Computational Mechanisms of Approach-Avoidance Conflict Predictively Differentiate Between Affective and Substance Use Disorders

Abstract

Psychiatric disorders are highly heterogeneous and often co-morbid, posing specific challenges for effective treatment. Recently, computational modeling has emerged as a promising approach for characterizing sources of this heterogeneity, which could potentially aid in clinical differentiation. In this study, we tested whether computational mechanisms of decision-making under approach-avoidance conflict (AAC) - where behavior is expected to have both positive and negative outcomes - may have utility in this regard. We first carried out a set of pre-registered modeling analyses in a sample of 480 individuals who completed an established AAC task. These analyses aimed to replicate cross-sectional and longitudinal results from a prior dataset (N = 478) - suggesting that mechanisms of decision uncertainty (DU) and emotion conflict (EC) differentiate individuals with depression, anxiety, substance use disorders, and healthy comparisons. We then combined the prior and current datasets and employed a stacked machine learning approach to assess whether these computational measures could successfully perform out-of-sample classification between diagnostic groups. This revealed above-chance differentiation between affective and substance use disorders (balanced accuracy > 0.688), both in the presence and absence of co-morbidities. These results demonstrate the predictive utility of computational measures in characterizing distinct mechanisms of psychopathology and may point to novel treatment targets.

Keywords: Anxiety; Approach-Avoidance Conflict; Computational Modeling; Depression; Predictive Classification; Substance Use Disorders.

Figure 1

The approach-avoidance conflict (AAC) task. Bottom: Each trial is divided into a decision phase, an affective stimulus phase, and a reward phase. Trials are separated by a variable intertrial fixation time. Top: During the decision phase, participants choose to move an avatar to one of nine positions on a runway. Pictures are presented on each side of the runway, indicating the types of stimuli that could be presented during the affective stimulus and reward phases. The sun and cloud images represented potential positive and negative affective stimuli, respectively (each being an image–sound combination). The height of the red fill in a rectangle signified the number of points that would be received in the reward phase (ranging from 0 to 6 points). Participants were instructed that the final position of the avatar determined the probability of each of these outcomes occurring (in increments of 10%, from 90% to 10% with each step away from the associated stimulus indicator images). All choices therefore resulted in a probabilistic outcome. For example, if a participant chose the highest probability option on a given side (i.e., the runway position closest to their preferred outcome), there was a 90% chance that the preferred outcome would be presented. However, there was still a 10% chance that the non-preferred outcome associated with the alternative side of the runway would be presented instead. In the CONF6 condition above, for instance, the preferred outcome might be the combination of a negative affective stimulus and 6 points, while the alternative outcome would be a pleasant affective stimulus associated with no points. At the end of each trial (i.e., in the reward phase), participants were informed of the points won on that trial (i.e., including when 0 points were earned) as well as the total number of points they had acquired in the task thus far. The five trial types and associated probabilities of each outcome at each runway position are also shown above. The task consisted of 60 trials, with 12 of each of the five trial types.

Figure 2

Overview of study design and key objectives. The current study is part of the larger Tulsa 1000 project (Victor et al., 2018). This project collected data from two independent samples, an exploratory sample and a confirmatory sample (Ns shown in bottom left panel), both at a baseline time point and a 1-year follow-up. At each time point, participants completed (among other measures) an Approach Avoidance Conflict (AAC) task and a set of clinical and dimensional scales (top left panel). These included the following: the Patient Health Questionnaire-9 (PHQ; depression), the Overall Anxiety Severity and Impairment Scale (OASIS), the Drug Abuse Screening Test (DAST-10), the Anxiety Sensitivity Index (ASI), the Behavioral Activation/Inhibition scales (BIS/BAS), the Patient-Reported Outcome Measurement Information System (PROMIS) depression and anxiety scales, the Positive and Negative Affect Schedule (PANAS), the State-Trait Anxiety Inventory (STAI), and the Temporal Experience of Pleasure Scale (TEPS). For inclusion in the study, the clinical groups were required to meet specific criteria across the PHQ, OASIS, and DAST-10 (top left). Formal diagnostic grouping was then performed in accordance with DSM-IV or DSM-5 criteria using the Mini International Neuropsychiatric Interview (MINI), administered at baseline by trained professionals. Building on prior work with this dataset, the pre-registered aims of the current study focused on 1-year follow-up data in the confirmatory sample (top right panel). Namely, we tested the replicability of several longitudinal results previously reported on the exploratory dataset. These centered on to two parameters derived from computationally modeling AAC task behavior: decision uncertainty (DU) and emotion conflict (EC). Following completion of all pre-registered analyses, we then combined the exploratory and confirmatory samples, affording power to test additional questions (bottom right panel). Check marks and Xs indicate which of these analyses were and were not successfully replicated, respectively.

