Machine Learning for CCHD Screening using Dynamic Data

项目总结/摘要 我建议开发和测试一种机器学习（ML）算法，该算法使用来自Pulse的动态数据 血氧测定法用于危重先天性心脏病（CCHD）筛查。基于血氧饱和度（SpO 2）的筛查是 CCHD筛查的现行标准;然而，它未能检测到50%的无症状新生儿， CCHD或近900新生儿在美国每年。SpO 2筛查漏诊的大多数新生儿 不引起低氧血症的系统性梗阻缺陷，如主动脉缩窄（CoA）。脉冲 血氧测定法还可以测量非侵入性测量结果例如灌注（例如灌注指数（PIx））， 射频股延迟、心率和其他波形特征。其他脉搏血氧仪介绍 这些功能有望改善CCHD和CoA检测。我最近的工作揭示了改进的CCHD检测 使用结合脉搏血氧测定特征的ML算法。该算法将CCHD检测提高到 至少93%，包括改进的CoA检测，同时保持高特异性。然而，模型 取决于两个单独的测量，包括同时在右侧和右侧的无伪影波形 手和脚。有一个模型与动态演示，允许婴儿的预测结果， 随着新数据的加入而改变可能会更好。此外，获得两个波形的时间量 在可能移动的婴儿中没有伪影的方法需要被理解以用于实现。 因此，我将开发和测试一种ML算法，该算法结合了脉搏血氧饱和度特征， 结合来自重复测量的动态数据，允许新生儿的预测分类（CCHD vs no-CCHD）随着新数据的加入而改变。我将以两种方式来做这件事。第一个将只利用 住院测量并将外部验证我们最近开发的ML算法。该第一方法 还将对任何初始“失败”的“重复”屏幕进行测试，这是一种模拟当前SpO 2标准屏幕的方法 并期望将假阳性率保持在1%以下。第二种方法将纳入测量 48小时后（包括门诊）。尚未对门诊CCHD筛查进行研究。 大多数新生儿在CoA变得更多的年龄附近被视为常规随访门诊 临床上明显，因此更可能通过非侵入性灌注评估检测到。 这项研究是重要的，因为包括灌注数据的动态筛选模型可以节省 数百名患有CCHD的新生儿的生命，这些新生儿每年未通过SpO 2筛查确诊。而且 是创新的，因为它利用了现成的非侵入性脉搏血氧饱和度数据，并将使用动态 数据（住院和门诊），允许新生儿的统计随着新数据的变化而变化， incorporated.从这个学习和职业计划中，我将获得机器学习的技能，重点是动态 方法和实施科学。我将使用这个建议的结果和技巧，然后研究一个 聚类随机试验我们的算法和评估的实施过程。

PROJECT SUMMARY/ABSTRACT I propose to develop and test a machine learning (ML) algorithm that uses dynamic data from pulse oximetry for critical congenital heart disease (CCHD) screening. Oxygen saturation (SpO2)-based screening is the current standard for CCHD screening; however, it fails to detect 50% of asymptomatic newborns with CCHD or nearly 900 newborns in the United States annually. Most newborns missed by SpO2 screening have defects with systemic obstruction, such as coarctation of the aorta (CoA), that do not cause hypoxemia. Pulse oximetry can also measure non-invasive measurements such as perfusion such as perfusion index (PIx), radiofemoral delay, heart rate, and other waveform characteristics. Introduction of other pulse oximetry features is expected to improve CCHD and CoA detection. My recent work revealed improved CCHD detection using ML algorithms that combined pulse oximetry features. The algorithms improved CCHD detection to at least 93%, including improved detection of CoA, while maintaining high specificity. However, the model depended on two separate measurements including simultaneously artifact free waveforms in both the right hand and a foot. Having a model with dynamic prognostication that allows for an infant’s predicted outcome to change as new data is incorporated could be better. Additionally, the amount of time to obtain two waveforms that are artifact free in a possibly moving baby needs to be understood for implementation. Therefore, I will develop and test a ML algorithm that combines pulse oximetry features and incorporates dynamic data from repeated measurements allowing a newborn’s predicted classification (CCHD vs no-CCHD) to change as new data is incorporated. I will do this in two ways. The first will utilize only inpatient measurements and will externally validate our recently developed ML algorithm. This first approach will also test a “repeat” screen for any initial “fails,” an approach that mimics the current SpO2 standard screen and is expected to keep the false positive rate below 1%. The second approach will incorporate measurements after 48 hours of age (including from the outpatient setting). Outpatient CCHD screening has not been studied. Most newborns are seen for routine follow up outpatient around the age at which CoA becomes more clinically apparent, and thus, more likely to be detected by non-invasive perfusion assessments. This study is significant because a dynamic screening model that includes perfusion data could save the lives of hundreds of newborns with CCHD that are not diagnosed by SpO2 screening annually. Additionally, it is innovative because it makes use of readily available non-invasive pulse oximetry data and will use dynamic data (inpatient and outpatient) that allows for a newborn’s prognostication to change as new data is incorporated. From this study and career plan, I will gain skills in machine learning with emphasis in dynamic approaches, and implementation science. I will use the results and skills from this proposal to then study a cluster randomized trial of our algorithm and assess implementation processes.