| dc.description.abstract | Substance dependence is a leading global health concern. Alcohol use is the primary risk-factor for deaths and disability-adjusted life years among those between the ages of 15-49 (Griswold et al., 2018). Opioid dependence accounts for the majority of drug use disorders worldwide (James et al., 2018), while tobacco-use alone results in over one in ten deaths per year (Reitsma et al., 2017). The chronically relapsing nature of substance-dependence renders challenges for successful abstinence, with some studies reporting more than two thirds of individuals relapse within months of treatment initiation (Moeller & Paulus, 2018; Sinha, 2011). A key characteristic of substance-dependence is the persistent and continued use of substances despite their negative consequences. Therefore, understanding the psychological and neurobiological mechanisms that give rise to such maladaptive decision-making is a key goal for the field of addiction science. Converging evidence suggests that aberrant value-based decision-making (e.g., reward processing) characterises substance-dependent individuals from healthy controls, and may predict future use. However, it remains unclear whether different substance types, misuse patterns, and treatment interventions differentially affect decision-making impairments.
This thesis investigated value-based decision-making in various substance-dependence phenotypes. Specifically, computational models of decision-making (reinforcement learning and drift-diffusion models; RLDDMs) were fit to choice and reaction time data from a popular reward learning task known to index fluctuations dopaminergic functioning and show sensitivity to various clinical disorders (Frank et al., 2004). Machine learning (ML) methods were leveraged to investigate if parameters derived from the computational models could successfully predict substance-dependence. Four ML models were compared: (i) a Summary model with mean choice accuracy from the reward learning task, (ii) a Computational model with parameter estimates from RLDDMs, (iii) a Personality model with self-reported impulsivity, and (iv) a Combined model with features from (ii) and (iii). Additionally, each individual model was compared with a null model including demographic features.
The literature on value-based decision-making and substance dependence was reviewed in Chapter 1. Chapter 2 described the general methods. Chapter 3 sought to predict hazardous alcohol-use risk (N=115). Chapter 4 sought to predict length of abstinence in a sample formerly dependent on heroin currently in methadone maintenance treatment (N=81). Chapter 5 sought to classify individuals based on their smoking group category (non-smokers, current smokers, ex-smokers, and vapers; N = 173). Chapter 6 aimed to assess behavioural and electrophysiological longitudinal changes in value-based decision-making during a smoking quit attempt (N = 112).
The results revealed reductions in response caution (indexed by the boundary separation parameter in drift-diffusion models) significantly predicted higher alcohol misuse risk, shorter lengths of opioid abstinence, and smoker versus non-smoker group membership. These findings suggest that response caution may be a task-general marker of substance-dependence that is sensitive to length of methadone treatment. Efficiency of evidence accumulation (i.e., the process of accumulating evidence for one option relative to another; drift-rate) was also a significant predictor across studies, however did not show a clear directional relationship with substance dependence, and was influenced by conflict in reward probabilities. For example, ex-smokers were classified by reduced evidence accumulation for high conflict trials with stimuli associated with negative feedback, and increased evidence accumulation for high conflict trials with stimuli associated with positive feedback.
Computational models performed similarly with Personality and Summary models, and outperformed demographic models overall. This indicates that self-reported impulsivity and mean choices in the decision-making task were as predictive compared with computational models fit to trial-by-trial choice and RT data. Overall, these findings highlight the utility of RLDDMs to investigate clinically relevant features of instrumental learning and decision-making, and identify features of value-based decision-making (i.e. evidence accumulation and decision threshold) that are predictive of substance-dependence. | en |
