Building normative models of Reinforcement Learning Decision Making Behavior

建立强化学习决策行为的规范模型

• 批准号: 10572615
• 负责人: POORNIMA KUMAR
• 金额: $ 24.6万
• 依托单位: MCLEAN HOSPITAL
• 依托单位国家: 美国
• 财政年份: 2022
• 资助国家: 美国
• 起止时间: 2022-12-01 至 2024-11-30
• 项目状态: 已结题
• 来源: https://reporter.nih.gov/project-details/10572615
• 关键词: Adopted; Age; Bayesian Method; Behavior; Behavior Disorders; Behavioral; Case/Control Studies; Clinical; Communities; Complex; Computer Models; Conflict (Psychology); Data; Data Reporting; Decision Making; Diagnosis; Diagnostic; Disease; Environment; Ethnic Origin; Failure; Family; Gender; Goals; Grain; Growth; Heterogeneity; Impairment; Individual; Learning; Link; Location; Longevity; Maps; Mathematics; Measures; Mental disorders; Modeling; Normal Range; Outcome; Participant; Patient Self-Report; Patients; Pediatrics; Population; Process; Psychiatry; Psychological reinforcement; Psychopathology; Research; Research Domain Criteria; Risk; Sample Size; Sampling; Shapes; Stimulus; Stratification; Structure; Symptoms; Testing; Time; Uncertainty; Update; aged; clinical phenotype; clinical practice; cohort; computer framework; cost; discounting; flexibility; improved; mathematical model; novel; precision medicine; psychiatric symptom; success; virtual

The main objective in the field of psychiatry is to be able to treat patients at an individual level. To reach this goal of precision medicine, large scale initiatives such as the RDoC have been developed to find new ways to parse heterogeneity in psychiatric disorders. Their success has been slow partly due to a slow transition away from case-control studies based on diagnoses and limitations due to small sample size. Therefore, there is a critical need to find alternative solutions at an affordable cost. One strategy is to identify a complex behavior such as Reinforcement Learning based decision making (RLDM) that is impaired across various psychiatric disorders and adopt a computational framework to explain heterogeneity at an individual level. RLDM is a multifaceted construct involving several sub-processes ranging from estimating values of different options in the environment (valuation), accumulating evidence for these options (sequential sampling), choosing the best option (explore- exploit behavior), estimating the outcome value (salience attribution) and lastly integrating relevant information about outcomes and updating the value of stimuli (learning rate). These sub-processes can be quantified by utilizing computational models. However, prior to building normative models of these RLDM constructs that can be potentially utilized in clinical practice, it is critical to assess the reliability of these RLDM model-derived parameters to avoid translational failures. Therefore, our main goals of this proposal are to (1) parse RLDM sub- processes into mathematically-defined parameters in a large sample using a diverse set of tasks; (2) assess test-retest reliability of these parameters; and finally (3) build normative models of the parameters and chart the heterogeneity at the level of the individual. To achieve this goal, we will acquire behavioral data from six RL tasks, including delay and effort discounting, probabilistic learning, virtual patch foraging, Pavlovian instrumental transfer and approach-avoid conflict tasks. These tasks are each optimized to measure at least two of the RL sub-processes separately. We will collect behavioral data and self-report assessments from a community sample from 1000 participants (aged 18-85). To estimate test-retest reliability of RLDM parameters, we will invite 500 healthy participants from this sample to complete the session again in one week’s time. First, we will apply start- of-art computational models to quantify RLDM behavior in each subject. Second, we will calculate the test-retest reliability of these parameters. Third, we will build normative models to link each of the RLDM construct with age and calculate each subject’s deviation from the norm. Lastly, we will conduct soft clustering on these deviations to identify clusters and investigate their differences in psychopathology and general functioning.

精神病学领域的主要目标是能够在个人层面上治疗患者。达到这个目标 在精确医学上，已经开发了大规模的大规模计划，以找到解析的新方法 精神疾病的异质性。他们的成功部分是由于远离缓慢的过渡 病例对照研究基于诊断和限制，由于样本量较小。因此，有一个关键 需要以负担得起的成本找到替代解决方案。一种策略是确定一个复杂的行为，例如 基于强化学习的决策（RLDM），这些决策受到各种精神疾病的损害 并适应一个计算框架以解释单个级别的异质性。 RLDM是多方面的 构建涉及几个子过程，从估计环境中不同选项的值 （估值），积累这些选项的证据（顺序抽样），选择最佳选择（探索 - 利用行为），估计结果价值（显着属性）和最后整合相关信息 关于结果并更新刺激的价值（学习率）。这些子处理可以通过 利用计算模型。但是，在构建这些RLDM构建体的正常模型之前 可能在临床实践中使用，评估这些RLDM模型衍生的可靠性至关重要 参数以避免翻译失败。因此，该提案的主要目标是（1）解析RLDM子 - 在大型样本中使用潜水员组合的任务中的数学定义参数中的过程； （2）评估 这些参数的测试可靠性；最后（3）构建参数的正常模型，并图表 个人水平的异质性。为了实现这一目标，我们将从六个RL获取行为数据 任务，包括延迟和努力打折，概率学习，虚拟补丁觅食，Pavlovian工具 转移和接近 - 避免冲突任务。这些任务均经过优化，以测量至少两个RL 子过程分别。我们将从社区样本中收集行为数据和自我报告评估 来自1000名参与者（18-85岁）。为了估计RLDM参数的重测可靠性，我们将邀请500 该样本的健康参与者在一周的时间内再次完成课程。首先，我们将应用启动 - ART计算模型以量化每个主题中的RLDM行为。其次，我们将计算重测 这些参数的可靠性。第三，我们将建立正常模型，将每个RLDM构造与年龄联系起来 并计算每个受试者偏离规范。最后，我们将在这些离开时进行软聚类 识别集群并研究其在心理病理学和一般功能方面的差异。