An insula-driven network computes decision uncertainty and promotes abstinence in chronic cocaine users

Abstract

The anterior insular cortex (AIC) and its interconnected brain regions have been associated with both addiction and decision-making under uncertainty. However, the causal interactions in this uncertainty-encoding neurocircuitry and how these neural dynamics impact relapse remain elusive. Here, we used model-based fMRI to measure choice uncertainty in a motor decision task in 61 individuals with cocaine use disorder (CUD) and 25 healthy controls. CUD participants were assessed before discharge from a residential treatment program and followed for up to 24 weeks. We found that choice uncertainty was tracked by the AIC, dorsal anterior cingulate cortex (dACC), and ventral striatum (VS), across participants. Stronger activations in these regions measured pre-discharge predicted longer abstinence after discharge in individuals with CUD. Dynamic causal modelling revealed an AIC-to-dACC directed connectivity modulated by uncertainty in controls, but a dACC-to-AIC connectivity in CUD participants. This reversal was mostly driven by early-relapsers (<30 days). Furthermore, CUD individuals who displayed a stronger AIC-to-dACC excitatory connection during uncertainty encoding remained abstinent for longer periods. These findings reveal a critical role of an AIC-driven, uncertainty-encoding neurocircuitry in protecting against relapse and promoting abstinence.

1. Introduction

A key computational process involved in decision-making is the estimation of action-outcome probabilities (De Martino et al., 2013; Payzan-LeNestour et al., 2013; Meyniel et al., 2015). That is, an individual must be able to accurately represent how likely her chosen action is to lead to a desirable result: we defined this quantity as choice uncertainty, following previous literature (Meyniel et al., 2015; Adler & Ma, 2018; Atiya et al., 2020). Previous studies have shown the involvement of several brain regions in the estimation of such probabilities (Payzan-LeNestour et al., 2013; Ma & Jazayeri, 2014; Morriss et al., 2019), including the anterior insular cortex (AIC) (Clark et al., 2008; Preuschoff et al., 2008), dorsal anterior cingulate cortex (dACC) (Rushworth & Behrens, 2008; Stolyarova et al., 2019), and ventral striatum (VS) (Berns et al., 2001). For instance, Preuschoff and colleagues used a gambling task where the probability of stimulus-reward associations changed over time. They found that in healthy controls, the AIC not only tracked this probability per se (termed “risk” in the paper) but also errors in such risk prediction (Preuschoff et al., 2008). The dACC, a brain region closely connected and often co-activated with the AIC, has been consistently implicated in decision making arbitrations based on conflict monitoring and the associated uncertainty estimations (Rushworth & Behrens, 2008; Stolyarova et al., 2019) across tasks such as Stroop, flanker or Simon tasks (Venkatraman et al., 2009; Shenhav et al., 2013).

The same AIC-related circuitry has also been commonly associated with substance use disorders (SUDs) and addiction-related phenotypes (Goldstein et al., 2009; Naqvi & Bechara, 2009; Koob & Volkow, 2016), with the VS particularly implicated in association in cocaine use disorder (CUD) and risk-preference behavioral traits (Ersche et al., 2011; Mitchell et al., 2014). For example, McHugh and colleagues used resting-state fMRI (rsfMRI) and found reduced insula-striatum coupling in individuals with CUD and especially those who relapsed early (<30 days) (McHugh et al., 2013) but enhanced rsfMRI connectivity in the AIC-dACC circuitry in those who managed to stay abstinent (McHugh et al., 2016). It has been hypothesized that the insula might drive a network of brain regions in integrating interoceptive states and cognitive control under uncertainty and risk in addiction (Naqvi et al., 2014), a hypothesis partially supported by animal models (Koob & Volkow, 2016). However, a causal role of the insula in computing uncertainty has not been directly examined or established.