Figure 3

Comparison between exploratory and confirmatory samples. Top: The left panels compare DU values between groups in each sample at baseline and follow-up. The right panel displays the associated change scores (i.e., follow-up scores minus baseline scores). Here, DEP/ANX and SUDs showed lower DU in the confirmatory sample compared to the exploratory sample at baseline, while HCs showed higher DU in the confirmatory sample compared to the exploratory sample at follow-up. Reductions in DU over time were greater in the exploratory sample than in the confirmatory sample for HCs and DEP/ANX, but not for SUDs. Bottom: The left and right panels show equivalent group by sample comparisons for EC scores and change score comparisons, respectively. No sample differences in EC were observed for any groups at baseline or follow-up. HCs in the exploratory sample showed an increase in EC from baseline to follow-up, while participants in the confirmatory sample instead showed no change in EC values. *p < 0.05, **p < 0.01, ***p < 0.001.

Figure 4

Pre-registered tests of group, time, and sex effects on model parameters. Spaghetti plots (left) show changes from baseline to follow-up and therefore only include participants who returned for the follow-up visit. Bar plots (right) include all baseline participants, including those who did not return for follow-up. These plots illustrate individual differences in the stability of parameter values over time. Here, thick lines within the spaghetti plots indicate group means and lighter lines indicate individual values. Shaded areas around the line for each group mean reflect the associated standard errors. Sex comparisons in the lower bar graphs illustrate how observed group differences in EC were mainly driven by females. *p < 0.05, **p < 0.01, ***p < 0.001.

Figure 5

Effects of group and sex on emotion conflict (EC) across specific diagnostic categories in the combined sample. Values are shown separately for healthy comparisons (HCs) and each specific affective and substance use disorder. For each group, the number of participants is provided in the white box. Top: Differences between HCs and each specific diagnosis are consistent with differences seen in the broader clinical groupings. Middle: The overall group differences appear to be driven by females at both baseline and follow-up. Bottom: Males with affective and substance use disorders did not differ from HCs. MDD: Major Depressive Disorder; GAD: Generalized Anxiety Disorder; SAD: Social Anxiety Disorder; PD: Panic Disorder; PTSD: Post-Traumatic Stress Disorder; Alc.: Alcohol; Can.: Cannabis; Stim.: Stimulant; Op.: Opioid; Sed.: Sedative. For illustrative purposes, Bayesian t-tests were performed to evaluate evidence for the presence or absence of group differences with HCs at each time point. Bayes Factors (BFs) greater than 3 were taken to support the presence of an effect, denoted by black stars (*): *BF > 3; **BF > 10, ***BF > 30, ****BF > 100. BFs between 0.33 and 3 were considered equivocal null results. BFs less than 0.33 were taken to support the absence of a group difference with HCs, denoted by red Xs: ^xBF < 0.33, ^xxBF < 0.1, ^xxxBF < 0.033, ^xxxxBF < 0.001.

Figure 6

Effects of group and sex on decision uncertainty (DU) across specific diagnostic categories in the combined sample. Values are shown separately for healthy comparisons (HCs) and each specific affective and substance use disorder. For each group, the number of participants is provided in the white box. Top: Group differences (between HCs and each specific diagnosis) were consistent with differences seen in DEP/ANX and SUDs more broadly. Middle and Bottom: Male but not female SUDs displayed greater DU at baseline. MDD: Major Depressive Disorder; GAD: Generalized Anxiety Disorder; SAD: Social Anxiety Disorder; PD: Panic Disorder; PTSD: Post-Traumatic Stress Disorder; Alc.: Alcohol; Can.: Cannabis; Stim.: Stimulant; Op.: Opioid; Sed.: Sedative. For illustrative purposes, Bayesian t-tests were performed to evaluate evidence for the presence or absence of group differences with HCs at each time point. Bayes Factors (BFs) greater than 3 were taken to support the presence of an effect, denoted by black stars (*): *BF > 3; **BF > 10, ***BF > 30, ****BF > 100. BFs between 0.33 and 3 were considered equivocal null results. BFs less than 0.33 were taken to support the absence of a group difference with HCs, denoted by red Xs: ^xBF < 0.33, ^xxBF < 0.1, ^xxxBF < 0.033, ^xxxxBF < 0.001.

Figure 7

Predictive classification of diagnostic status in the combined clinical sample. In the combined clinical sample (excluding HCs), multiple stacked classification algorithms predictively classified (left panel) individuals with only affective disorders (DEP/ANX) vs. only substance use disorder (SUDs), (middle panel) individuals with vs. without SUDs, and (right panel) individuals with vs. without DEP/ANX. Top-left: Histograms show the number of participants in each group. Up-sampling was used to minimize model biases. Top-right: Confusion matrices illustrate classification accuracy for the winning model. Middle-left: Bar plots indicate variable importance (VI), which is the relative contribution each predictor made to the winning model. This metric was computed using the STACK approach to reflect the contribution of predictors across algorithms. Middle-right: Receiver operating characteristic (ROCs) curves are shown for the winning model. ROC curves provide information about true positive and false positive rates at various categorization thresholds. Associated area-under-the-curve (AUC) values show the probability of a sample being assigned to the correct group. Bottom: Model performance was evaluated using AUCs, balanced accuracy, sensitivity (true positive rate), and specificity (true negative rate). Balanced accuracy was used due to unbalanced sample sizes. Performance metrics were averaged across the five folds of cross validation. ENET had the highest balanced accuracy (highlighted) for each predictive classification and was chosen as the best performing algorithm. Algorithms: ENET = elastic net; KNN = k-nearest neighbors; ADABAG = Bagged AdaBoost.

