Predictive and explanatory models of cigarette smoking: Computational approaches to understanding nicotine addiction

Abstract

Smoking is the leading cause of preventable death worldwide, causing 6 million deaths every year (WHO, 2011). Most people try smoking for the first time in adolescence (O?Loughlin et al., 2014), making this a critical period for research regarding risk factors for progressing into nicotine addiction. As with other substance use disorders, much is known about how nicotine-induced changes in neurotransmitter systems and sensitivity to drug- and non-drug rewards lead from recreational to habitual and finally to compulsive use. Differences in personality, life history, environment, behavioural responding, and neurobiology between non-smokers, smokers, and smokers who manage to quit are also known. However, there is very little evidence as to what pre-existing neurobiological factors make adolescents vulnerable to smoking behaviour. Using a large sample of 548 14-year old non-smokers, machine learning was used to predict smoking behaviour in the next four years. The analysis framework was chosen based on a rigorous empirical examination of 13 machine learning analysis pipelines for use with neuroimaging data. This revealed that the Elastic Net (Zou & Hastie, 2005), a form of regularized regression, allows use of large quantities of correlated variables in prediction models without the decline in accuracy seen with other approaches when large amounts of data are examined. Of the participants, 59 became regular smokers before age 16, and 33 became regular smokers before age 18. Using only MRI and fMRI data prediction accuracy was poor. Using personality, life history, psychopathology, substance use, and family/peer environment, classification into smoking trajectories was good with only a small reduction in accuracy when neuroimaging and non-imaging measures were combined. In line with previous research, behavioural and trait impulsivity were strong predictors of smoking. Extending previous knowledge, the ability of impulsivity metrics, particularly novelty-seeking, to predict smoking behaviour differed strongly by age of smoking onset. The absence of behavioural expressions of impulsivity such as conduct disorder was hypothesized to be a protective factor delaying onset of smoking. Using fMRI measures of reward processing, inhibitory control, affective processing, and mathematical and semantic processing, a predictive phenotype indicating risk for smoking onset between the ages of 16 and 18 emerged. The presence of deficits in processing of facial affect indicated by altered activity in regions including the temporal pole was a predictor of future smoking, and is consistent with previous accounts linking similar deficits to future binge-drinking behaviour (Whelan et al., 2014). Grey matter volume in the temporo-parietal junction and function of associated networks including the default-mode-network were also observed as risk factors for smoking. The primary finding regarding atypical brain function putting adolescents at risk for smoking was seen in networks underlying reward processing and cognitive control. Reduced sensitivity to cues signalling non-drug rewards in regions involved in attribution of saliency ? including the orbitofrontal cortex (OFC) and anterior cingulate, and increased activity in these regions upon receipt of a reward were strong indicators for long-term smoking risk. To further examine this effect, functional connectivity patterns of the reward system were examined in a sample of 206 14-year old adolescents who had already begun smoking. A Psychophysiological Interaction analysis coupled with Elastic Net regression revealed patterns of altered ventral striatum functional connectivity associated with lifetime frequency of smoking. Adolescents who had smoked more showed stronger functional connectivity between reward system nodes including the OFC and the ventral striatum. In addition, heightened smoking frequency was associated with lower functional connectivity between reward system nodes and regions involved in cognitive control and inhibition, including the right inferior frontal gyrus. To determine whether reward-related changes in cognitive control and sensitivity to rewarding stimuli would still be evident after adolescence and when using a different behavioural measure, a sample of adult smokers, non-smokers, and ex-smokers was recruited. Participants completed the Iowa Gambling Task (Bechara et al., 1994), which is known to engage the same brain regions for which smoking-related effects were found in the previous studies. Behavioural responses in this task indicate reinforcement learning, sensitivity to positive and negative feedback, and ability to anticipate future outcomes. These elements of task performance were quantified using computational models. As conclusively proving the validity of such models is challenging, EEG data from a separate sample was used to confirm the neurobiological validity of computational model calculations. Findings showed that both smokers and ex-smokers displayed strong preferences for immediate rewards with a disregard for the long-term negative consequences of choices. A phenotype characterized by reduced anticipatory sensitivity to non-drug reinforcers and increased attribution of salience when receiving non-drug rewards is suggested. This phenotype appears to put adolescents at risk for future smoking, has an association with smoking frequency, and persists in adulthood and after smoking cessation. The use of predictive modelling was shown to be a valuable tool to extend knowledge of aetiology and pathophysiology of maladaptive behaviour. The combination of neuroimaging and psychometric data made it possible to create a holistic model of smoking risk that took into account diverse facets of psychological, neurobiological, and environmental vulnerabilities. The neurobiological insights and behavioural indicators of decision-making and executive function identified here as risk factors for smoking behaviour have the potential to be translated into cognitive training or neurofeedback tools to be used in prevention or intervention efforts. Given the demonstration of the neurobiological validity of computational models of cognitive mechanisms, such tools could be used as cost-effective means of approximating reward system function in risk assessment or progress monitoring.