A Predictive Analytics Approach to the Optimization of Diagnosis, Treatment, and Ambulatory Management of Major Depressive Disorder and Bipolar Disorder

与重度抑郁症（MDD）和双相情感障碍（BD）患者合作的临床医生面临着多重挑战，从确定及时正确的诊断和设计最佳治疗策略到以最大限度地减少疾病进展和个体痛苦的方式管理情绪发作。在这里，我们现在将利用新兴的心理健康预测分析领域的进展来开发，调整和实施最先进的机器学习算法，直接解决应用情感障碍研究中三个最紧迫的临床目标。我们的目标是构建诊断支持模型：1）区分MDD和BD患者，2）预测MDD患者对电休克治疗（ECT）的个体反应，3）能够基于智能手机数据进行动态实时复发风险预测。为此，我们将设计一种原则性的方法来处理MDD和BD的大量多变量和多模态性质，并提供一个算法框架，旨在预测实时复发风险。具体来说，我们将首先开发一种用于自动化特征工程的无监督深度学习方法，目标是在低维流形上表示磁共振成像数据。这不仅可以缓解Curse of Expersionality，而且还可以提供对扫描仪站点和采集协议更恒定的功能，从而能够无缝构建更强大的多中心模型。其次，我们将采用最先进的数据集成方法来融合遗传，心理测量和神经影像学信息，使MDD与BD分类和ECT反应预测成为可能。建立一个多模态模型，包括三个最严重的调查数据源在情感障碍研究将-第一次-提供经验证据的问题，在何种程度上结合患者的特点，通常在精神病学提高模型的性能。第三，我们将开发一个基于智能手机数据的实时复发风险预测模型，评估算法，这些算法将复发风险概念化为1）与无复发期间存在的模式的偏差，以及2）从复杂系统的角度来看，作为一个关键的过渡。建立在多模态神经成像，遗传和心理数据已经提供给我们（总N> 41，000）以及当前在相关多中心项目中获取的基于智能手机的数据集（总N > 2，000），我们将全力推进...艺术机器学习方法在情感障碍的研究，并促进预测模型的建设与直接相关的临床实践将最大的可用数据集和当今出现的最强大的机器学习算法结合在一起，为促进精神病学的转化工作提供了独特的机会，将该领域从概念验证研究转向首次临床应用。

Clinicians working with Major Depressive Disorder (MDD) and Bipolar Disorder (BD) patients face multiple challenges, ranging from determining a timely and correct diagnosis and devising an optimal treatment strategy to managing mood episodes in a way that minimizes disease progression and individual suffering. Here, we will now utilize advances in the emerging field of Predictive Analytics in Mental Health to develop, adapt, and implement state-of-the-art machine-learning algorithms directly addressing three most pressing clinical objectives in applied affective disorder research. We aim to construct diagnostic support models 1) differentiating MDD and BD patients, 2) predicting individual response to Electroconvulsive Therapy (ECT) in MDD patients and 3) capable of dynamic real-time relapse-risk prediction based on smartphone data. To this end, we will devise a principled approach dealing with the massively multivariate and multimodal nature of MDD and BD and provide an algorithmic framework aiming to predict real-time relapse-risk. Specifically, we will first develop an unsupervised Deep Learning approach for automated feature-engineering with the goal of representing Magnetic Resonance Imaging data on a lower-dimensional manifold. This will not only alleviate the Curse of Dimensionality, but also provide features which are more invariant to scanner sites and acquisition protocols, enabling the seamless construction of more robust multi-center models. Second, we will employ state-of-the-art data integration methodology to fuse genetic, psychometric, and neuroimaging information, enabling MDD vs. BD classification and ECT response prediction. Building a multimodal model comprising the three most heavily investigated data sources in affective disorder research will – for the first time – provide empirical evidence regarding the question to what extent combining patient characteristics commonly measured in psychiatry improves model performance. Third, we will develop a model for real-time relapse-risk prediction based on smartphone data by evaluating algorithms which conceptualize relapse-risk 1) as a deviation from the patterns present during symptom-free periods and 2) as a critical transition from a Complex Systems perspective. Building on multimodal neuroimaging, genetic and psychometric data already available to us (total N>41,000) as well as on the smartphone-based dataset currently acquired in an associated multicenter project (total N > 2,000), we will fully focus on advancing state-of-the-art machine learning methodology in affective disorders research and facilitate the construction of predictive models with direct relevance for clinical practice. Bringing together the largest datasets available and the most powerful machine learning algorithms emerging today opens up the unique opportunity to catalyze translational efforts in psychiatry, moving the field from proof-of-concept studies towards first clinical applications.