Figure 2

Overview of study design and key objectives. The current study is part of the larger Tulsa 1000 project (Victor et al., 2018). This project collected data from two independent samples, an exploratory sample and a confirmatory sample (Ns shown in bottom left panel), both at a baseline time point and a 1-year follow-up. At each time point, participants completed (among other measures) an Approach Avoidance Conflict (AAC) task and a set of clinical and dimensional scales (top left panel). These included the following: the Patient Health Questionnaire (PHQ; depression), the Overall Anxiety Severity and Impairment Scale (OASIS), the Drug Abuse Screening Test (DAST-10), the Anxiety Sensitivity Index (ASI), the Behavioral Activation/Inhibition scales (BIS/BAS), the Patient-Reported Outcomes Measurement Information System (PROMIS) depression and anxiety scales, the Positive and Negative Affect Schedule (PANAS), the State-Trait Anxiety Inventory (STAI), and the Temporal Experience of Pleasure Scale (TEPS). For inclusion in the study, the clinical groups were required to meet specific criteria across the PHQ, OASIS, and DAST-10 (top left). Formal diagnostic grouping was then performed in accordance with DSM-IV or DSM-5 criteria using the Mini International Neuropsychiatric Interview (MINI), administered at baseline by trained professionals. Building on prior work with this dataset, the pre-registered aims of the current study focused on 1-year follow-up data in the confirmatory sample (top right panel). Namely, we tested the replicability of several longitudinal results previously reported on the exploratory dataset. These centered on to two parameters derived from computationally modeling AAC task behavior: decision uncertainty (DU) and emotion conflict (EC). Following completion of all pre-registered analyses, we then combined the exploratory and confirmatory samples, affording power to test additional questions (bottom right panel). Check marks and Xs indicate which of these analyses were and were not successfully replicated, respectively.

Figure 1

The approach-avoidance conflict (AAC) task. Bottom: Each trial is divided into a decision phase, an affective stimulus phase, and a reward phase. Trials are separated by a variable intertrial fixation time. Top: During the decision phase, participants choose to move an avatar to one of nine positions on a runway. Pictures are presented on each side of the runway, indicating the types of stimuli that could be presented during the affective stimulus and reward phases. The sun and cloud images represented potential positive and negative affective stimuli, respectively (each being an image–sound combination). The height of the red fill in a rectangle signified the number of points that would be received in the reward phase (ranging from 0 to 6 points). Participants were instructed that the final position of the avatar determined the probability of each of these outcomes occurring (in increments of 10%, from 90% to 10% with each step away from the associated stimulus indicator images). All choices therefore resulted in a probabilistic outcome. For example, if a participant chose the highest probability option on a given side (i.e., the runway position closest to their preferred outcome), there was a 90% chance that the preferred outcome would be presented. However, there was still a 10% chance that the non-preferred outcome associated with the alternative side of the runway would be presented instead. In the CONF6 condition above, for instance, the preferred outcome might be the combination of a negative affective stimulus and 6 points, while the alternative outcome would be a pleasant affective stimulus associated with no points. At the end of each trial (i.e., in the reward phase), participants were informed of the points won on that trial (i.e., including when 0 points were earned) as well as the total number of points they had acquired in the task thus far. The five trial types and associated probabilities of each outcome at each runway position are also shown above. The task consisted of 60 trials, with 12 of each of the five trial types.

Supplementary Figure 3

Association between model parameters (top: DU and bottom: EC) and (left) select self-report questionnaire items (Q2 and Q3 in Table 5, main text) and (right) response time (RTs) in approach (top) and avoid (bottom) conditions.

Figure 4

Pre-registered tests of group, time, and sex effects on model parameters. Spaghetti plots (left) show changes from baseline to follow-up and therefore only include participants who returned for the follow-up visit. Bar plots (right) include all baseline participants, including those who did not return for follow-up. These plots illustrate individual differences in the stability of parameter values over time. Here, thick lines within the spaghetti plots indicate group means and lighter lines indicate individual values. Shaded areas around the line for each group mean reflect the associated standard errors. Sex comparisons in the lower bar graphs illustrate how observed group differences in EC were mainly driven by females. *p < 0.05, **p < 0.01, ***p < 0.001.