Identification and Prediction of Peripartum Depression from Natural Language Collected in a Mobile Health App

项目总结 背景：妊娠期和产后期抑郁影响高达15%的美国母亲，给 母亲、孩子和社会。及早发现可以显著减少抑郁的发生率，但抑郁症状 在产前检查期间经常被遗漏，产前检查往往侧重于孕产妇和胎儿的身体健康，留下的时间较少 产妇的心理健康。即使在产前护理中讨论了心理健康问题，女性可能也不会舒服地回答 被认为令人尴尬或冒犯他人的问题。更有可能的是，在 产后期，因为婴儿出生后很少去看医生。以日记的形式进行测量， 可以使用自然语言处理进行分析，可以促进更早和更频繁地发现抑郁症 在怀孕期间和产后期间。 研究目的：1)语言使用的动态特征随时间的变化最能预测抑郁状态变化的模型 孕期和产后期，产生抑郁风险的表型；2)检查语言模式是如何 预测非裔美国人和白人女性抑郁的不同；以及3)确定 围产期抑郁妇女的言语和求医行为特征。 创新：拟议的研究在使用高频自然语言测量方面具有创新性， 使用智能手机应用程序结合自然语言处理模型的进步来评估发病情况的每日日记 以及孕期和产后期抑郁的轨迹。这是第一个前瞻性的纵向研究。 在临床人群中使用自然语言集合进行风险预测，首先要：1)描述关键主题 女性在围产期使用开放式期刊讨论一段时间；2)评估语言的多个方面 为了更全面地了解语言和抑郁之间的关系；3)使用纵向 设计方法，允许对与抑郁症发病相关的语言变化进行最佳建模。 方法和预期结果：根据爱丁堡产后抑郁量表确定每月抑郁风险。 将通过MyHealthy Pregnancy智能手机应用程序收集，这是一款通过Close开发的移动健康应用程序 决策科学家、临床医生、统计学家和当地围产期妇女之间的合作。一份嵌入的日报 在MyHealthy Pregnancy应用程序中，将收集参与者的自然语言文本10个月(从他们的第一个月开始 产后两个月的产前访视)。使用三种不同的自然语言处理算法方法， 这项研究将描述围产期妇女在日常日记中使用的自然语言是如何联系起来的 围产期抑郁的发生和经历，通过每月给药的抑郁量表进行测量。团体- 然后，基于轨迹的建模将根据女性抑郁得分随时间的变化模式对女性进行分类。 潜在影响：这项工作为开发和评估实时干预奠定了基础 大规模部署给使用预示着高抑郁风险的语言的女性。

PROJECT SUMMARY Background: Depression during pregnancy and the postpartum period affects up to 15% of US mothers, imposing costs on mother, child, and society. Early detection can significantly reduce the incidence of depression, yet depressive symptoms are often missed during prenatal visits, which tend to focus on maternal and fetal physical health, leaving less time for maternal mental health. Even if mental health is addressed during prenatal care, women may not feel comfortable answering questions that are perceived to be embarrassing or invasive. Failing to detect depression is even more likely during the postpartum period due to infrequent physician visits once the baby has been born. Measurement in the form of daily journals, which can be analyzed using natural language processing, can promote early and more frequent detection of depression during pregnancy and the postpartum period. Study Aims: 1) Model which dynamic features of language used over time best predict changes in depression status in the pregnancy and postpartum periods, creating phenotypes of depression risk; 2) examine how the language patterns that predict depression differ for African-American and Caucasian women; and 3) identify the relationship between the characteristics of what depressed peripartum women say and their treatment-seeking behavior. Innovation: The proposed research is innovative in its use of high frequency natural language measurements, captured in daily journals using a smartphone app, combined with advances in natural language processing models, to assess the onset and trajectory of depression during pregnancy and the postpartum period. This is the first prospective longitudinal study using natural language collection for risk prediction in a clinical population and the first to: 1) characterize the critical topics women discuss during the peripartum period over time using open-ended journals; 2) evaluate multiple facets of language to gain a more comprehensive understanding of the relationship between language and depression; 3) use a longitudinal design approach allowing for optimal modeling of language changes associated with depression onset. Methodology and Expected Results: Monthly depression risk identified from the Edinburgh Postnatal Depression Scale. will be collected through the MyHealthyPregnancy smartphone app, a mobile health application developed through close collaboration between decision scientists, clinicians, statisticians, and local peripartum women. A daily journal embedded in the MyHealthyPregnancy app will collect natural language text from the participants for 10 months (from their first prenatal visit through two months postpartum). Using three distinct natural language processing algorithmic approaches, this study will characterize how the natural language used by peripartum women in their daily journal entries is connected to the onset and experience of peripartum depression, as measured through monthly-administered depression scales. Group- based trajectory modeling will then classify women according to the patterns in their depression scores over time. Potential Impact: This work lays the foundation for developing and evaluating real-time interventions that could be deployed at scale to women who are using language that signals high depression risk.