Abstract
Keywords
INTRODUCTION
Parkinson’s disease (PD) is characterized by rest tremor, akinesia or bradykinesia and cogwheel rigidity. These symptoms result from dopaminergic nigrostriatal denervation. Patients with PD also often present with non-motor symptoms, including cognitive impairment and behavioral disorders (such as dementia, depression, apathy, anxiety, and impulse control disorders (ICDs)). The defining characteristic of ICDs is failure to resist an impulse or temptation to perform an act that is harmful to the person or to others, such as pathological gambling and compulsive shopping, eating or sexual behavior [1]. In a multicenter study of a large group of PD patients, the prevalence of ICDs was 13.6% [2]. Impulse control disorders are strongly associated with the use of dopamine agonists [2, 3] and usually disappear after the latter are discontinued. Known risk factors for ICDs include young age at PD onset, unmarried status, cigarette smoking, caffeine use, a family history of gambling problems, a novelty-seeking temperament and impulsivity [2, 4]. Polymorphisms in the DRD3, GRIN2B, and HTR2A genes have also been associated with ICD in PD [5, 6].
Behavioral studies revealed that PD patients with ICD make more impulsive [7] and risky choices [8], and have a greater reflection impulsivity (making decision before gathering enough information to reduce uncertainty) than PD patients without ICD, suggesting impairment in mapping actual decision to future reward [9]. Metabolic studies of the dopaminergic pathways at rest [10] or during reward visualization [11, 12] have highlighted deregulation of the dopaminergic mesolimbic pathway in ICDs. Activation functional MRI (fMRI) studies (in which adults perform tasks involving risk-taking [8], probabilistic learning [13] or passive visualization of reward cues [14, 15]) have highlighted disturbances in limbic cortical or striatal regions. This suggests that mesolimbic dopaminergic deregulation by dopamine agonists leads to difficulties in reward evaluation (i.e. the outcomes of actions are perceived to be better than they really are). These impairments supposedly bias goal-directed behavior towards more immediate rewards with negative consequences, and thus cause ICDs.
Reward processing can also be investigated by analyzing event-related potentials (ERPs) or event-related spectral perturbation (ERSP). These time-domain approaches complement the fMRI approach and offer better time resolution (albeit at the expense of spatial resolution). Feedback-related negativity (FRN) is an ERP component that peaks between 200 to 300 ms after presentation of a negative outcome [16]. It is thought to reflect the impact of the dopamine signal on neurons in the anterior cingulate. According to the theory of reinforcement learning, FRN may reflect a quantitative reward prediction error [17]. ERPs may reflect partial phase alignment, frequency synchronization or a power increase of cortical oscillations. Therefore, time-frequency analyses are able to more precisely discriminate between the various mechanisms underlying feedback-related changes in EEG signals. By using this type of approach, several groups have reported an increase in frontocentral theta power following a negative outcome [18–25]. This increase is thought to reflect prediction error calculations in the medial prefrontal cortex [24]. In some (but not all) studies, other spectral perturbations in the delta, beta and gamma frequency bands have been associated with the outcome’s valence. The outcome’s amplitude [18] and probability [19] are known to modulate brain oscillations. Surprisingly, very few time-frequency analyses have focused on pathological conditions (such as addictions). Only one study found that alcoholic patients displayed a smaller increase in theta power following a negative outcome than healthy controls (HCs) did [22].
The primary objective of the present study was to determine whether changes in the characteristics of cortical oscillations may be a marker of changes in outcome processing in PD patients with ICDs. To this end, we compared outcome-related ERSP in PD patients with ICDs and without ICDs during the performance of a simple “two choice – single outcome” gambling task. Our starting hypothesis was that PD patients with ICDs would display alterations in cortical oscillations following the presentation of the outcome (i.e. reflecting impaired reward processing).
