling?" 2) "How appropriately does this image represent one or more 5 gambling games?" 3) "How appropriately does this image represent possible negative effects 6 of gambling?" 4) "How appropriately does this image represent possible positive effects of 7 gambling abstinence?". All scales were operated via a slider from 0 to 100. 8 All cue ratings were z-standardized within subject. Ratings were analyzed one-by-one using 9 linear mixed-effects regression, using lmer from the lme4 package in R 37 , where cue category 10 (and clinical group) denoted the fixed effects and subjects and cues denoted the sources of 11 random effects. 12 Estimating subject-specific parameters from behavioral choice data 13 We modeled each subject's behavioral data by submitting dichotomized choices (somewhat no, 14 no: 0; somewhat yes, yes: 1) into logistic regressions. We dichotomized choices to increase the 15 precision when estimating behavioral parameters, in line with previous studies using the mixed 16 gambles task 23,35 . Regressors for subject-wise logistic regressions were gain (mean-centered) 17 and absolute loss (mean-centered) from the mixed gamble, as well as gamble simplicity (ed), 18 loss-gain ratio and cue category of the stimulus in the background of the mixed gamble. We 19 defined different logistic regressions by using different trial-based definitions of gamble value 20 ( ) (see Tab. S1), submitted to the logistic function: 21 Different trial-based definitions of gamble value ( ) reflected two things: 23 1) Different ways of modeling LA may be adequate to distinguish a GD from a HC 1 subject 23,25,27,35 (Tab. S1). 2 2) Different ways of incorporating cue effects on decision-making (PIT effects) may be 3 adequate to distinguish a GD from a HC subject. For example, the model lac assumes 4 … 5 where 0 is the intercept, the objective gain value of the gamble, the 9 regression weight for (same holds for and , respectively), and c the 10 dummy-coded column vector indicating the category of the current cue and a column 11 vector holding the regression weights for the categories. Lac thus is a weighted linear 12 combination of objective gain, objective loss with an additive influence of cue category. 13 That is, some influence of cue category on decision-making (PIT) is modeled. Note that 14 we have multiple PIT effects here, because is a vector of length three, reflecting the 15 three affective categories (gambling, negative, positive) different from neutral. There 16 were also models that did not incorporate any influence of loss aversion or category 17 Overall reasoning in building the classifier 18 The main interest of our study was to assess whether cue-induced changes in decision-making 19 during an affective mixed gambles task can be used to distinguish GD from HC subjects. We 20 hypothesized that shifts in loss aversion that depend on what cues are shown in the background 21 ("loss aversion PIT") should best distinguish between GD and HC subjects. This means, the laci model's parameter set should have been the most effective in distinguishing between GD 1 and HC subjects. To test this hypothesis, we used a machine learning algorithm based on 2 regularized logistic regression that selected among various competing parameter sets (from the 3 21 different models, la, lac, laci, etc.) the set that best distinguished HC and GD subjects. 4 To assess the generalizability of the resultant classifier, we used cross-validation (CV) 30,32,39,40 . 5 Generalizability estimates the predictive power, and hence replicability, of a classifier in new 6 samples 28 . Note that machine learning algorithms are designed to generalize well to new 7 samples by inherently avoiding overfitting to the training data 41(p9) . We computed a p-value of 8 the algorithm denoting the probability that its classification performance was achieved under a 9 baseline model (predicting using only smoking severity as predictor variable). 10 Beyond cross-validation, which uses only one data set (splitting it repeatedly into training and 11 test data set), validation of a classifier on a completely independent sample is the gold-standard 12 in machine learning to assess the quality of an estimated model 28 . Hence, we estimated the 13 generalization performance also via application of our classifier to a completely independent 14 sample of HC and GD subjects, who had performed a slightly adapted version of the task in an 15 fMRI scanner. A p-value was computed, as above, with random classification as the baseline 16 model. For detailed information on estimating the classifier, please see Supplements (1.4 and 17 Gambling cues were seen as more appropriately representing one or more gambling games than 3 any other cue category: gambling > neutral (β = 1.589, p < 0.001), gambling > negative (β = 4 1.197, p < 0.001), gambling > positive (β = 1.472, p < 0.001). They elicited more craving in GD 5 subjects (β = 0.71, p < 0.001). Negative cues were seen as evoking more negative feelings in 6 both groups (β = -0.775, p < 0.001) and were seen as representing negative effects of gambling, 7 more than any other category (Supplements 2.1). Positive cues were indeed seen as more 8 representative for positive effects of gamble abstinence than any other category (Fig. S2). 9

