社会与工程心理学研究室知识库

药物成瘾是一种具有高社会危害性的慢性复发性脑疾病，但目前却缺少有效的诊断和治疗方法。在药物成瘾过程中大脑的奖赏系统和控制系统遭到破坏，并影响多种眼动行为，形成独特的可测量的信号，这种眼动特征有潜力成为成瘾诊断的客观标准。成瘾者的眼动特征反映在静息态眼动模式和线索诱发的眼动模式两方面。静息态，即无注视目标的条件下，瞳孔对光线变化的调节能力反映了视网膜的多巴胺神经元的受损情况；线索诱发条件下，即对成瘾物质相关事物的注 视、瞳孔等特征，反映了注意偏向。但目前缺少研究在自然状态下对成瘾者的静息态眼动模式和线索诱发的眼动模式进行全面的探索。因此本研究计划利用在虚 拟现实环境下的眼动追踪技术，对轻重度药物成瘾者和健康个体的两种眼动模式进行测量，探究成瘾造成的脑损伤在眼动上的表现形式，并通过机器学习算法利用轻重度成瘾者和健康个体眼动不同的眼动模式对他们进行分类和预测，从而发现能够辅助药物成瘾的诊断和治疗的客观标准。

本研究共纳入了 405 名被试（药物成瘾者 275 名，健康对照 130 名）。首先，针对长期使用成瘾性药物可能导致多巴胺受体下调的假设，我们进行了静息态眼动特征测量，包含 3 种亮度（0.04lux、4lux 和 50lux）下短时光刺激和稳定光刺激下瞳孔光反应，并计算了瞳孔光反应潜伏、最小瞳孔直径、收缩幅度、收缩率、收缩时间和收缩速度以及稳定的瞳孔直径作为眼动指标。另外，我们记录了被试自由浏览包含药物相关线索的 VR 便利店时的眼动数据，并计算了在这个过程中被试对 5 类线索（近端药物线索、远端药物线索、烟酒、自然奖赏和普通商品）的多种眼动指标，包含首次注视时间、凝视时间、注视次数、注视时间、回视次 数、回视时间和瞳孔直径。最后，在对两种眼动模式下的眼动数据进行数据清洗 后首先使用递归特征消除和交叉验证的方法进行了特征选择，然后利用网格搜索进行参数调优，之后使用三种机器学习算法（逻辑回归、支持向量机和随机森林）对是否成瘾（健康 vs. 成瘾）、成瘾程度（轻度 vs. 重度）和使用的药物类型（甲 基苯丙胺 vs. 海洛因）进行二元分类，另外也利用随机森林算法对是否成瘾及成 瘾程度进行三元分类。

实验结果显示，成瘾被试在暴露于 50lux 短时光照刺激时表现出较小的瞳孔反应，而在 50lux 稳定的光照刺激下成瘾被试的瞳孔直径与黑暗环境中的差值显著大于健康被试。这提示成瘾个体的光敏感性受到了损害，光适应能力显著减弱，验证了药物成瘾对多巴胺系统的影响。另外，在 VR 便利店中成瘾者对药物线索表现出更高的关注度和认知资源投入，尤其是重度成瘾者，成瘾人群的瞳孔扩大程度明显高于健康人群，这些数据支持了药物成瘾引发的无意识的注意偏向假设。最后，针对个体是否成瘾的分类，支持向量机模型表现最佳，具有 82.22%的分类准确率，85.12%的精确度，89.45%的召回率，0.88 的 ROC 曲线下的面积和 0.94 的 PR 曲线下的面积，相较于逻辑回归和随机森林模型有显著优势。针对健康、轻度成瘾和重度成瘾的区分，随机森林模型能够较好地区分出健康组，具有 74.51% 的精确度，但对于轻度和重度成瘾的区分效果不佳，精确度分别只有 59.09%和 63.04%。进一步针对成瘾者的两种成瘾程度建立二分类模型，结果显示支持向量机分类器的性能更好，准确率为 74%，ROC 曲线下面积达到了0.78，这一结果为成瘾程度的准确识别提供了有力支持。然而，对于不同成瘾物质的成瘾者进行分类时，模型效果不理想，最佳准确率（支持向量机模型）只有 69%，说明单独使用本研究中测量的眼动指标难以区分使用不同成瘾物质的人群。

综上所述，本研究为理解药物成瘾的认知机制和生物标志物提供了重要线索，为成瘾诊断和治疗提供了新思路。未来研究可探索更多模态融合特征和新算法，提高成瘾状态和程度的准确识别率。