MATERIAL AND METHODS
Participants
Twenty-four patients with PD defined by international criteria [26]) participated in the study. All were outpatients at Lille University Medical Center’s Neurology and Movement Disorders Department (Lille, France). Twelve of the patients had one or more self-reported ICDs, as judged by their answers to the Questionnaire for Impulsive-Compulsive Disorders in PD (QUIP) [27]. This diagnosis was confirmed during a semi-structured interview with a trained clinician (KD), using standardized criteria for compulsive gambling and eating [28], buying [29] or sexual behavior [30]. Patients with ICDs were matched with (i) 12 PD patients who had never experienced an ICD (as judged by the absence of a “yes” answer in the QUIP, and confirmed in the semi-structured interview) and (ii) 14 HCs. Subjects were matched for gender, age ( ± 5 years) and the duration of PD duration ( ± 1 year). None of the participants in the two control groups had a history of psychiatric disorders or addiction. None of the patients had a neurological disease other than PD.
All participants gave their informed consent to participation in the study, which had been approved by the local institutional review board (CPP Nord-Ouest IV, Lille, France; reference 2010-A00594-35).
The clinical assessment
The severity of PD was rated on the Hoehn and Yahr scale [31]. Motor disability and cognitive status were rated using the motor part of the Unified Parkinson’s Disease Rating Scale (UPDRS-III) [32] and the Mini Mental State Examination (MMSE) [33], respectively. Participants were excluded from the study if they scored less than 27 in the MMSE. The Mini-International Neuropsychiatric Interview was used to screen for psychiatric comorbidities [34].
All patients were assessed in the “on-drug” state, after having taken their usual antiparkinsonian medication(s). The mean L-dopa equivalent daily dose (LEDD) was calculated according to international guidelines [35]. Patients with ICDs were assessed during an ICD episode and prior to any treatment changes. For all PD patients with ICDs, the dose of dopamine agonist was reduced immediately after the study. All PD patients without ICDs were being treated with a dopamine agonist.
The task
The single-outcome gambling task used in the study is depicted in Fig. 1. At the start of each trial, a choice stimulus with two numbers (“5” in the left box and “25” in the right box, corresponding to the value of the bet) was displayed in white against a black background until the participant pressed on the corresponding response button on a joystick (the left button for “5” and the right button for “25”). The outcome stimulus (the outcome of the bet) appeared immediately after the button was pressed and remained on the screen for 2000 ms. The outcome stimulus was shown in green for a gain or in red for a loss, with the overall score displayed in white just below. During standard trials (which accounted for 80% of all trials), the outcome was either a gain or a loss corresponding to the value of the bet. Boost trials (accounting for 10% of all trials) were randomly intermixed within thestandard trials. In a boost trial, the outcome was either a gain or a loss of 125. In order to control for the effect of surprise induced by the boost, a third type of trial (again accounting for 10% of all trials) was randomly intermixed. In this case, the outcome was either a gain or a loss of 7 or 27. In all types of trial, the probability of a gain or a loss was 0.50. The standard, boost and surprise trials and the trial valences were displayed in semi-random order. Patients were informed of neither the frequency of gains/losses nor the existence of boost trials and surprise trials. Patients were told that they would start with a score of zero and would have to accumulate as many points as possible (with the current best score being 200). The task was presented in four different runs of 100 trials each. Subjects were allowed a resting period between each run and the score was reset to zero at the beginning of each run.
For each participant, the frequency with which 25 was chosen (the “risky choice”) was computed for the task as a whole and as a function of the previous outcome (–5, –25, +5, +25, +125, –125, +7, –7,+27, –27).
EEG recording
An EEG was recorded continuously at 128 scalp locations, using a DC amplifier (ANT Software BV, Enschede, The Netherlands) and a Quick-cap 128 AgCl electrode cap (ANT Software BV) placed according to the 10/05 international system [36] with a linked mastoid reference. A vertical electro-oculogram (EOG) was recorded using two electrodes placed 1.5 cm above and below the axis of the right pupil, in order to detect artifacts related to eye movements. We used Advanced Source Analysis (ASA) software (ANT Software BV, Enschede, the Netherlands) for data acquisition. The EEG and EOG signals were digitized with a sampling rate of 1024 Hz. Electrode impedances were kept below 5 kOhms.
