Deep Learning Based Natural Language Processing Markers of Anxiety and Depression

项目摘要/摘要 严重抑郁障碍(MDD)和广泛性焦虑障碍(GAD)是主要的 造成全球健康负担的原因。MDD是导致残疾的主要原因，与道德水平的提高有关 风险，MDD和GAD都会导致相当大的经济成本、功能损失和质量下降 生活的一部分。响应当前人口水平筛查呼声的最大挑战之一是监测 大规模的MDD和GAD，同时最大限度地减少评估负担。然而，现有的评估方法， 依赖主观测量，基于诊断方法，在需要的程度上是繁重的 描述MDD和GAD的异质性，这将需要对所有 症状。需要新的方法来准确评估行为健康，克服监测障碍 和关怀，并促进对抑郁和焦虑的科学理解。 拟议的研究旨在通过将MDD和GAD解构为数字来解决这些差距 基于大型语言模型识别的语言特征的生物标志物(DB)。最先进的人工 智能和自然语言处理方法允许数据库的表征学习来自认知和 从语言信息中捕捉到的情感领域。虽然有效的、被动的和大规模的监控是 数据库的主要好处，我们还将使用它们来研究相关的研究领域标准(RDoC)， 包括负价系统反应和正价特性。研究的目标是：1)设计数据库 使用深度学习方法，通过训练基于注意力的语言模型 非常大的去身份心理治疗治疗成绩单语料库；2)检查初步表现 在高度特征化的MDD和GAD患者样本中，DB模型的可行性，并比较 结果与临床医生评分；3)探索基于研究范式的DB模型的改进 与RDoC结构一致，以进一步完善DB模型管道和未来在临床环境中的部署。 在这个以患者为导向的研究生涯中描述的研究和培训计划 发展奖旨在开发系统的数字健康方法，以允许维度 MDD和GAD的概念化与RDoC一致，增强了检测的简易性和一致性 最终支持有针对性的干预。拟议的项目得到了一个多学科的大力支持 团队包括Naomi Simon博士和Kyunghyun Cho博士的指导，以及Dr. 保罗·格里姆彻、蒂姆·阿尔索夫、哲·陈和坦泽姆·乔杜里。从奖项中获得的经验将 支持未来的R级研究，重点是先进的计算精神病学方法 进一步改进数据库模型，以改进对行为健康的被动和客观评估，并最终 提高我们对抑郁和焦虑的经验性理解。

PROJECT SUMMARY / ABSTRACT Major Depressive Disorder (MDD) and Generalized Anxiety Disorder (GAD) are among the primary causes of health burden worldwide. MDD is a leading cause of disability associated with increased morality risk, and both MDD and GAD result in considerable economic costs, loss of functioning, and decreased quality of life. One of the biggest challenges in responding to current calls for population-level screening is to monitor MDD and GAD at a large scale while minimizing assessment burden. Existing assessment methods, however, rely on subjective measures, are based on diagnostic approaches, and are burdensome in the extent needed to characterize MDD and GAD in their heterogeneity, which would require combined evaluation of all symptoms. New methods are needed to accurately assess behavioral health, overcome barriers to monitoring and care, and advance the scientific understanding of depression and anxiety. The proposed study aims to address these gaps by deconstructing MDD and GAD into Digital Biomarkers (DB) based on linguistic features identified by large language models. State of the art artificial intelligence and Natural Language Processing methods allow representation learning of DB from cognitive and emotional domains captured from linguistic information. While effective, passive, and at-scale monitoring are the primary benefits of DB, we will also use them to study relevant Research Domain Criteria (RDoC), including negative valence system reactions and positive valence traits. The study goals are to: 1) Design DB of MDD and GAD symptoms using deep learning methods, by training an attention-based language model on a very large corpus of de-identified psychotherapy treatment transcripts; 2) Examine preliminary performance and feasibility of the DB model in a highly characterized sample of MDD and GAD patients, and compare results with clinician ratings; 3) Explore improvements to the DB model based on research paradigms consistent with RDoC constructs, to further refine DB model pipeline and future deployment in clinical settings. The program of research and training described in this mentored patient-oriented research career development award is aimed at developing systematic digital health approaches to allow dimensional conceptualization of MDD and GAD consistent with RDoC, enhancing the ease and consistency of detection to ultimately support targeted interventions. The proposed project is strongly supported by a multidisciplinary team including the mentorship of Drs. Naomi Simon and Kyunghyun Cho, and the domain expertise of Drs. Paul Glimcher, Tim Althoff, Zhe Chen, and Tanzeem Choudhury. The experience gained from the award will enable the pursuit of future R-level studies focusing on advanced computational psychiatry approaches to further refine DB models to improve passive and objective assessment of behavioral health, and ultimately improve our empirical understanding of depression and anxiety.