Based on this literature, we formulated two hypotheses: first, that the AIC drives the activity of its interconnected regions (e.g. dACC) in encoding choice uncertainty. Second, we hypothesized that alterations in the neural dynamics of this AIC-driven network is associated with phenotypic differences in SUD severity, measured in terms of time to relapse, or the reinstatement of compulsive drug-seeking behaviors after drug abstinence (Kelley, 2004). We tested these hypotheses by examining 25 healthy controls (HCs) and 61 participants with CUD. All CUD participants were assessed right before discharge from a residential treatment program and followed up for 24 weeks, or until the first day of any use of cocaine or amphetamine. We defined the categories of early and late relapse a priori, with a cutoff set at 30 days after discharge (Adinoff et al., 2015; McHugh et al., 2016). We used a Bayesian model to quantify choice uncertainty in a motor decision-making task (i.e. stop signal task; Fig. 1) and dynamic causal modelling (DCM, Friston et al., 2003) was applied to model effective connectivity (i.e. directed influence between neuronal populations) subserving uncertainty encoding. The stop signal task has been traditionally used to examine response inhibition and impulsivity (Rubia et al., 2003; Schall et al., 2017; Verbruggen et al., 2019). However, participants also form beliefs representing the probability for a chosen action to be correct (e.g. left-go vs. stop), given an observed sequence of stimuli and the possibility to receive a stop signal after an unknown time interval (cf. Chikazoe et al., 2009; Zandbelt & Vink, 2010). In Bayesian terms, newly acquired information, such as the symbolic visual stimuli of the stop signal task, is used to update prior beliefs (Friston, 2010; Friston et al., 2016), which can be consolidated or weakened. Thus, each choice selection can be associated with a degree of uncertainty, representing the estimated probability the chosen action might not be correct (Meyniel et al., 2015; Adler & Ma, 2018; Atiya et al., 2020). A few studies have adopted a similar Bayesian approach to examine prediction error encoding in methamphetamine-dependent individuals using the stop signal task (Harle et al., 2015; Harle et al., 2016). Here, we will instead examine subject-specific uncertainty encoding and network dynamics to highlight their contribution to CUD relapse, consistent with the computational psychiatry approach to characterize ‘deep’ cognitive processes (Montague et al., 2012; Adams et al., 2016).

Fig. 1. Task details.

In this stop signal task, the initial inter-trial interval (ITI) was followed by either a left or a right arrow symbol on the monitor (equal probability for both symbols) for a duration of .5 sec. These initial stimuli were followed, after a variable delay (stop signal delay; SSD), by an upward arrow in 20% of the trials and indicated that the participant should not respond. The VI varied as a function of the behavior of the participants: with a minimum of 250 ms, the VI increased by 50 ms after a successful response inhibition and decreased by 50 ms if any response was produced in a no-response trial.

2. Materials and Methods

2.1. Participants

We recruited 61 two-to-four weeks abstinent CUD participants, and 25 non-using HC participants with no lifetime history of SUD or other psychiatric disorders (see Table 1). Exclusion criteria for all participants were any major medical illness, an IQ below 70 as assessed by the Wechsler Test of Adult Reading 4, any neurological or active and non-substance related Axis I disorder, or use of any psychotropic medication (SUD comorbidity for the CUD group are reported in Table 1). All participants provided informed consent prior to the study. The research protocol was reviewed and approved by the Institutional Review Boards of the University of Texas Southwestern Medical Center and the Veterans Administration North Texas Health Care System.

Table 1.

Demographic data (mean values ± standard deviation)

	Healthy controls	CUD participants
		All CUD participants	Early-relapse	Late-relapse
Number of people	25	57†	26	30
Days to relapse	-	67 ± 67.97	8.04 ± 5.74	118.23 ± 53.43
Duration of treatment		25 ± 10.4	25.08 ± 7.65	25.55 ± 13.14
Age (Years)	39.52 ± 9.29	44.04 ± 6.82	43.38 ± 7.65	44.70 ± 6.18
Sex
Male	80 %	85.96 %	80.77 %	90.32 %
Female	20 %	14.04 %	19.23 %	9.68 %
Ethnic groups
African-American	56 %	71.93 %	76.92 %	80.77 %
Caucasian	32 %	24.56 %	23.08 %	30.77 %
Asian/other	8 %	0 %	0 %	0 %
Hispanic	4 %	3.51 %	0 %	7.69 %
Education (Years)	12.56 ± 1.76	13.05 ± 1.89	12.96 ± 1.68	13.17 ± 2.10
Number of people with other SUD
Barbiturate		4	2	2
Alcohol		39	18	20
Tobacco		19	12	7
Amphetamine		7	3	3
Cannabis		27	11	15
Cocaine Craving Questionnaire (CCQ)	-	2.49 ± 0.85	2.57 ± 0.86	2.46 ± 0.86
Cocaine use (Days)
Life time	-	3414.22 ± 2215.91	3426.61 ± 2299.52	3471.76 ± 2185.21
Previous 90 days	-	65.21 ± 27.09	61.69 ± 25.89	63.33 ± 28.77
Money spent for cocaine ($)	-
Life time	-	376068.93± 410219.51	436529.52 ± 517231.28	333971 ± 293126.06
Previous 90 days	-	6211.32 ± 5512.29	6246.12 ± 5553.98	6340.87 ± 5592.23

Note:

^†:The numbers of participants from the two subgroups do not add up because there is one CUD participant with missing relapse data.