EEG analysis
The EEG signal was band-pass filtered between 0.1 and 30 Hz and ocular artifacts were detected and removed off-line. The time-frequency analysis was computed using EEGLAB software [37]. At the Fz, FCz and Cz electrodes, the power spectra were computed between 200 ms before the outcome to 600 ms after the outcome by using a sinusoidal Morlet wavelet transform (window length: 696 ms; 2.5 cycles). We focused on the theta, alpha and beta band. The output frequencies from 4 Hz to 30 Hz were log-transformed. A baseline signal was measured over the 200 ms preceding the display of the outcome. For each participant, power spectra were averaged for each of the following conditions: standard gain (+5/+25), standard loss (–5/–25), high-amplitude outcome (+25/–25), low-amplitude outcome (+5/–5), boost (+125/–125) and surprise (+7/–7/+27/–27). The mean baseline log power spectrum was subtracted from each spectral estimate, to produce the baseline-normalized ERSP. The maximum power in the theta (4–7 Hz) and beta (15–30 Hz) frequency band between 200 and 500 ms after the outcome stimulus were computed for each participant and each condition. As outcome presentation was associated with a decrease in alpha power, minimum power in the alpha (8–14 Hz) frequency band between 200 and 500 ms after the outcome stimulus was also computed. The mean ERSP values for each group were computed by averaging the individual ERSP values.
Statistical analysis
Given that the Kolmogorov-Smirnov test revealed that some variables had a significantly non-normal distribution, non-parametric statistical tests were applied. Intergroup comparisons were performed with a Kruskall-Wallis test. For behavioral and ERSP power data, within-group comparisons of conditions were also investigated using Wilcoxon’s signed rank test.
ERSP maps were compared across groups and conditions by applying a permutation test [38]: participants (for intergroup comparisons), conditions (for inter-condition comparisons) or participants and conditions (for assessments of group × condition interactions) were iteratively shuffled, and analyses of variance were computed for each data point in each of 1000 different permutations. A histogram was built from the most positive and most negative t-values in each random partition. A t-test was then performed on the original trial sets. The proportion of random partitions that resulted in a greater t-value (i.e. more negative or more positive) than the observed value was calculated and corresponded to the p-value.
The threshold for statistical significance was set to p < 0.05 for all analyses.
RESULTS
All data are quoted as the mean ± standard deviation (SD).
Demographic and clinical characteristics
The three groups’ mean clinical and demographic characteristics are shown in Table 1. The groups of PD patients with and without ICDs did not differ significantly in terms of age, PD duration, LEDD, UPDRS score or MMSE score. Of the 12 PD patients with ICDs, eight had one ICD and four had two or more ICDs. Nine patients had compulsive sexual behavior disorders, 4 presented with pathological gambling, 2 had binge-eating disorder and 3 presented with compulsive buying. No patient had punding.
Behavioral variables
Figure 2 show the mean (SD) frequencies of making a risky choice (choosing 25) as a function of the outcome of the previous trials. There were no significant difference between the three groups.
Overall, subjects changed their bet (e.g. choosing 25 after choosing 5, or choosing 5 after choosing 25) more frequently after a loss than after a gain (respectively 22 ± 11%, 22 ± 16% and 27 ± 11% after a gain and 48 ± 17%, 45 ± 17% and 46 ± 8% after a loss in PD patients with ICDs, PD patients without ICDs and HCs; p = 0.003, p = 0.012 and p = 0.001, respectively). Hence, participants displayed a win-stay strategy but not loss aversion; they made random choices because the likelihood of a risky choice did not differ significantly from 0.5 after a loss).