10
The classification algorithm yielded an AUC-ROC of 68.9% (under 0-hypothesis, i.e. with only 11 smoking as predictor: 55.1%, p = 0.002) (Fig. 2B, S4). The most often selected model was the 12 "acceptance rate per category" (ac) model (90.7% of the rounds). Combined with the models 13 laec, lac in 95.8% of the rounds a model was used that incorporated PIT, i.e. an effect of cue 14 category on decisions (Fig. S5). In only 9.3% of the rounds a model was selected that 15 incorporated loss aversion (i.e. gain and loss sensitivities). Validating the estimated classifier in 16 the independent sample, the classifier yielded an AUC-ROC of 65.0% (under random 17 classification: 55.3%, p = 0.047) (Fig. 2C). Adjusting for cue repetition and equalizing the 18 number of trials across cue categories lead to very similar AUR-ROC scores, the ac model was 19 still the most often chosen model (42%), otherwise laec_LA and lac were chosen very often 20 (Supplements 2.4). 21

Inspection of classifier
1 Inspecting the classifier's logistic regression weights, we saw that the classifier places most 2 importance on the shift in gambling acceptance during gambling cues (see Fig. 2D). Note 3 further that the classifier places also some importance on the sensitivity to the negative cues but 4 deselects the sensitivity to positive cues. 5

Acceptance rate and loss aversion under cue conditions
6 Overall acceptance rate between groups was not significantly different (HC: 53%, GD: 58%, p 7 = 0.169, ΔAIC = 0). Across all subjects there was a significant effect of cue category on 8 acceptance rate (p < 0.001, ΔAIC = 648), driven by the effect of positive and negative cues. 9 There was a significant interaction with group (p = 0.002, ΔAIC = 9). There, GD subjects 10 showed significantly higher acceptance rate during gambling cues than HC subjects (HC: 49%, 11 GD: 68%, pWaldApprox = 0.003) ( Fig. 2A), and there were no more cue effects in the HC group 12 and no other significant cue effect differences between HC and GD. 13 The fixed effects for gain sensitivity, absolute loss sensitivity, and LA over all trials for HC 14 (0.26, 0.42, and 1.64) were descriptively larger than for GD (0.19, 0.22, and 1.13). Testing the 15 interaction between group, gain, and loss (i.e. testing for difference of LA between groups) via 16 nested model comparison, yielded p < 0.001, ΔAIC = 93, with sensitivity to loss being 17 significantly smaller in GD subjects pWaldApprox = 0.011. Loss aversion was significantly smaller 18 in GD than in HC (pperm < 0.001). Loss aversion shifts due to category did not differ between 19 groups (Supplements 2.2).

9
Standardized regression parameters and their confidence intervals (percentiles across cross-validation rounds).

10
The algorithm mainly used the shift in acceptance rate in response to gambling cues in order to detect GD 11 subjects.

