Creating an adaptive screening tool for detecting neurocognitive deficits and psychopathology across the lifespan

精准医学设想的将行为测量纳入大规模研究的努力是 受制于所需的时间和专业知识。纸笔测试目前主导着临床评估 而神经心理测试显然是不可行的。NIH工具箱包含许多计算机化的测试和 临床评估工具的可行性各不相同。工具箱中独一无二的是宾夕法尼亚大学计算机化的神经认知 电池(CNB)，它包含14个测试，需要一个小时来管理。CNB已通过以下验证 功能神经成像，并在全球范围内的多个正常和临床人群中使用，以及 可免费用于研究。临床评估工具通常专门用于特定的障碍，和量表 他们的注意力集中在特定于疾病的症状上。我们已经制定了一项广泛的评估 工具(GOASSESS)，它目前需要大约一个小时来管理。这些仪器是建造的， 使用经典心理测量理论(CTT)进行优化和验证，并在CTT允许的情况下保持效率。 然而，基因组研究需要更省时的工具，可以大规模应用。 基于项目反应理论(IRT)的新方法可以极大地提高测验的效率和 临床评估。IRT通过估计条目将重点从测试转移到构成测试的条目上 在一般特质水平范围内的“难度”和“辨别”等参数。IRT有助于缩短 在不影响数据质量的情况下管理的时间长度，并且对于许多领域导致计算机自适应 测试(CAT)，进一步优化对个人能力的测试。我们建议开发和验证自适应 CNB和GOASSESS的版本，导致神经认知和临床筛查，使用机器 学习工具将不断优化，在部署时变得更短、更精确。该工具将 在公共领域中可用的工具箱中。我们有要对其执行IRT分析的项目级信息 现有数据，并使用此信息开发CAT实施并为自适应 测试。我们的具体目标是：1.使用宾夕法尼亚州立大学和GOASSESS上可用的逐项数据，并添加 新的测试和项，用于生成项池以扩展范围，同时使用IRT-CAT和 其他方法。将扩大当前项目库，以允许在CAT期间进行大量项目选择 管理和添加临床项目到GOASSESS。新产品将通过众包进行校准。2. 制作涵盖主要RDoC的神经认知和临床评估单元的模块化CAT版本 领域和各种精神症状。我们已经在一些CNB测试中实施了这一程序 和临床量表，并将酌情将类似程序应用于剩余和新的测试。3.验证 100例精神病谱系障碍(PS)和100例抑郁症/焦虑症患者的CAT版本 (DA)和100名健康对照(HC)。我们将使用该数据集来实现和测试数据挖掘算法 优化对特定结果的预测。所有测试、算法和标准数据都将放在工具箱中。

Efforts to include behavioral measures in large-scale studies as envisioned by precision medicine are hampered by the time and expertise required. Paper-and-pencil tests currently dominating clinical assessment and neuropsychological testing are plainly unfeasible. The NIH Toolbox contains many computerized tests and clinical assessment tools varying in feasibility. Unique in the Toolbox is the Penn Computerized Neurocognitive Battery (CNB), which contains 14 tests that take one hour to administer. CNB has been validated with functional neuroimaging and in multiple normative and clinical populations across the lifespan worldwide, and is freely available for research. Clinical assessment tools are usually devoted to specific disorders, and scales vary in their concentration on symptoms that are disorder specific. We have developed a broad assessment tool (GOASSESS), which currently takes about one hour to administer. These instruments were constructed, optimized and validated with classical psychometric test theory (CTT), and are efficient as CTT allows. However, genomic studies require even more time-efficient tools that can be applied massively. Novel approaches, based on item response theory (IRT) can vastly enhance efficiency of testing and clinical assessment. IRT shifts the emphasis from the test to the items composing it by estimating item parameters such as “difficulty” and “discrimination” within ranges of general trait levels. IRT helps shorten the length of administration without compromising data quality, and for many domains leads to computer adaptive testing (CAT) that further optimizes tests to individual abilities. We propose to develop and validate adaptive versions of the CNB and GOASSESS, resulting in a neurocognitive and clinical screener that, using machine learning tools, will be continually optimized, becoming shorter and more precise as it is deployed. The tool will be in the Toolbox available in the public domain. We have item-level information to perform IRT analyses on existing data and use this information to develop CAT implementations and generate item pools for adaptive testing. Our Specific Aims are: 1. Use available itemwise data on the Penn CNB and the GOASSESS and add new tests and items to generate item pools for extending scope while abbreviating tests using IRT-CAT and other methods. The current item pool will be augmented to allow large selection of items during CAT administration and add clinical items to GOASSESS. New items will be calibrated through crowdsourcing. 2. Produce a modular CAT version of a neurocognitive and clinical assessment battery that covers major RDoC domains and a full range of psychiatric symptoms. We have implemented this procedure on some CNB tests and clinical scales and will apply similar procedures to remaining and new tests as appropriate. 3. Validate the CAT version in 100 individuals with psychosis spectrum disorders (PS), 100 with depression/anxiety disorders (DA), and 100 healthy controls (HC). We will use this dataset to implement and test data mining algorithms that optimize prediction of specific outcomes. All tests, algorithms and normative data will be in the toolbox.