The residential treatment program for CUD participants lasted two to four weeks (the precise number of days could only be retrieved for N=47 CUD participants, for a mean of 25 days and standard deviation of 10.4). This time period was chosen (1) to avoid the unstable withdrawal period during the first week of abstinence, (2) it is a relatively circumscribed time period just prior to Early Abstinence [reference DSM-IV], and (3) it was a realistic time period to have individuals in treatment remain on a residential unit. To the best of our knowledge, small variations in residential treatment do not affect subsequent time to relapse, therefore we considered only pre-determined clinical and demographic variables in our statistical analysis. The program utilized a Minnesota Model psychosocial treatment approach and it often incorporated cognitive-behavioral techniques, although no neurocognitive training was utilized. No other treatments known to directly influence neurocognitive functioning was provided. All participants remained on a supervised residential unit throughout their treatment; observed urine drug screens were conducted throughout residential treatment to verify abstinence. During the first two weeks, CUD participants underwent a comprehensive medical history and physical examination, a general laboratory panel, a urine substance screen, and a guided interview of lifetime substance use history (Kelly et al., 2009). The fMRI session took place immediately before CUD participants were discharged from treatment, to assure abstinence at the time of scanning. The fMRI sessions were conducted in the morning at approximately the same time of day for all subjects. Participants were asked to restrict caffeine and nicotine intake to at least one hour before the fMRI session to control for potential acute or withdrawal effects. After discharge, CUD participants were followed for 24 weeks, or until the first day of any use of cocaine or amphetamine, with two appointments per week: one by phone and one in person. A structured interview about substance use and the urine drug screen took place during the in-person clinical appointment. The “days to relapse” measure was defined a priori as either the first day of use of cocaine or amphetamine after discharge (self-report or urine drug screen) or the day of their first missed appointment if participants missed two consecutive appointments and all attempts to contact participant (e.g., phone, mail, collaterals) were unsuccessful.

All participants received $400 for the screening assessment, mock scan, fMRI, and neurocognitive testing. To maximize effort, participants were told they would receive an additional $15 if they did not perform very well on the fMRI task, $30 if they performed moderately well, and $45 if they performed extremely well. However, all participants received $30 regardless of their performance. Participants who participated in follow-up assessments received $25 for each follow-up visit and $10 for each follow-up phone call. Reimbursement was paid by gift card. To further encourage adherence with follow-up visits, these latter participants were also given tokens at each visit that would be selected at random for a prize, using the fish-bowl contingency management technique (Peirce et al., 2006; Walker et al., 2010).

We determined a priori to exclude subjects showing sustained head movement beyond 3 mm over any axis. Only one subject in the HC group was found to briefly pass this threshold, and no significant difference in head motion was found in comparing HC and CUD groups or early and late relapse groups, therefore we did not exclude any subject due to head motion. Four CUD participants were excluded due to corrupt data archive, which resulted in incomplete data, and one further CUD participant was excluded due to image artifacts. In relapse-related analyses, one CUD participant was also excluded due to the absence of follow-up data. Nine of the CUD participants did not relapse by the end of the follow-up period; their relapse data are here reported as equivalent to 168 days in correlation analyses (i.e. 24 weeks, the duration of the entire follow-up period). After the exclusions, the CUD group consisted of 57 individuals for the behavioral analysis (early relapse: 26 and late relapse:30) and 56 individuals for the fMRI analysis (early relapse: 25 and late relapse: 30).

2.2. Experimental Design

In the stop signal task (adapted from: Rubia et al., 2003), a left or right arrow (randomized) was presented on the screen as the “go signal” in randomly jittered inter-trial intervals of ~ 1.8s (1.6–2s, Fig. 1). The participants were tasked to respond as fast as possible by pressing a left or right button, accordingly, unless the left/right arrow was followed by an upward arrow (in 20% of the trials), which represented the “stop signal”. To ensure equal task difficulty (50% success rate of inhibition) across different participants, the variable interval between the initial left/right arrow and the stop signal (stop signal delay; SSD) changed idiosyncratically by a 50 ms step size, decreasing the interval if the probability of inhibition was below 50%, to make future inhibitions easier for the participant, or increasing the interval if the probability of inhibition was over 50%, to achieve the opposite effect. The task consisted of 362 trials: 288 go and 74 stop trials. The task was presented using the Presentation® software.

2.3. Behavioral Analysis

2.3.1. Model-agnostic behavioral analysis

We calculated the standard behavioral indices (Rubia et al., 2003) for each group 1) the mean length of the SSD; 2) the mean reaction times (RTs) to go trials and stop failures; 3) The stop-signal RT (SSRT) (Logan et al., 2014) for both groups was computed following the Integration Method (Verbruggen et al., 2019). First, we excluded the participants who violated the race model [cf. Recommendation 7, (Verbruggen et al., 2019)]; specifically, we excluded the participants that have longer RT on unsuccessful stop trials than on go trials. This screening procedure left us with 19 HC and 43 CUD participants for further SSRT analysis. In the Integration Method, the SSRTs were estimated by subtracting the mean delay between stimuli and signals (mean SSD) from the RT that corresponds to p(respond|signal) = 0.5. This RT was estimated by multiplying the number of go trials (including go trials with a choice error) by the conditional probability of p(respond|signal). In this estimation, the go trials on which the participants failed to respond were assigned with their maximum RT.