ERSP
Effect of the outcome’s valence
Figure 3 shows the grand average ERSP at Fz, FCz and Cz after “gain” trials and “loss” trials in the three groups. Overall, after the outcome, the power in the theta and beta band increased and the power in the alpha band decreased. In HCs and in PD patients without ICDs, the outcome-related theta power was greater after a gain than after a loss. This was most obvious in the 4-5 Hz band at Cz (between 300 and 600 ms after the outcome in HCs and between 200 and 400 ms after the outcome in PD patients without ICDs). In PD patients with ICDs, valence had no effect on the theta power, except for a small, borderline-significant difference at Fz 250 ms after the outcome. In HCs, an early decrease in theta power (around 150 ms after the outcome) and a late decrease (after 500 ms after the outcome) were observed at Fz and FCz. “Loss” trial were associated in PD patients with ICD but not in the two other groups with an early (before 200 ms) decrease in beta power. Figure 4 shows the maximum power in the theta and beta bands and the minimum power in the alpha band at Fz, FCz and Cz, as a function of the outcome’s valence. In PD patients with ICDs, the maximum theta power was not significantly related to the outcome (i.e. gain or loss). PD patients without ICDs and/or HCs variously displayed a higher theta power following gain than following a loss at Cz (PD without ICDs: 2.95 dB ± 1.35 after a gain vs. 1.94 dB ± 1.29 after a loss, p = 0.019; HCs: 2.38 dB ± 1.27 after a gain vs. 1.79 dB ± 0.96 after a loss, p = 0.006), Fz (HCs: 1.87 dB ± 1.16 after a gain vs. 1.26 dB ± 1.34 after a loss, p = 0.016) and FCz (HCs: 1.80 dB ± 1.18 after a gain vs. 1.26 dB ± 1.19 after a loss, p = 0.013).
Effect of the outcome’s amplitude
Figure 5 shows the grand average ERSP at Fz, FCz and Cz in each of the three groups after trials with a high gain or loss (i.e. 25) or a low gain or loss (5). In all conditions, the power in the theta and beta band increased after the outcome and the power in the alpha band decreased. In PD patients with ICDs, the increase in theta and alpha power was greater after high-amplitude outcomes than after low-amplitude outcomes at Fz, FCz, and Cz. The early increase in theta power after high-amplitude outcomes was greater in PD patients with ICDs than in the two other groups at FCz (with a significant group effect: p < 0.02) and resulted in a group×amplitude interaction effect at Fz. In HCs, the theta power 300 ms after the outcome increased more after high-amplitude outcomes than after low-amplitude outcomes at Fz and FCz. In PD patients without ICDs, the theta power was not significantly related to the outcome’s amplitude. Figure 6 shows the maximum power in the theta and beta bands and the minimum power in the alpha band at Fz, FCz and Cz, as a function of the outcome’s amplitude. HC displayed higher beta power following high-amplitude outcomes than low-amplitude outcomes at Fz (1.84 dB ± 0.92 after a high amplitude outcome vs. 1.52 ± 0.77 after a low amplitude outcome, p = 0.022) and FCz (2.09 dB ± 1.11 after a high amplitude outcome vs. 1.67 ± 0.91 after a low amplitude outcome, p = 0.022), whereas PD patients with ICD showed the opposite pattern at Cz (1.64 dB ± 0.93 after a high amplitude outcome vs. 2.21 ± 0.80 after a low amplitude outcome, p = 0.012).
Effect of boost trials and surprise trials
Figure 7 shows the grand average ERSP at Fz, FCz and Cz after boost trials (i.e. a gain or loss of 125) and surprise trials (i.e. a gain or loss of 7 or 27 when betting 5 or 25, respectively) in the three groups. In all groups, the theta and beta band power increased after the outcome was presented, and the alpha band power decreased. In HCs, the “surprise” outcomes were associated with a decrease in theta power at Fz and FCz between 300 and 400 ms, although the “boost” outcomes induced an increase. In PD patients with ICDs, the theta power was greater after boost outcomes than after “surprise” outcomes at Cz between 150 and 300 ms. The theta power was not significantly related to the outcome in PD patients without ICDs. At Fz, PD patients without ICD displayed an increase in the beta band power 200 ms after the outcome that was not observed in patients without ICD nor HC. Figure 8 shows the maximum power in the theta and beta bands and the minimum power in the alpha band at Fz, FCz and Cz in the boost trials and the surprise trials. In PD patients without ICDs, boost and surprise outcomes did not differ in terms of the maximum theta power. In PD patients with ICDs, the theta power at FCz was greater after boost outcomes than after surprise outcomes (2.82 dB ± 1.37 after a gain vs. 1.94 dB ± 1.07 after a loss, p = 0.023), whereas HCs displayed greater theta power at Fz after boost outcomes than after surprise outcomes (2.17 dB ± 1.10 after a gain vs. 1.44 dB ± 1.53 after a loss, p = 0.035).