DISCUSSION 1
Gambling disorder (GD) is characterized by impaired decision making 4 and craving in response 2 to gambling associated images 9 . However, it is unclear whether specific cue-induced changes 3 in loss aversion exist that distinguish GD from HC subjects. In order to better understand the 4 basic mechanisms of impaired decision-making in addiction, we thus used a machine-learning 5 algorithm to determine the relevance of cue-induced changes on loss aversion ("loss aversion 6 PIT") in distinguishing GD from HC subjects. We hypothesized that cue-induced changes in 7 gamble acceptance and especially a strong shift of loss aversion by gambling and other affective 8 cues should distinguish GD from HC subjects (i.e. the model representing this effect should 9 have been chosen most often by the algorithm to distinguish GD from HC subjects). To our 10 knowledge, our study is the first to investigate the classificatory power of addiction-relevant 11 behavioral task parameters when distinguishing GD from HC subjects. Moreover, we are not 12 aware of any study specifically investigating the relevance of behavioral PIT effects in 13 characterizing addicted subjects using predictive modeling. 14 Our algorithm was significantly better in distinguishing GD from HC subjects than the control 15 model, which only used smoking severity as a predictor variable (cross-validated AUC-ROC of 16 68.9% vs. 55.1%, p = 0.002). In an independent validation sample the classifier was almost as 17 accurate (AUC-ROC of 65.0% vs. 55.3%, p = 0.047). When classifying subjects, in 93% of the 18 estimation rounds, our algorithm chose a model with some influence of the cue categories on 19 choices. The most frequently chosen model was the ac model (85%), i.e. a model only 20 accounting for mean shifts in acceptance rate depending on cue category. PIT-related variables 21 could therefore successfully discriminate between GD and HC subjects. We saw that especially 22 the tendency of subjects to gamble more during the presentation of gambling cues was indicative 23 of the subject belonging to the GD group. Contrary to what we expected, "loss aversion PIT" was not useful in distinguishing between GD and HC subjects. In other words, the algorithm 1 never selected the laci model, which included the modulation of gain and loss sensitivity by cue 2 categories. We also did not see this in univariate group comparisons. "Loss aversion PIT" might 3 thus not play a role in distinguishing GD from HC subjects. However, small sample size, as in 4 the present study, may limit the possible complexity of a classifier 42(p237) . It cannot be ruled out 5 that larger and more diverse samples in future studies may produce classifiers allocating at least 6 minor importance to "loss aversion PIT". 7 We observed that both GD and HC subjects perceived the cues as intended. GD subjects 8 reported higher craving for gambling in response to gambling stimuli as seen in other studies 9 . 9 Our results may thus be interpreted as cue reactivity leading to more automatic decision-making 10 in GD subjects. Note that this does not mean that GD subjects simply show higher vigor or more 11 disinhibition to press a button, as in some PIT designs 43 . Instead, since the required motor 12 response for saying yes or no changed randomly, gamblers seemed to be indeed more inclined 13 to decide in favor of gambling when gambling cues were shown in the background. Especially 14 because cue influence on LA was not relevant for distinguishing GD from HC subjects, but 15 instead cue influence on general acceptance rate, this may be seen as GD subjects responding 16 more habitually and in a less goal-directed manner 15 when gambling cues are visible. 17 In the current study, the classifier also put some importance on behavior under negative cues, 18 and, descriptively but not significantly, GD subjects tended to reduce gambling more in the face 19 of negative cues than HC subjects. Future studies should explore the possible power of negative 20 images to inhibit gambling in larger and more heterogeneous GD samples. 21 Our results show the gambling promoting effects of gambling cues in GD subjects. Alcohol and 22 tobacco advertisement promote alcohol and tobacco use 44 and advertisement bans and counter-23 active labels on alcohol and tobacco goods help reduce consumption 45 . Our results suggest that much like advertisement for these substances, visual stimuli in gambling halls and on slot 1 machines may also increase PIT effects. Policy makers may consider our results as another 2 piece of evidence that gambling advertisement is not different from alcohol and tobacco 3 advertisement and that respective advertisement regulation perhaps should be extended. 