Evaluating the feasibility of an innovative point-of-care screening tool for detection of infant motor delay within the newborn period

许多慢性神经发育疾病的发病，如脑瘫，神经肌肉 疾病和其他神经遗传性疾病发生在生命的围产期阶段。检测这些 在婴儿期，当大脑可塑性最高时，有可能显着改善 长期成果。在美国（U.S.）诊断脑瘫等疾病的年龄 对于已知围产期危险因素的儿童，约为2岁。多达一半的 然而，患有残疾的儿童来自简单的妊娠，诊断通常 发生在生命的后期。令人担忧的是，对一个来自服务不足社区的孩子的诊断可能是 迟至5岁。诊断这些疾病的传统模式是“观望” 直到孩子最终错过了运动里程碑，因此婴儿错过了关键的几个月。 可能的早期干预。对于有围产期危险因素的婴儿，有时可以作出诊断 通过使用昂贵的技术，如磁共振成像;或专门的评估 一般运动评估（GMA）。然而，这些测试通常仅 在三级护理中心提供。因此，一个护理新生儿筛查系统点可以改变 诊断过程，让每个孩子都有机会早期发现异常运动 新生儿时期的模式。我们的创新系统BabySure使用人工智能 模型，以评估婴儿运动的特点，从一个简单的，非侵入性的视频记录。我们的飞行员 数据显示，BabySure可以评估自发运动并将其分类为健康或 在6个月之前的婴儿中异常。为了确保每个婴儿都有机会获得 早期，准确的诊断，我们将在社区新生儿托儿所试点新生儿筛查计划。 BabySure使用自动骨骼跟踪来提取33个骨骼标志的位置坐标 从婴儿自发运动的每个视频帧（包括四肢上的点，头， 脸和躯干）。我们项目的一个重要部分是继续努力在一个 从历史上看，计算机视觉在深色皮肤上不够准确。到 训练我们的模型来识别异常运动，我们将使用金标准GMA来对婴儿进行分类， 异常或健康运动训练组。我们的分析计划将三个人工智能的原型 集成单个分类器的基本模型（每个分类器有两个头：回归和分类） 包括为图像处理和定制训练而预先训练的微调卷积神经网络 模型最终模型将输出运动功能评分，该评分将婴儿运动分类为典型运动 或异常此外，我们还将评估结果的日内和日间变异性，以告知未来 建议在新生儿时期进行运动筛查的最佳时间或顺序。

The onset of many chronic neurodevelopmental conditions, such as cerebral palsy, neuromuscular conditions, and other neurogenetic disorders occur within the perinatal phase of life. Detection of these conditions, during infancy, when brain plasticity is at its highest, has the potential to dramatically improve long term outcomes. In the United States (U.S.) the age of diagnosis of conditions such as cerebral palsy is around 2 years of age for children who had known perinatal risk factors. As many as half of the children with a disability, however, are from an uncomplicated pregnancy and the diagnosis typically occurs even later in life. Alarmingly, a diagnosis for a child from an underserved community can be as late as 5 years of age. The traditional paradigm for diagnosis of these conditions, uses a ‘wait-and-see’ approach until the child eventually misses motor milestones, thus the infant has missed critical months of possible early intervention. For infants with perinatal risk factors, a diagnosis is sometime achieved through the use of costly technologies, such as magnetic resonance imaging; or specialized evaluation techniques, such as the General Movements Assessment (GMA). However, these tests are typically only available in tertiary care centers. Thus, a point of care newborn screening system could transform the diagnostic process, allowing every child the opportunity for early detection of aberrant movement patterns within the newborn period. Our innovative system, BabySure uses an artificial intelligence model to evaluate infant movement characteristics from a simple, non-invasive video recording. Our pilot data has shown that BabySure can evaluate and categorize spontaneous movements as healthy or aberrant in infants before the age of 6 months. To ensure every baby has the opportunity to receive an early, accurate diagnosis, we will pilot a newborn screening program in a community newborn nursery. BabySure uses automated skeletal tracking to extract positional coordinates of 33 skeletal landmarks from each video frame of the infant’s spontaneous movements (including points on the extremities, head, face, and trunk). An essential part of our project is to continue our efforts to validate our system on a variety of skin tones as historically, computer vision has not been sufficiently accurate on darker skin. To train our model to identify aberrant movement we will use the gold-standard GMA to classify infants in aberrant or healthy movement training groups. Our analysis plan will prototype three artificial intelligence base models to ensemble a single classifier (each with two heads: regression and classification) including fine tuning convolutional neural networks pre-trained for image processing and custom-training models. The final model will output a motor function score that will classify the infant movement as typical or aberrant. In addition, we will evaluate the intra- and inter-day variability of results to inform future recommendations for optimal timing or sequence of movement screening across the newborn period.