2.3.2. Computational modeling of behaviors

We used an ideal observer Bayesian model to estimate the trial-by-trial action-outcome distribution of probabilities or beliefs associated with the action performed. Given an environment (e.g. visual stimuli) and a set of possible interactions (e.g. press the left-right button), a belief represents the estimated likelihood an action produced in response to a perceived stimulus will result in a known outcome (Friston, 2010; Friston et al., 2016), denoted by p(θ|a):

$$p\left(\theta |a\right)=\frac{p\left(a|\theta \right)p\left(\theta \right)}{p\left(a\right)},$$

where a stands for the action and θ for the outcome, i.e. in the present task complying vs not-complying to the task instructions. In this equation, p(θ) represents the prior probability and p(θ|a) the posterior probability, which is updated depending on each observed outcome, given action a. We assumed participants were characterized by normally distributed, subject-specific, beliefs about the distribution of available events/interactions in the task environment p(a):

$$f(a|\sigma )=\frac{1}{\sigma \sqrt{2\pi }}{e}^{\frac{-{\left(a\right)}^2}{2{\sigma }^2}}$$

where a different value for the standard deviation (σ) was estimated for each participant to determine the pace of belief updates. The model estimated probabilities over two independent dimensions: 1) left/right and 2) go/stop trials, with visual stimuli used as evidence on a trial-by-trial basis to update beliefs about the action to perform (left vs right) and whether to perform an action or not (go vs stop). As a result, the model estimated the subject-specific trial-by-trial probability that each of the available actions (go-left, go-right, stop) would be correct. This estimated quantity determined the confidence (c) associated with the chosen action. The complement probability (1-c) was used to estimate the probability the chosen action was incorrect: this value was used to represent subject-specific trial-by-trial uncertainty in choice selections. The choice behavior of this task was simulated by the model with a mean prediction accuracy of 90.21±2.59% for HCs with mean σ_left/right = 0.27±0.05 and σ_go/stop = 0.44±0.07. The simulation for CUD participants had a mean accuracy of 89.05±5.12% with mean σ_left/right = 0.29±0.13 and σ_go/stop = 0.44±0.08.

2.4. fMRI data analysis

2.4.1. fMRI acquisition and preprocessing

Anatomical and functional images were collected on a Philips 3T MRI scanner. High-resolution T1-weighted anatomical images were acquired with a 3D magnetization prepared rapid gradient-echo (MPRAGE) sequence with a repetition time (TR) = 8.2 ms, an echo time (TE) = 3.8 ms, a flip angle = 12°, and a 1 × 1 × 1 mm³ spatial resolution. Functional images were acquired with a single-shot echo-planar imaging (EPI) sequence with a flip angle of 70°, a TR of 1700 ms, a TE of 25 ms, 36 slices, the field of view (FOV) = 208 × 208 mm², and the voxel size = 3.25 × 3.25 × 3.25 mm³. The functional scans were preprocessed using standard statistical parametric mapping (SPM12, Wellcome Department of Imaging Neuroscience; www.fil.ion.ucl.ac.uk/spm/) algorithms, including slice timing correction, co-registration, normalization with resampled voxel size of 2 mm × 2 mm × 2 mm, and smoothing with an 8 mm Gaussian kernel. A temporal high-pass filter of 128 Hz was applied to the fMRI data, and the temporal autocorrelation was modeled using a first-order autoregressive function.

2.4.2. Model-based fMRI data analysis

The first-level general linear model (GLM) included the onsets of button presses as a sensorimotor event (in the case of go trials and unsuccessful stop trials) and the onset of the stop signal (in the case of successful stop trials). Trial-by-trial uncertainty estimated from the Bayesian model was entered as a parametric modulator. The reaction time of each responded trial (or the onset of the stop signal in the successful stop trials) and six motion parameters were included as regressors-of-non-interest in first-level GLMs.

For the ROI analysis, we created masks using the peak coordinates of activation related to uncertainty (p < .05, corrected using the SPM default family-wise error [FWE], based on Gaussian random field theory) across healthy and CUD groups and examined within their respective CUD groups how these activations were related to relapse. ROI masks were generated as spheres with an 8-mm radius using the MarsBar toolbox (marsbar.sourceforge.net/). Peak coordinates are: bilateral AIC (X = 33/−33, Y = 20, Z = −4), dACC (X = 8, Y = 23, Z = 32), and bilateral VS (left: X = −12, Y = 11, Z = −8; right: X = 13, Y = 13, Z = −8). The mean parameter estimates of GLM were extracted for each participant from each ROI.