Drug addiction is a chronic relapsing brain disorder with high societal harm, yet effective diagnostic and therapeutic methods are currently lacking. The reward and control systems of the brain are disrupted during the addiction process, affecting various eye movement behaviors, thereby forming unique measurable signals. These eye movement characteristics have the potential to become objective criteria for addiction diagnosis. Eye movement characteristics of addicts are reflected in both resting-state eye movement patterns and cue-induced eye movement patterns. In the resting state, the regulation of pupil size in response to changes in light reflects the damage to dopamine neurons in the retina. Under cue-induced conditions, such as viewing addiction-related stimuli, features like gaze and pupil responses reflect attentional biases. However, comprehensive exploration of resting-state eye movement patterns and cue-induced eye movement patterns in addicts in natural states is currently lacking. Therefore, this study plans to use eye tracking technology in virtual reality environments to measure two types of eye movement patterns in individuals with mild and severe drug addiction and healthy individuals. It aims to explore the manifestations of brain damage caused by addiction in eye movements and classify and predict addicts and healthy individuals based on their different eye movement patterns using machine learning algorithms, thus discovering objective criteria to assist in the diagnosis and treatment of drug addiction.

A total of 405 participants (275 drug addicts and 130 healthy controls) were included in this study. Firstly, based on the hypothesis that long-term use of addictive drugs may lead to downregulation of dopamine receptors, we measured resting-state eye movement characteristics, including pupil light responses under short-term light stimuli and stable light stimuli of three intensities (0.04lux, 4lux and 50lux), and calculated pupil light response latency, minimum pupil diameter, constriction amplitude, constriction rate, constriction time, constriction velocity, and stable pupil diameter as eye movement indicators. Furthermore, we recorded participants' eye movement data while freely browsing a virtual reality convenience store containing drug-related cues and calculated various eye movement indicators for five categories of cues (proximal drug cues, distal drug cues, tobacco and alcohol cues, natural rewards, and common goods), including initial fixation time, gaze time, fixation count, fixation time, return count, return time, and pupil diameter. Finally, after cleaning the eye movement data for the two types of eye movement patterns, we first performed feature selection using recursive feature elimination and cross-validation, then optimized parameters using grid search, and used three machine learning algorithms (logistic regression, support vector machine, and random forest) to perform binary classification for addiction (healthy vs. addicted), addiction severity (mild vs. severe), and type of drug used (methamphetamine vs. heroin). Additionally, a random forest algorithm was used for ternary classification of addiction and addiction severity.

The results showed that addicted participants exhibited smaller pupil responses when exposed to 50 lux short-term light stimuli, while the difference in pupil diameter between addicted participants and healthy participants under stable 50 lux light stimuli was significantly greater than in a dark environment. This suggests that addicts' photosensitivity is impaired, and their light adaptation ability is significantly reduced, confirming the impact of drug addiction on the dopamine system. Moreover, addicts showed higher attention and cognitive resource allocation to drug cues in the virtual reality convenience store, especially severe addicts, with addicts' pupil dilation being significantly higher than that of healthy individuals, supporting the hypothesis of unconscious attentional bias induced by drug addiction. Lastly, for individual addiction classification, the support vector machine model performed the best, with an accuracy of 82.22%, precision of 85.12%, recall of 89.45%, an area under the ROC curve of 0.88, and an area under the PR curve of 0.94, showing significant advantages over logistic regression and random forest models. For distinguishing between healthy, mild addiction, and severe addiction, the random forest model could effectively distinguish the healthy group with a precision of 74.51%, but had poor performance in distinguishing between mild and severe addiction, with precisions of only 59.09% and 63.04%, respectively. Furthermore, for binary classification of addiction severity, the support vector machine classifier performed better, with an accuracy of 74% and an area under the ROC curve of 0.78, providing strong support for accurate identification of addiction severity. However, when classifying addicts using different addictive substances, the model performed poorly, with the best accuracy (support vector machine model) being only 69%, indicating that it is difficult to distinguish between populations using different addictive substances solely based on the eye movement indicators measured in this study.

In conclusion, this study provides important clues for understanding the cognitive mechanisms and biomarkers of drug addiction, offering new insights for addiction diagnosis and treatment. Future research could explore more feature fusion modalities and new algorithms to improve the accuracy of addiction status and severity identification.