Using Population Contrast Sensitivity Function Data to Develop Tunable Test Procedures

摘要 视觉对比敏感度代表视觉系统的核心处理能力，可用于诊断 各种视觉障碍。量化这种能力的最简单、最容易、最廉价和最便携的方法是 直接查询-提供适当的视觉刺激并记录行为反应。和所有人一样 然而，心理物理测试估计对比敏感度函数(CSF)需要连续数据采集， 这导致了超长的收购时间。因此，尽管完整的CSF可能具有显著的临床价值， 缺乏定量精确度的快速心理物理筛查通常出于实际原因而被使用。 该方案的目标是将机器学习算法和高质量的回溯CSF相结合 数据以设计可调的诊断估计器，既可以是快速的(用于筛查)也可以是彻底的(用于筛查 诊断)，视需要而定。我们的方法将是训练一个多维贝叶斯主动机器学习 先前已针对视野视野检查和听力测试进行验证的估计器-共享 许多属性都带有对比敏感度测试。在目标1中，我们将实现并验证机器学习 脑脊液估计器(MlCSF)。这种类型的估计器适应灵活的假设，并允许优化 收集数据以最大限度地获得信息。在目标2中，我们将使用种群CSF提高mlcsf效率。 数据。Mlcsf的贝叶斯性质允许先前来自总体的经验发现来改进先前 对新测试对象的信念。将使用从以前的脑脊液检测程序得出的总体汇总 为mlcsf估计者建立提供信息的先验信念。在目标3中，我们将扩展mlcsf模型以包括 相关单项措施。其他视觉测试的结果是与个人的脑脊液相关的测量结果。 在先前从同一视频处理数据收集的视觉处理数据中这些额外预测者之间的关系 个体将被用来改进mlcsf估计器的先验信念。完成后，这项研究将有 开发了一个用于评估概率对比敏感度的尖端主动机器学习框架 使用相对较少的测量即可运行。这种估计器的灵活性将允许实验者和 临床医生将理论假设和经验先验信念结合起来，解决各种临床问题 从筛查到诊断的各种问题都采用了同样的程序。

ABSTRACT Visual contrast sensitivity represents a core processing ability of the visual system useful for diagnosing a variety of visual disorders. The simplest, easiest, cheapest and most portable way to quantify this ability is by querying directly—delivering appropriate visual stimuli and recording behavioral responses. As with all psychophysical tests, however, estimating contrast sensitivity functions (CSFs) requires serial data acquisition, leading to impractically long acquisition times. While full CSFs can therefore have significant clinical value, quick psychophysical screenings that lack quantitative precision are often used instead for practical reasons. The objective of this proposal is to combine machine learning algorithms and high-quality retrospective CSF data to design tunable diagnostic estimators that can be either quick (for screening) or thorough (for diagnostics), as desired. Our approach will be to train a multidimensional Bayesian active machine learning estimator that has been validated previously for visual field perimetry and audiometric testing—tests that share many properties with contrast sensitivity testing. In aim 1 we will implement and validate a machine learning CSF estimator (mlCSF). This type of estimator accommodates flexible assumptions and allows optimization of data collection for maximizing information gain. In aim 2 we will improve mlCSF efficiency with population CSF data. The Bayesian nature of mlCSF allows for previous empirical findings from a population to refine prior beliefs for new test subjects. Population summaries derived from previous CSF testing procedures will be used to establish informative prior beliefs for the mlCSF estimator. In aim 3 we will extend mlCSF models to include related individual measures. Other visual tests result in measurements that correlate with an individual’s CSF. Relationships among these extra predictors in previously collected visual processing data from the same individuals will be used to refine the prior beliefs of the mlCSF estimator. When complete, this study will have produced a cutting-edge active machine learning framework to estimate probabilistic contrast sensitivity functions using relatively few measurements. The flexibility of this estimator will allow experimenters and clinicians to combine theoretical assumptions and empirical prior beliefs to address a variety of clinical questions ranging from screening to diagnosis with the same procedure.