2.4.3. Dynamic causal modelling

We used SPM12 to estimate the effective connectivity among the three key ROIs of the AIC, dACC and VS. We relied on the same spherical, 8-mm radius, masks centered on peak of activity as described above for the model-based fMRI data analysis, limited to the left hemisphere (i.e. AIC: X = −33, Y = 20, Z = −4; dACC: X = −8, Y = 23, Z = 32; VS: X = −12, Y = 11, Z = −8). The left hemisphere was chosen for the DCM analysis because model-based results indicated an increased BOLD response to the signal of uncertainty-and therefore better signal-to noise ratio across participants-, in comparison with the right hemisphere (Figure 2). A further node was also added to include the visual cortex (VC) as a network input, with the ROI centered on coordinates (X = −24, Y = −76; Z = 26), again with 8-mm radius and spherical shape. Contrarily to the ROIs of AIC, dACC and VS, which are part of the network found active in association with the uncertainty signal and are included to estimate the causal relation among them, the VC is not included to test a specific hypothesis. Instead, this node serves the function to drive the activation of the circuit, when the visual stimulus is presented to the participants. Thus, the principal eigenvariates of these four ROIs were used to extract fMRI time series, using uncertainty-related parametrically modulated signal for the 3 ROIs of the AIC, dACC and VS, and non-modulated baseline activity for the VC.

Fig. 2. Dynamic causal modeling (DCM) model space.

The eight network architectures illustrate the tested DCM models, assuming a fully connected network for all models (illustrated by the presence of black and red arrows) and different connectivity targets for the modulatory signal of model-estimated uncertainty (red arrows). The eight structures illustrate different combinations of these targets of the modulatory signal, to allow for the pair-wise investigation of the directed connectivity among AIC, dACC, and VS. These eight models were grouped into different configuration of four models characterized by a single common directed connectivity, to allow for family-wise comparison. For example, models 1–4 and models 5–8 are respectively characterized by a modulated AIC-to-dACC or dACC-to-AIC connectivity, and varying directed connectivity involving the VS. A comparison between the former and the latter family would then indicate the complementary exceedance probability for AIC-to-dACC and dACC-to-AIC directed connectivity. Similar analyses have been run for the remaining pairs of regions of interest. AIC: anterior insular cortex; dACC: dorsal anterior cingulate cortex, VC: visual cortex; VS: ventral striatum.

Eight competing models (Fig. 2) were defined to be used for family-wise comparisons using a Bayesian random effect approach to determine the causal relations among the selected ROIs in uncertainty computations. This method relies on random effect Bayesian model comparison to estimate the probability that a model (or a family of models) has to explain (i.e. simulate/replicate) the changes in BOLD activity recorded for each participant. In this sense, each model embodies an hypothesis and DCM establishes a competition among all hypotheses to estimate their relative probability of being correct (Friston et al., 2003; Stephan et al., 2007; Stephan et al., 2010). This analysis is summarized by the exceedance probability of each model or family of models, representing how more likely each model or family is to generate the data of a randomly chosen subject, in comparison with any competing model or family.

We assumed a single input region in the VC and a fully connected network for the A-matrix (all nodes connected in both directions with all remaining nodes, plus self-connectivity), across all tested models. Eight different architectures were used for the B-matrix, determining differences in the connections targeted by the modulatory signal of the estimated Bayesian uncertainty. The modulatory signal was assumed to always target all the six connections - inwards and outwards - characterizing the VC, plus three connections (single direction) among the remaining three nodes in the network. The resulting eight configurations for the B-Matrix represented all possible target combinations (2 possible directions per 3 pairs of nodes: AIC-dACC, AIC-VS, dACC-VS), allowing grouping the models in families of four, each characterized by one common directed connectivity. Finally, these families were used in three pair-wise comparisons, to test which direction of the modulated connectivity would be characterized by the highest exceedance probability. For instance, the family including the four models characterized by an AIC-to-dACC connectivity was compared with the family characterized by the dACC-to-AIC connectivity.

2.5. Statistical methods

Given the unbalanced sample size, the group difference between HCs and CUD participants was examined by a one-tailed two-sample Welch’s t-test. Next, a correlation analysis was conducted to examine how neural activation and connectivity measures correlated with relapse time in CUD patients. To compute the exact p value associated with the correlation of coefficient (r), we used a non-parametric permutation test (Gu et al., 2012; Gu et al., 2015). In the permutation test, the null distribution of r was first obtained from 10,000 iterations of the permuted data and then used to compute the p value associated with the observed r. The α level was set at p = .05.

3. Results

3.1. Model-agnostic results

We followed previous studies (Rubia et al., 2003) in measuring for each group: 1) the mean length of SSD; 2) the mean RTs to go trials and stop failures; and 3) the mean SSRT for both groups, after subtracting the mean length of SSD from the mean RT to the “go trials”. No significant between-group differences were found in the mean length of SSD (Welch’s t₄₀ = 0.56, p = .58), in RTs to the go trials (Welch’s t₄₇ = 0.18, p = .86) or in SSRTs between the two groups (Welch’s t₃₀ = 0.67, p = .51). Similarly, in the comparison between early and late relapse CUD groups, we found no significant difference for the measures of SSD (Welch’s t₄₈ = 1.28, p =.21), RTs (Welch’s t₄₀ = 1.98, p=.054) and SSRTs (Welch’s t₄₀ = 0.76, p = .45). In conclusion, we found no observable group difference in conventional analyses of the stop signal task.

