Personalized Deep Learning Models of Rapid Changes in Major Depressive Disorder Symptoms using Passive Sensor Data from Smartphones and Wearable Devices

摘要 重度抑郁症（MDD）是一种高度流行的疾病，是全球疾病负担的主要原因。 与1，000多种不同的症状特征相关，MDD具有高度异质性。大多数MDD 症状在数小时内发生变化。因此，需要将MDD研究的重点越来越多地集中在 这些快速症状波动的个性化评估。迄今为止，MDD的个性化模型已经 显示出希望，但仅依赖于自我报告的措施。因此，迫切需要开发个性化的 结合客观信号的MDD模型。从智能手机被动收集信息， 可穿戴传感器可以连续且不引人注目地跟踪与核心相关的行为和生理信号， 与MDD相关的障碍，包括精神发育迟滞、睡眠障碍、社会接触， 行为激活、心率变异性和屏幕时间。初步数据显示， 人工智能（即，个人加权的深度学习模型）非常适合创建新颖的 这些被动指标的个性化数字生物标志物，这些生物标志物可以预测快速 MDD症状的变化。该提案将调查开发个性化深度学习的能力 120例治疗的全国代表性样本中MDD症状的快速变化模型 使用从智能手机和可穿戴设备被动收集的数据寻找90天内患有MDD的成年人 传感器.这个提议旨在测试个性化、子类型化和基于队列的建模的准确性 技术，并发现MDD症状中每时每刻变化的个性化数字生物标志物。 该项目提出了以下创新：（1）进行MDD的第一次被动感知研究， 具有全国代表性的队列;（2）利用深度学习模型来帮助发现新的维护 MDD症状变化的因素;（3）使用MDD的个性化多模式评估来解决 MDD的异质性。根据NIMH研究领域标准（RDoC）的目标，该项目将 跨多个分析单元研究MDD症状变化并集成多个系统。本研究将 为发现MDD症状变化的新型个性化维护模式迈出了关键一步 在日常生活中此外，它将允许使用技术开发可扩展的个性化治疗， 在MDD症状快速变化之前立即进行行为干预。

ABSTRACT Major depressive disorder (MDD) is highly prevalent and the leading cause of global disease burden. Associated with over 1,000 different symptom profiles, MDD is highly heterogeneous. The majority of MDD symptom change occurs across hours. Consequently, there is a need to increasingly focus MDD research on personalized assessment of these rapid symptom fluctuations. To date, personalized models of MDD have shown promise, but relied solely on self-report measures. There is thus a critical need to develop personalized models of MDD that incorporate objective signals. Passively collected information from smartphones and wearable sensors can continuously and unobtrusively track behavioral and physiological signals related to core disturbances associated with MDD, including psychomotor retardation, sleep disturbances, social contact, behavioral activation, heart rate variability, and screen time. Preliminary data suggest that personalized artificial intelligence (i.e., personally weighted deep learning models) are well suited for creating novel personalized digital biomarkers of these passive indicators, and that these biomarkers can predict rapid changes in MDD symptoms. This proposal will investigate the ability to develop personalized deep learning models of rapid changes in MDD symptoms among a nationally representative sample of 120 treatment seeking adults with MDD across 90 days using passively collected data from smartphones and wearable sensors. This proposal aims to test the accuracy of personalized, subtyped, and cohort-based modeling techniques and uncover personalized digital biomarkers of moment-to-moment changes in MDD symptoms. The project proposes the following innovations: it will (1) conduct the first passive-sensing study of MDD in a nationally-representative cohort; (2) utilize deep learning models to aid in the discovery of novel maintenance factors of MDD symptom changes; and (3) use personalized multimodal assessments of MDD to address the heterogeneity in MDD. In line with the aims of the NIMH Research Domain Criteria (RDoC), this project will study MDD symptom changes across multiple units of analysis and integrate multiple systems. This study will provide a critical step towards uncovering novel personalized maintenance patterns of MDD symptom changes in daily life. Further, it will allow for scalable personalized treatments to be developed using technology to deliver behavioral interventions in the moments immediately preceding rapid MDD symptom changes.