4 We are not aware of any machine learning studies that have assessed the relevance of a 5 behavioral task measure in characterizing GD. Using this approach, we observed a cross-6 validated classification performance of AUC-ROC = 0.68. We are aware of one machine 7 learning study that built and tested a classifier in 160 GD patients and matched controls based 8 on personality questionnaire self-report, reaching an AUC-ROC = 0.77 31 . Studies in the field of 9 substance-based addiction, using behavioral markers and machine learning for classification, 10 report cross-validated AUC-ROC's of 0.71 to 0.90 for cross-validated classification 11 performance 30,39 . However, the mentioned studies used a whole array of different informative 12 variables while the current studied tried to carve out the relevance of one basic behavioral 13 mechanism while controlling for all covariates of no-interest. 14 Our results may be a first building block in creating more advanced and more multivariate 15 diagnostic tools for GD and other addictive disorders, especially when combined with other 16 high-performing discriminating features, such as personality profiles and scores from other 17 decision-making tasks. Further, our results invite more in-depth scrutiny of decision-making in 18 GD subjects during the presence of cues, e.g. on neural level 34 . Moreover, the above machine 19 learning studies did not use an independent validation sample to corroborate their results. Our 20 independent validation yielded an AUC-ROC of 0.65. This supports the validity of our findings 21 of increased PIT in GD. 22 STRENGTHS AND LIMITATIONS 1 When carving out the relevance of PIT, we did not match for depression score (BDI) because, 2 epidemiologically, GD is associated with high depression scores 46 , meaning it could be seen as 3 a feature of GD. Further, the evidence on the association of PIT and depression is 4 inconclusive 47,48 . However, PIT might play some role in depression and thus also in GD 5 subjects. Future studies should thus address the modulatory effect of depressive symptoms in 6 GD on PIT 49 . 7 The current classifier was slightly less effective in the independent validation sample than 8 estimated using cross-validation (AUC = 65.4% vs. 68.0%). This might have occurred due to 9 the use of an fMRI version of the affective mixed gambles task in the validation sample. It 10 included an additional decision-making period, during which subjects could not yet answer. 11 This may have led to slight changes in responses with respect to the cue categories. However, 12 this could be seen as a strength since our classifier still performed better than chance. And 13 classifiers that are robust against slight changes in the experimental set-up allow arguably more 14 general conclusions than classifiers that only work with data from the same experimental set-15 up. Future studies should also use validation samples 40 . 16 Cues were repeated and trial numbers were not perfectly balanced across categories. We 17 adjusted for this in our analyses and results were stable. Here, model selection geared also 18 towards reduced loss aversion additionally characterizing GD, in line with 23,24 . CONCLUSION 1 Our results propose that GD subjects' acceptance of mixed gambles is cue-dependent and that 2 this cue-dependency even lends itself to distinguishing GD from HC subjects in out-of-sample 3 data. However, we did not observe that cues specifically shift loss aversion, neither on average, 4 nor in a way relevant to classification. We saw that especially gambling cues lead to increased 5 gambling GD subjects. Observing increased PIT in GD suggests that PIT related to an addictive 6 disorder might not depend on the direct effect of a substance of abuse, but on related learning 7 processes 17 or on innate traits 18 . The here reported effects should be explored further in larger, 8 more diverse and longitudinal GD samples as they could inform diagnostics, therapy 50 and 9 public health policy.

ONLINE MATERIAL 1
You can find the data and R Code to reproduce the analyses here: 2 https://github.com/pransito/PIT_GD_bv_release 3 AUTHORS' CONTRIBUTION: 1 AG designed the experiment, collected the data, analyzed the data, and wrote the manuscript. 2 MA implemented the ratings and questionnaire electronically, analyzed the ratings data, and 3 revised the manuscript. KB collected data and revised the manuscript. CM reviewed the 4 machine-learning algorithm and revised the manuscript. AH revised the manuscript, and 5 oversaw manuscript drafting and data analyses. AW revised the manuscript and oversaw 6 implementation of experiment in the lab. NK revised the manuscript and, advised first author. 7 NRS designed and supervised study and experiment, and oversaw manuscript drafting and data 8 analyses