3.2. Uncertainty-related neural activations predict cocaine relapse

We found significant activation in the AIC, dACC (extending into the supplementary motor area), and VS that encoded our model-estimated choice uncertainty (Fig. 3a; Table 2) at p_FWE < .05. The group comparison between CUD participants and HCs revealed a significant difference in the right VS (Welch’s t₆₈ = 2.46, p = .02), but not in bilateral AIC (left: Welch’s t₆₈ = 1.96, p = .05; right: Welch’s t₆₀ = 1.54, p = .13), dACC (Welch’s t₅₂ = 1.92, p = .06), or left VS (left: Welch’s t₆₀ = 0.65, p = .52). Finally, we tested whether this response to uncertainty across multiple ROIs could be predictive of relapse time in CUD individuals. Extracted beta values from the ROIs in the dACC, left/right AIC, and left/right VS were found to be all significantly correlated with one another (r=[+.46 +.69], p≤.0005). Consistently, ROI analysis revealed that uncertainty-related activity positively correlated with participants’ days to relapse in left AIC (Fig. 3b; r = .276, p = .0182), dACC (Fig. 3c; r = .281, p = .0187) and bilateral VS (Fig. 3d–e; left: r = .301, p = .0139; right: r = .305, p = .0125).

Fig. 3. Uncertainty encoding in the stop signal task.

(a) Across all participants, whole-brain analysis showed uncertainty-related BOLD activation in dorsal anterior cingulate (dACC; extending into supplementary motor area), anterior insula (AIC), and ventral striatum (VS). The image shows BOLD activity FWE corrected p<.05, k=30. (b-e) Beta values extracted from the regions of interest (ROIs) using spherical volumes (r=8mm). This analysis showed significant positive correlations between days to relapse and uncertainty-related activation in (b) left AIC (peak coordinates: X = −33, Y = 20, Z = −4), (c) right dACC (coordinates: X = 8, Y = 23, Z = 32), (d) left VS (coordinates: X = −12, Y = 11, Z = −8), and (e) right VS (coordinates: X = 13, Y = 13, Z = −8), suggesting reduced encoding of uncertainty associated with early relapse.

Table 2.

Peaks of regions related to uncertainty in the stop signal task across all participants (p < .05 with FWE correction)

Coordinates
Region	L/R	x	y	Z	T	Z	k
Insula	R	36	20	−4	10.97	Inf	3268
	L	−48	14	−4	9.82	Inf	-
	L	−33	23	2	9.78	Inf	-
Dorsal anterior cingulate gyrus/supplementary motor area	R	9	14	56	10.17	Inf	-
	L	−3	20	56	9.31	7.64	-
	L	−3	11	50	8.88	7.38	-
Middle cingulate gyrus	R	6	23	38	10.08	Inf	-
Inferior frontal gyrus	R	51	17	8	9.2	7.57	-
Precentral gyrus	R	48	14	32	7.87	6.75	-
Inferior parietal lobule	R	33	−43	41	8.1	6.9	1197
Middle temporal gyrus	R	48	−61	2	7.59	6.57	-
	R	57	−49	5	6.6	5.88	-
	R	48	−46	14	6.56	5.85	-
Fusiform gyrus	R	39	−46	−16	7.15	6.27	-
	R	39	−58	−16	6.98	6.15	-
	R	30	−46	−16	6.48	5.79	-
Supramarginal gyrus	R	60	−28	35	7.05	6.2	-
Middle occipital gyrus	R	30	−67	32	7.03	6.18	-
Superior temporal gyrus	R	57	−43	23	6.83	6.04	-
Supramarginal gyrus	L	−54	−40	32	8.06	6.88	996
Middle temporal gyrus	L	−48	−64	11	7.84	6.73	-
	L	−42	−67	2	7.51	6.52	-
	L	−57	−49	14	6.83	6.05	-
Superior occipital gyrus	L	−24	−76	26	6.85	6.06	-
Fusiform gyrus	L	−39	−43	−19	6.13	5.53	-
	L	−30	−70	−13	5.79	5.27	-
	L	−36	−58	−16	5.78	5.27	-
Cuneus	L	−18	−76	35	5.98	5.42	-
Calcarine sulcus	L	−9	−70	8	5.99	5.43	36
Lingual gyrus	L	−18	−58	2	5.46	5.02	0
Superior frontal gyrus	L	−24	47	23	6.26	5.63	19

3.3. AIC-to-dACC connection protects against relapse risk

Next, we examined our central question of whether uncertainty-modulated AIC effective connectivity might be associated with relapse. Our family-wise DCM analysis revealed that the family of models where AIC drove dACC activity during uncertainty encoding had an exceedance probability of 98.63% in HCs (Fig. 4a). Similar family-wise analysis was used to estimate the exceedance probability associated with the remaining pairs of ROIs, revealing that uncertainty was more likely to modulate AIC-to-VS (90.55%) and VS-to-dACC (66.45%) directed connections in HCs (Fig. 4a).

