Developing statistical image analysis tools for non-invasive monitoring of anemia in low birth weight infants

项目摘要 我们的建议是出于需要开发非侵入性的工具，监测贫血，在非常低的出生 体重（极低出生体重;出生体重<1，500克），并减少常规疼痛，侵入性血液采样的数量 可能改变婴儿神经发育和行为的程序（放血术）。近日，一款新的智能手机 应用[Mannino等人，Nature Communications，9，4924（2018）]，收集和分析临床苍白， 已开发出源自患者的指甲照片和图像元数据来预测血红蛋白水平。的 应用程序使用一个强大的多元线性回归模型，该模型包含汇总颜色强度值（平均值 以及由捕获图像的设备生成的图像元数据， 预测病人的血红蛋白水平虽然当前的app算法简单且易于实现，但 明显的局限性。首先，它没有充分利用指甲照片中丰富的空间信息， 计算简单平均值。其次，目前的算法只使用成年人进行训练，其临床 特征与婴儿有很大不同。应用程序预测和血液之间的95%一致性限度 成人样本血红蛋白水平报告为2.4 g/dL，高于临床实验室 改善修正案的质量标准方差为1.0 g/dL，鉴于VLBW婴儿 微小的非特异性指甲床这种严格的误差要求和群体的异质性要求更多 准确和定制的算法比目前的应用程序采用。最后，将应用程序应用于 目前缺乏在VLBW婴儿的纵向护理连续体中最大限度地减少抽血。 基于这些考虑，我们提出（目标1）开发一种新的图像分析算法（IAA）， 无创、准确、稳定地预测血红蛋白水平。IAA将基于一个新的原则 成分分析方法，提供了一个非参数和简约的手段，共同建模高， 三维照片和图像元数据，同时充分利用其空间结构和共同变化的模式。 我们还将考虑一种新的偏最小二乘方法作为替代方法。我们将培训和验证 IAA基于成人数据以及VLBW婴儿数据。在目标2中，我们将开发一种新的聚类方法， 研究指甲照片和图像元数据的子群体结构，并研究它们与 贫血的潜在生理机制。我们可以用这个方法来制作一个非侵入性的图像- 通过识别具有高贫血风险的极低出生体重婴儿群，在目标3中，我们将开发 数据驱动的工具，利用纵向、患者水平的临床数据和IAA预测， 总体临床目标是在整个护理过程中尽量减少VLBW婴儿的抽血次数 连续体我们的建议将使用在三级新生儿重症监护室监测的极低出生体重婴儿的数据 单位在亚特兰大所提出的方法通常适用于各种各样的环境，具有不同的和 临床数据的复杂形式。

Project Summary Our proposal is motivated by the need to develop non-invasive tools for monitoring anemia in very low birth weight (VLBW; birth weight < 1,500 grams) and reduce the number of routine painful, invasive blood sampling procedures (phlebotomy) that may alter infant neurodevelopment and behavior. Recently, a new smartphone application [Mannino et al., Nature Communications, 9, 4924 (2018)] that collects and analyzes clinical pallor in patient-sourced fingernail photos and image metadata has been developed to predict hemoglobin levels. The app uses a robust multi-linear regression model that incorporates summary color intensity values (average across pixels) of fingernail photos well as the image metadata generated by the device capturing the image to predict patient's hemoglobin level. While the current app algorithm is simple and easy to implement, there are notable limitations. First, it does not fully leverage the rich spatial information available in fingernail photos by calculating a simple average value. Second, the current algorithm is trained using only adults, whose clinical characteristics are vastly different from infants. The 95% limit of agreement between the app-predicted and blood sample-based hemoglobin level for adults is reported as 2.4 g/dL, which is higher than the Clinical Laboratory Improvement Amendments specification variance of 1.0 g/dL, and will likely increase in VLBW infants given their tiny, non-specific fingernail beds. Such strict error requirements and heterogeneity in populations demand more accurate and tailored algorithms than what the current app employs. Lastly, a framework for applying the app to minimize blood draws across the longitudinal care continuum for VLBW infants is currently lacking. With these considerations, we propose (Aim 1) to develop a new image analysis algorithm (IAA) that produces non-invasive, accurate and stable prediction of hemoglobin level. The IAA will be based on a novel principal component analysis method that provides a non-parametric and parsimonious means to jointly model high- dimensional photos and image metadata, while fully leveraging their spatial structures and co-varying patterns. We will also consider a new partial least squares approach as an alternative method. We will train and validate the IAA based on adult data as well as VLBW infant data. In Aim 2, we will develop a new clustering method to study sub-population structures of fingernail photos and image metadata and study their relationships with the underlying physiological mechanisms of anemia. This approach will allow us to formulate a non-invasive image- based screening tool by identifying clusters of VLBW infants with high anemia risk. In Aim 3, we will develop data-driven tools that leverage longitudinal, patient-level clinical data and IAA predictions to achieve the overarching clinical goal of minimizing the number of blood draws in VLBW infants throughout the care continuum. Our proposal will use the data of VLBW infants monitored at three level III neonatal intensive care units in Atlanta. The proposed methods are generally applicable to a wide variety of settings with diverse and complex modalities of clinical data.