DISCUSSION
The primary objective of the present study was to compare outcome-locked, event-related spectral perturbation in a gambling task in PD patients with and without impulse control disorders and therefore to identify EEG markers of impaired reward processing in PD patients with ICDs. Our main finding was that in PD patients with ICDs, the theta band brain oscillations were less strongly modulated by the valence of the outcome in a gambling task (relative to PD patients without ICDs and to HCs). Secondly, PD patients without ICDs displayed a smaller increase (relative to HCs) in theta band power following unexpected high outcomes (i.e. boost trials). In PD patients with ICDs, the modulation by boost trials was similar to that observed in HCs.
Effect of the outcome’s valence
We observed a greater increase in theta power following gains than following losses in PD patients without ICDs and in HCs but not in PD patients with ICDs. This finding was in line with our starting hypothesis and suggests that reward processing is impaired in PD patients with ICDs. All but one [39] of the previous studies in this field reported that loss outcomes elicited higher theta power than gain outcomes [18–25, 40–42]. There are several possible explanations for this discrepancy. Firstly, none of the studies included PD patients. Secondly, our participants were older than those in other ERSP studies of outcome integration (most of which recruited undergraduate students). To the best of our knowledge, no study has reported that age has an impact on EEG spectral perturbation following the display of an outcome. However, ERP studies have clearly identified changes in cortical markers of reward monitoring with age, since older people having less marked FRN following losses than younger people do [43–48]. Moreover, the present study was based on a simple gambling task that included infrequent boost trials and surprise trials. To the best of our knowledge, the effect of boost trials has never been examined in time-frequency studies of reward processing. In ERP studies, the amplitude of the FRN depends on the range of possible outcomes in the task [49]. Similarly, our inclusion of a boost that represented five times the high-amplitude bet may have reduced (i) the subjective value of the high-amplitude gains, (ii) the aversive value of the high-amplitude losses and thus (iii) the effect of valence in standard trials. Lastly, the inclusion of infrequent boost and surprise trials necessarily led to the use of a large number of trials (400), which could have induced fatigue and decreased motivation in the participants. However, to avoid this type of effect, the task was divided into four blocks of 100 trials and the participants were asked to try to improve their performance in each run.
Effect of the outcome’s amplitude
Whereas many studies have reported that theta band EEG power is modulated by the outcome’s valence, few have reported data on the effect of the outcome’s amplitude. In a study of 61 young HCs, it was shown that theta power increased after a small-amplitude outcome and beta power increased after a high-amplitude outcome [18]. In contrast, a study on 40 HCs found that high amplitude outcome yield higher theta power [39]. A third study of HCs and alcoholic patients failed to identify a significant effect of the outcome’s amplitude [22]. In the present study, theta power increased more after high-amplitude outcomes than after low-amplitude outcomes in PD patients with ICDs. This increase occurred soon after the outcome (within 100 ms). It is noteworthy that our high-amplitude and low amplitude outcomes corresponded to a risky choice and a non-risky choice, respectively. Early modulation of theta band oscillation by the reward amplitude in PD patients with ICDs might therefore constitute a marker of risk rather than of feedback processing. The increase in theta power after a risky choice (relative to a non-risky choice) was only observed in PD patients with ICDs (and, at a greater latency, in HCs) but not in PD patients without ICDs. These results suggest that risk modulates theta power less intensely in PD patients without ICDs than in PD patients with ICDs. In the beta band, we found an increase in power following a high-amplitude outcome in HC, as previously reported [18], but PD patients with ICD displayed an opposite modulation at a more posterior scalp location. Globally, beta band oscillations have been linked with the maintenance of a motor or cognitive state [50], but its functional role during reward processing are largely unknown.