Fig. 4. DCM family-wise comparison results.

(a) The family-wise DCM analysis showed that Bayesian uncertainty is more likely to modulate AIC-to-dACC directed connectivity (98.63% exceedance probability) in healthy controls. (b) In contrast, CUD participants showed a dominant dACC-to-AIC directed connectivity (72.76%). This reversal is not present in (c) the late-relapse group, which show a mixed result in terms of AIC-dACC connectivity, and it is driven by (d) the early-relapse group (with a dACC-to-AIC exceedance probability of 90.56%). Finally, panel (e) illustrates that stronger positive uncertainty-modulated AIC-to-dACC modulated connectivity is related to longer abstinence of cocaine (r = .245, p = .0385). Estimated values close to zero represent subjects for whom AIC-dACC connectivity is estimated to be weakly modulated by the signal of uncertainty.

In the CUD group, in sharp contrast, family-wise comparisons revealed that the signal of uncertainty primarily modulated the directed connectivity from dACC to AIC (72.76%; Fig. 4b). We also identified higher exceedance probabilities of models with AIC-to-VS (80.33%) and dACC-to-VS (92.28%) directed connectivity (Fig. 4b). To further identify which subgroup drove such pattern, we repeated our family-wise model comparisons after separating CUD early relapsers (<30 days) and CUD individuals with longer abstinence periods, as defined a priori (Adinoff et al., 2015; McHugh et al., 2016). This analysis showed that the reported reversed connection between AIC and dACC was primarily driven by the early-relapse group (dACC-to-AIC: 90.56%; Fig. 4d), in comparison with those who maintained abstinence beyond 30 days (dACC-to-AIC: 38.34%; Fig. 4c). Furthermore, the analysis of the two CUD subgroups also revealed differences in AIC-to-VS (early: 36.731%, late: 93.34%) and the dACC-to-VS connectivity (early: 97.91%, late: 47.49%). Notably, in both connectivity analyses the early relapse group showed the most distant estimates (of the two CUD groups), from HCs estimated connectivity.

These findings highlight a strong link between relapse and the inversion of directed connectivity between AIC and dACC during uncertainty encoding. They also suggest a putative protective role of the AIC-to-dACC connectivity during uncertainty encoding; if this is the case, we speculated we would observe longer abstinence time in those who indeed showed a stronger ACI-to-dACC connection. Therefore, we further tested the relationship between relapse time and subject-specific values estimated for the AIC-to-dACC modulated connectivity. This final analysis confirmed that CUD individuals with stronger positive AIC-to-dACC connection weights remained abstinent for longer periods (r = .245, p = .0385; Fig. 4e). Taken together, these results suggest a protective function of the directional influence from AIC to dACC in cocaine relapse under uncertainty.

4. Discussion

Drug addiction is marked by the repeated selection of suboptimal choices despite their adverse consequences (Volkow & Morales, 2015; Everitt & Robbins, 2016). A recent computational model suggests that one of the reasons that can lead an individual to systematically select these suboptimal policies is one’s inability to correctly generate an internal model of the environment (Ognibene et al., 2019). This theory postulates that addiction can emerge when environment complexity exceeds, either permanently or temporarily, the cognitive capabilities of an agent, leading to faulty estimations of action-outcome probabilistic associations. In terms of relapse after treatment, this theory predicts that the more these representations are impoverished, the higher the likelihood to reinstate the addictive behavior. Consistent with this prediction, our empirical findings revealed a direct relationship between the neural dynamics of an AIC-driven, uncertainty-computing circuitry and the duration of abstinence post treatment. Specifically, we found that early return to drug use in CUD individuals was predicted by reduced uncertainty-related responses in the AIC and its interconnected regions of the dACC and VS. Notably, the activity of the VS, though less frequently found in association with uncertainty (Zandbelt & Vink, 2010; Ersche et al., 2011; Mitchell et al., 2014), is well studied in association with SUD and CUD in particular, as a result of over-activation of mesolimbic dopamine signals (Everitt & Robbins, 2016; Koob & Volkow, 2016). Importantly, while healthy volunteers showed a predominantly AIC-to-dACC directed connectivity during uncertainty encoding, CUD individuals were characterized by the inverse dACC-to-AIC influential relation, which was mostly driven by early relapsers. Among CUD individuals, those who did have stronger AIC-to-dACC excitatory connections were more likely to remain abstinent for longer periods of time. Collectively, these findings pinpoint the AIC as a key node in driving an uncertainty-encoding neurocircuitry and protecting against relapse.

