Computational Mechanisms of Learning and Forgetting Differentiate Affective and Substance Use Disorders

Abstract

Depression and anxiety are common, highly co-morbid conditions associated with a range of learning and decision-making deficits. While the computational mechanisms underlying these deficits have received growing attention, the transdiagnostic vs. diagnosis-specific nature of these mechanisms remains insufficiently characterized. Individuals with affective disorders (iADs; i.e., depression with or without co-morbid anxiety; N=168 and 74, respectively) completed a widely-used decision-making task. To establish diagnostic specificity, we also incorporated data from a sample of individuals with substance use disorders (iSUDs; N=147) and healthy comparisons (HCs; N=54). Computational modeling afforded separate measures of learning and forgetting rates, among other parameters. Compared to HCs, forgetting rates (reflecting recency bias) were elevated in both iADs and iSUDs (p = 0.007, η ² = 0.022). In contrast, iADs showed faster learning rates for negative outcomes than iSUDs (p = 0.027, η ² = 0.017), but they did not differ from HCs. Other model parameters associated with learning and information-seeking also showed suggestive relationships with early adversity and impulsivity. Our findings demonstrate distinct differences in learning and forgetting rates between iSUDs, iADs, and HCs, suggesting that different cognitive processes are affected in these conditions. These differences in decision-making processes and their correlations with symptom dimensions suggest that one could specifically develop interventions that target changing forgetting rates and/or learning from negative outcomes. These results pave the way for replication studies to confirm these relationships and establish their clinical implications.

Keywords: active inference; affective disorders; computational psychiatry; decision-making; explore-exploit dilemma; reinforcement learning; substance use disorders.

Figure 1.

Task structure and computational model. a) The winning computational (active inference) model included 5 parameters: Initial Decision Noise (β₀), Learning Rates for Wins (η_win) and Losses (η_loss), Forgetting Rate (ω), and Pessimism (ρ_p). Higher β₀ values decrease the initial precision of the probability distribution from which actions are sampled (π). This value then changes across trials, leading to noisier choice after unexpected negative outcomes and less noisy choice after positive outcomes. Higher learning rates increase the degree to which confidence increases about a particular action-outcome probability after each new win or loss. Higher ω values decrease confidence in current beliefs about all action-outcome probabilities after each trial, leading to a recency bias in learning. Higher ρ_p values deter exploratory choices, as they entail that unchosen actions will likely lead to negative outcomes. Note that, in the equations shown, A is a matrix of action-outcome probabilities, which corresponds to the normalized columns of a-matrix. This latter matrix counts the number of times each outcome has been observed after each action (where these counts are scaled by learning and forgetting rates). G_π assigns lower values to actions expected to both generate more reward and lead to greater reductions in uncertainty about action-outcome probabilities. F_π quantifies the unexpectedness of an outcome under a chosen action, while G_error quantifies how strongly beliefs about the best action change after a new observation. Note that σ indicates a softmax function that converts a vector of values into a proportional probability distribution. ⨂ represents matrix multiplication. b) Three-armed bandit task interface. Participants played 20 games of 16 trials each. c-e) Correlations between generative and estimated parameter values in simulations, demonstrating recoverability of each parameter in the winning model.

Figure 2.

Model-free correlations and group differences in computational model parameters. a) Correlations between model parameters and model-free measures. b-d) Results of Parametric Empirical Bayes (PEB) analyses, including posterior means and credible intervals for effects of group (i.e., when accounting for effects of age and sex). Positive numbers indicate greater values in iADs relative to HCs, iSUDs relative to HCs, and iADs relative to iSUDs. Learning and forgetting rates are in logit space, while all other parameters are in log space. e-i) Top: Mean and standard error for computational parameter values in each group. Stars indicate significant differences in linear models reported in the main text (*p<.05, **p<.01, **p<.001). Bottom: Raincloud plot of computational model parameters for each group. RT=mean reaction time.

Figure 3.

Association between latent factors and computational measures. a) Heatmap of factor loadings across measures. The 7 factors were interpreted as reflecting: Negative Affect, Positive Affect, Physical Functioning, Interoceptive Awareness, Early Adversity, Impulsivity, and Working Memory. Detailed descriptions of each measure are provided in Supplementary Materials. b-e) Significant relationships between factors and computational parameters. Early Adversity was positively correlated with learning rate for losses. Impulsivity was negatively correlated with pessimism. Working memory was negatively correlated with the Pearson correlation (r) values for within-subject associations between reaction times and belief update magnitudes. Note that, while significant at uncorrected thresholds for exploratory purposes, none of these relationships survived correction for 35 comparisons (p<.0014). f-g) Bar graphs showing the average and standard error of Pearson correlation values reflecting within-subject correlations between reaction times and both belief update magnitudes and choice uncertainty. *p<.05, **p<.01, **p<.001, uncorrected.