Effect of the boost trials
Given that our starting hypothesis predicted an impairment in reward processing in PD patients with ICDs, we included unexpected, high-amplitude trials (boosts) because this type of rare gain or loss reportedly induces an increase in the phasic activity of midbrain dopaminergic neurons [51]. In PD patients with ICDs and HCs, theta band power was greater after boost trials than after surprise trials at Fz (HCs) or FCz (PD patients with ICDs). In contrast, there was no difference in theta power for PD patients without ICDs when comparing boost and surprise outcomes. Unlike surprise and standard trials, boost trials constitute highly salient signals and (according to models of reinforcement learning) induce major errors in reward prediction. The latter is known to be correlated with the theta band power after presentation of the outcome. This finding supports the hypothesis whereby theta band oscillations reflect prediction error calculations in the medial prefrontal cortex [24] - presumably following activation of the mesolimbic dopaminergic pathway. Our results thus suggest that PD patients with ICDs are more sensitive to unexpected, salient outcomes (relative to PD patients without ICDs) and that PD patients without ICDs are less sensitive (relative to HCs).
Behavioral results
In the present study, PD patients (regardless of the presence of ICDs) and HCs did not differ in their response patterns: all tended to adopt a win-stay strategy and made a more random choice after losses. This pattern is analogous to that observed in a previous study of a similar gambling task [18] in HCs under standard and stressed conditions.
Study limitations
The present study had several limitations. Firstly, the number of participants in each group may have be too small to reliably identify intergroup differences. However, the absence of literature data on PD patients made an a priori power calculation impossible to perform. Moreover, the three groups were strictly matched for demographic characteristics, and this limited the enrollment rate. Secondly, EEG data are sensitive to artefacts - especially those related to movement, which is a major concern when evaluating PD patients. In order to limit movement-related artefacts, we evaluated patients in the on-drug state and applied strict, automatic procedures for the correction of visual artefact. Thirdly, the groups of PD patients with ICDs comprised various ICD subtypes, and the sample size was too small to enable stratification of the data for each of these. It is likely that the mechanisms underlying (for example) compulsive gambling and compulsive sexual behavior are not identical. The inclusion of patients with different ICDs probably introduced variability into our data. Finally, theta band oscillations have been linked with various processes, including sensory-motor integration [52] and motor control [53]. Motor control is impaired in PD and could therefore be a confounding factor in our analysis. Nevertheless, PD patients with and without ICD were carefully matched and there were no inter-group differences in terms of treatment, disease duration or severity of the motor symptoms.
Taken as a whole, our results show that PD patients with ICDs are (i) less sensitive to the outcome’s valence, (ii) more sensitive to risk and (iii) more sensitive to reward prediction errors. Given that learning was not possible in the study’s task and that the probability of a gain was set to 0.50, impaired prediction error monitoring did not induce behavioral modifications during the task. Although tasks based on probabilistic learning in fMRI experiment have already identified this type of alteration in reward learning [13], they involve working memory and executive functions – both of which are impaired in PD. Hence, we decided to use a simple gambling task to focus on reward integration and to avoid bias related to cognitive impairment in the patients. However, it would also be interesting to analyze EEG signals during more complex, reward-learning tasks.
CONCLUSIONS
In patients with PD, ICDs are associated with (i) decreased modulation of cortical markers of reward by the outcome’s valence and (ii) increased modulation of these markers by risk and an unexpected, salient outcome. Our results support the hypothesis whereby prediction error computation in the medial prefrontal cortex is impaired (as reflected by frontal theta band oscillations) – probably as a consequence of deregulation of the mesolimbic dopaminergic pathway.
CONFLICT OF INTEREST
The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be considered as potential conflicts of interest.
Footnotes
ACKNOWLEDGMENTS
David Fraser provided editorial support; Marie Delliaux and Anne-Sophie Carette helped with the clinical evaluations; the Clinical Research Federation at Lille University Medical Center provided technical support.