关键词: 社交焦虑障碍, 认知行为疗法, 神经预测因子, 机器学习, 个体化预测

Abstract: Objective: Social anxiety disorder (SAD) is a common chronic psychological disorder and an important cause of depression and autism. Cognitive behavioural therapy (CBT) is the gold-standard psychological treatment for SAD; however, some patients fail to respond to it. Therefore, identifying neurobiological predictors of the treatment response to CBT has implications for precision medicine. Methods: This paper systematically reviewed the relevant studies on neural predictors of individual responses to CBT in SAD in the core collection of Web of Science, MEDLINE, SciELO, PubMed and PsychoInfo databases from January 2013 to July 2020. The study samples met the diagnostic criteria for SAD in the diagnostic system of mental disorders. SAD was taken as the main symptom, and patients with other comorbid mood or anxiety disorders (such as mood disorders and generalised anxiety disorders) were not strictly excluded, but the participants had no history of major mental disorders (such as schizophrenia and bipolar disorder). Results: Preliminary evidence suggests that functional activation of the higher-order visual cortex (e.g. the dorsal and ventral occipito-temporal cortex and superior and middle temporal gyrus), the dorsal anterior cingulate cortex and parts of the prefrontal cortex (e.g. dorsolateral prefrontal cortex, dorsomedial prefrontal cortex and medial orbitofrontal cortex) may be potential predictors of the success of CBT treatment for SAD. In contrast, the predictive effects of activation of the amygdala and insula were not consistent across studies. The structural or functional connections between the amygdala and areas associated with emotional regulation, such as the prefrontal cortex, predicted therapeutic responses. There were few electrophysiological studies, and preliminary studies found that larger late positive potential (LPP) amplitude for aversive distractors is associated with more significant symptom improvement after CBT. No potential predictors were identified for brain structure, genetics, demography or clinical variables except for initial anxiety. Studies showed that neuroimaging predictors are superior to demographic and clinical indicators, which indicates the necessity of searching for neural predictors. Using machine learning methods, researchers made individualised-level predictions for the suitability of individuals with SAD for CBT therapy. Support vector machine and logistic regression were two of the most commonly used machine learning algorithms. The accuracy of prediction was between 69% and 92%, and the sample size was about 20–50 people. However, for most of the studies, feature selection or feature extraction methods were not combined to further optimise the classification performance of machine learning. In addition, most of the research used cross-validation to investigate the generalisability of the model. In this process, over-fitting caused by model comparison and parameter selection may also occur. Conclusion: Great progress has been made in the study of neural predictors for the treatment response to CBT for individuals with SAD, which will be helpful for the development of personalised treatment plans and the effective use of medical resources. Future research should consider cooperation and sharing among research institutions to obtain big data, conduct data collection under multi-modal and multi-task conditions, further enhance the ecology of experimental tasks and verify the effectiveness of individualised-level prediction in independent samples.

Key words: social anxiety disorder, cognitive behavioral therapy, neural predictors, machine learning, individual prediction