The insula, in particular the anterior insula, has been under the spotlight in addiction research (Garavan, 2010; Naqvi et al., 2014; Koob & Volkow, 2016) primarily due to its role in interoception (Wang et al., 2019; Livneh et al., 2020) and craving (Garavan, 2010). In a seminal study by Naqvi and colleagues (Naqvi et al., 2007), patients with insula damage showed reduced craving for smoking and greater likelihood for quitting, compared to those with lesions outside of the insula. Despite conflicting results from one other human lesion study (Bienkowski et al., 2010), the majority of animal models (Contreras et al., 2007; Forget et al., 2010) and human neuroimaging studies on this topic have independently supported a role of the insula in craving (Filbey et al., 2009; Gu et al., 2016). In these studies, insula activity is often considered ‘unhealthy’. For example, drug cues induce stronger neural responses in the insula than neutral cues; and consequently, by disrupting the insula, craving reduces and so do drug-seeking behaviors (Contreras et al., 2007; Naqvi et al., 2007; Forget et al., 2010). Different from these previous findings, our results suggest that stronger insula activity and directional influence towards the dACC are ‘healthy’ in the sense that they promote abstinence and protect against relapse.

We reconcile these seemingly disparate findings by reexamining another important line of work on the insula, centering on its role in cognition and decision-making. As introduced earlier, the insula has been consistently involved in the computation of risk, uncertainty, and conflict detection (Clark et al., 2008; Preuschoff et al., 2008; Bossaerts, 2010; Limongi et al., 2013). The central computational element modeled in our study is an individual’s estimation of decision uncertainty, or how likely it is that a choice will lead to any outcome but the desired outcome. This computational process is crucial for optimal decision-making (Fiorillo et al., 2003; Niv et al., 2005; Platt & Huettel, 2008; Rushworth & Behrens, 2008; Singer et al., 2009) and it is distinct from craving, which represents a subjective interoceptive state. Thus, neural responses in the insula might not be simply deemed ‘healthy’ or ‘unhealthy’; instead, the fitness and meaning of insula activity might crucially depend on the task context (e.g. uncertain choices or craving). When drug cues are present, too much insula activity could be detrimental as they signal interoceptive changes related to craving. But when the environment involves no drug stimuli, as is the case in the stop signal task (or in real life when CUD individuals just exited a treatment program), being able to monitor uncertainty requires strong involvement of the insula. In these situations, a dysfunctional AIC-neurocircuitry would then lead to mistakes in uncertainty estimations and therefore increased probability to select actions that are very likely to be associated with ‘unhealthy’ outcomes (Ognibene et al., 2019), such as deciding to return to drug-related environments and eventually, relapse.

We acknowledge a limitation in the interpretation of our neuro-cognitive findings in the lack of significant behavioral differences in the comparison between healthy control and CUD subjects. This is likely due to the structure of the stop signal task, which was designed to adjust dynamically the interval preceding the stop signal, so to maintain each participant’s response/inhibition accuracy to 50%,. This design resulted in a ceiling of behavioral performance, which led to similar behaviors across groups, even in presence of different underlying neural processes. We hypothesize that the observed neural differences would correlate with significant behavioral differences under conditions of increased complexity of choice behavior. The more the ramifications of each choice and the associated uncertainty increase, the higher the resources engaged by the model based system control and therefore the higher the chances to trigger divergent behaviors (cf. Ognibene et al., 2019). The choices leading abstinent SUD individuals to avoid relapse are indeed likely to require such high computational resources in considering the long lasting consequences associated with both relapse and abstinence. Future studies based on more ecologically plausible tasks may validate this interpretation, enhancing both behavioral and neuro-cognitive differences between groups, possibly also improving our current connectivity estimations for the group of late relapsers.

Conclusions

Our findings suggest that an insula-driven neurocircuitry that computes decision uncertainty can be an important mechanism and biomarker for CUD early relapse detection and intervention. These results could inform potential early detection of relapse (e.g. by measuring this circuitry during treatment) and intervention (e.g. by giving more intensive treatment and care to individuals with this vulnerability). These findings could have far-reaching implications for a wide range of addictive disorders beyond CUD. However, the study was designed specifically to assess cocaine use disorder, so leaving open a question about specificity of the results. Furthermore, the DCM estimations did not provide a clear picture concerning cortico-striatal relations. Thus, a different study design with separate cohorts of individuals for different primary substance use disorders will be needed to understand whether this mechanism is valid across drug use disorders and to provide a complete picture of the neuro-computational dynamics underlying choice uncertainty.

Abbreviation list

AIC

Anterior Insular Cortex

dACC

dorsal Anterior Cingulate Cortex

VS

Ventral Striatum

VC

Visual Cortex

SUD

Substance use disorder

CUD

Cocaine use disorder

HC

healthy control

RT

reaction time

SSD

stop signal delay

SSRT

stop-signal reaction time