Longitudinal trajectory identification of Post-Traumatic Stress Disorder using Machine Learning for high-dimensional cognitive, emotional and biological data

道路交通事故或袭击等造成的创伤影响全世界数以百万计的人，并构成重大的健康负担。一些人在潜在的精神创伤事件后出现精神障碍症状，而另一些人则没有。为了尽快提供有针对性的治疗服务，重要的是可靠地识别那些已经处于早期阶段的精神障碍高危人群。以前对预后因素的研究一直很有限，不同的研究往往发现不同的预测因素。这有两个方法论上的原因。首先，当患者在精神障碍发作很长一段时间后进行回顾检查时，很难清楚地区分精神疾病的危险因素和精神疾病的后果。其次，当应用传统的统计方法利用广义线性模型孤立地研究某些风险因素的焦点假设时，很难发现变量之间的虚假关系(省略相关变量)或变量冗余。为了克服第一个问题，有必要在潜在的创伤事件发生后不久对患者进行检查，以便为未来的精神病理学确定最近的风险因素。为了解决第二个问题，有必要检查包含所有相关信息的高维数据，或者至少比之前检查的更多的观测数据，并使用机器学习等统计方法，允许在一段时间内仔细检查这些高维数据。该项目的目标是通过对潜在创伤事件发生后立即以及1、3、6和12个月后在纽约市贝尔维尤医院中心急诊室的患者的高维数据应用机器学习来识别最具预测性的风险因素集。对几个测量点的高维纵向数据的前瞻性收集将允许在12个月的过程中识别创伤后应激障碍(PTSD)的不同轨迹。由于共病是创伤后应激障碍的规则而不是例外，收集的高维数据集也将被用于从遗传学、生理学、心理学和社会学特征识别不同的共病轨迹。因此，另一个目的是预测这些特定轨迹的成员资格，并在独立样本中验证风险因素。总而言之，准确预测集的发现将使人们能够在潜在创伤事件后的急性期识别属于不同精神病理学轨迹的人。这具有重要的公共卫生影响；它为具体针对性治疗选择的新战略提供了信息，并优化了治疗服务的有效和高效率分配。

Traumatic injuries resulting e.g. from road traffic accidents or assaults affect millions of people worldwide and constitute a major health burden. Some individuals develop symptoms of mental disorders after potential traumatizing events while others do not. To offer targeted therapeutic services as soon as possible, it is important to reliably identify those individuals that are at high risk for developing mental disorders already at an early stage. Previous research on prognostic factors has been limited, with different studies often finding different predictors. There are two methodological reasons for this. First, when patients are examined retrospectively, long time after the onset of a mental disorder, it is difficult to clearly distinguish between risk factors for mental illness and the consequences of mental illness. Second, when conventional statistical methods are applied utilizing a Generalized Linear Model to investigate focused hypothesis on certain risk factors in isolation, spurious relationships (omitting relevant variables) or redundancy of variables are hard to detect. To overcome the first problem, it is necessary to examine people shortly after a potential traumatic event in order to identify proximal risk factors for future psychopathology. To solve the second problem, it is necessary to examine high-dimensional data that comprises all relevant information, or at least much more observations than previously examined, and to use statistical methods such as Machine Learning that allow to scrutinize such high-dimensional data over the course of time.The aim of the project is to identify maximally predictive sets of risk factors by applying Machine Learning on high-dimensional data from patients at the emergency room of Bellevue Hospital Center in New York City immediately after a potentially traumatic event as well as 1, 3, 6 and 12 months later. The prospective collection of high-dimensional longitudinal data across several measurement points will allow identifying different trajectories of posttraumatic-stress disorder (PTSD) over the course of 12 months. Since comorbidity is the rule rather than the exception in PTSD the collected high-dimensional dataset will also be used to identify different trajectories of comorbidity from genetic, physiological, psychological, and sociological characteristics. A further aim is therefore to predict membership to these specific trajectories and to validate the risk factors in an independent sample.Taken together, the discovery of accurate predictor sets will enable the identification of people during the acute phase after a potential traumatic event who belong to different trajectories of psychopathology. This has important public health implications; it informs new strategies for specifically targeted treatment selection and optimizes the effective and efficient allocation of treatment services.