Novel computer vision-based assessment of infant-caregiver synchrony as an early level II screening tool for autism

项目总结 R21解决了对能够检测自闭症谱系的准确且可扩展的筛查工具的迫切需求 生命第一年内的精神障碍(ASD)。该项目将试验一种创新的数字表型筛选 该方法使用计算机视觉和机器学习来测量简单婴儿照顾者的同步性 互动。同步性是指婴儿自发地和动态地协调他们的 及时与照顾者交流行为。这一关键的和早期出现的发展进程可以提供 关于婴儿患自闭症风险的独特而准确的信息，同时也为早期了解提供了一个镜头 社交差异是ASD的核心。意义：该项目代表了ASD的范式转变 筛选，超越行为评级标准，转向更适合捕捉细微之处的方法 房间隔缺损的早期指标。照顾者评级量表缺乏检测体征所需的粒度和客观性 在第一年缓慢而微妙地出现的ASD。方法：跨学科研究小组将 利用尖端技术，以游戏为基础，在5分钟内客观、精确地测量同步 婴儿与照顾者之间的互动。无标记计算机视觉将被用来量化面部运动，捕获 使用小型双向摄像头，不引人注意。婴儿和照顾者面部的二元同步性 然后，将在整个交互过程中计算动作，作为自动机器学习管道的一部分。 初步数据：我们在简短的对话中对患有和不患有自闭症的年轻人评估了这种方法 与研究人员的互动。在机器学习分析管道中，对同步特征进行分类 诊断准确率为91%-显著优于评估相同视频的专业临床医生。相同 一组同步特征显著地预测了ASD组的症状严重程度，表明该方法 对于个体差异的诊断分类和维度预测都是有效的。重要的是， 流水线还以同样高的准确率对儿童诊断进行分类，证明了 不同年龄段的结果。目标。该项目将这些基于计算机视觉的方法扩展到婴儿， 首要目标是评估它们作为ASD II级筛查者的效用。目标1将评估并发 通过评估交互同步性的计算度量与 建立了临床医生对早期ASD标志物的评估。目标2将评估我们的 在12个月时，通过测试其预测未来的能力，将交互同步性衡量为II级筛查工具 ASD诊断具有较高的特异性。影响：这款R21将为一种新颖的计算机视觉提供初步验证- 基于ASD婴儿期筛查。通过以婴儿和照顾者之间的自然互动为目标，这种方法 甚至在典型的自闭症症状之前就有可能识别出社会发展中断的非常早期的迹象 浮出水面。此外，这种基于快速交互的筛选器很容易适应常规儿科护理的背景， 作为可在全球筛查框架内部署的II级筛查人员，前景看好。

PROJECT SUMMARY This R21 addresses a critical need for accurate and scalable screening tools able to detect autism spectrum disorder (ASD) within the first year of life. This project will pilot an innovative digital phenotyping screening method, which uses computer vision and machine learning to measure synchrony within simple infant-caregiver interactions. Synchrony refers to the tendency for infants to spontaneously and dynamically coordinate their behaviors with their caregivers in time. This critical and early-emerging developmental process may provide unique and precise information about an infant’s risk for ASD, while also offering a lens for understanding early social interaction differences at the core of ASD. Significance: This project represents a paradigm shift in ASD screening, moving beyond behavior rating scales toward methods that are better suited to capture the subtle early indicators of ASD. Caregiver rating scales lack the granularity and objectivity necessary for detecting signs of ASD that emerge slowly and subtly throughout the first year. Approach: The interdisciplinary study team will leverage cutting-edge technology to objectively and granularly measure synchrony within 5-minute, play-based infant-caregiver interactions. Markerless computer vision will be used to quantify facial movements, captured unobtrusively with small, bidirectional cameras. The dyadic synchrony among infants’ and caregivers’ facial movements will then be calculated throughout the interaction, as part of an automated machine learning pipeline. Preliminary Data: We evaluated this approach in young adults with and without ASD during brief conversational interactions with research staff members. In a machine learning analysis pipeline, synchrony features classified diagnosis with 91% accuracy - significantly better than expert clinicians assessing the same videos. The same set of synchrony features significantly predicted symptom severity in the ASD group, suggesting that this method is effective for both diagnostic classification and dimensional prediction of individual differences. Importantly, the pipeline also classified diagnosis in children with similarly high accuracy, demonstrating the reproducibility of results across age groups. Aims. This project extends these computer vision-based methods to infants, with the overarching goal of evaluating their utility as a Level II screener for ASD. Aim 1 will evaluate the concurrent validity of our computational measures of interactional synchrony by evaluating their relationships with an established clinician-administered assessment of early ASD markers. Aim 2 will assess the utility of our interactional synchrony measure as a Level II screening tool at 12 months, by testing its ability to predict future ASD diagnosis with high specificity. Impact: This R21 will provide initial validation for a novel, computer vision- based screener for ASD in infancy. By targeting the dynamics of natural infant-caregiver interactions, this method has the potential to identify very early signs of disrupted social development, even before classic ASD symptoms emerge. Moreover, this quick interaction-based screener would fit easily into the context of routine pediatric care, holding promise as a Level II screener deployable within a universal screening framework.