Combining computational techniques with movement data to predict adult autism diagnosis

自闭症是一种终生发育状况，会影响一个人与人沟通和互动的方式。除了这些社交方面的因素外，约80%的自闭症患者还存在协调困难，如眼手协调不良、平衡不稳定和步态异常。该项目的医疗保健目标是揭示这些协调困难是否可以用于诊断成人自闭症。目前，自闭症成年人的诊断很困难，而且很耗时，自闭症成年人已经将早期诊断和改善诊断的需要放在了他们的十大研究优先事项中。这是因为现有的诊断标准尚未在成人人群中得到验证，自闭症成人已经制定了补偿策略，观察性清单的主观性质意味着不同的临床医生的诊断可能不同。因此，获得宝贵支持的机会被推迟了。该项目将把在运动任务期间收集的运动跟踪数据与数据科学方法结合起来，以调查基于协调技能的自动化测试在与当前的观察清单结合使用时是否可以为诊断精度提供附加值。使用运动任务来诊断成年人比目前的方法更有优势，因为在整个生命周期中都会出现协调困难，并且可以定量和客观地测量运动，为识别区分特征提供了丰富的数据集。我们最近发表的EPSRC资助的工作展示了这种方法的潜力：机器学习(ML)技术成功地应用于44名自闭症和非自闭症参与者的运动跟踪数据。我们现在需要在更广泛的动作曲目和更大的样本量上创建新的模型，以便能够识别一致的运动模式，从而提高分类精度。目标1：收集关于更广泛的运动任务的数据，并开发稳健的分类工具。目标2：收集数据并测试模型在更大范围的自闭症个体以及运动障碍患者(帕金森氏病(PD)、发育协调障碍(DCD))中的稳健性。目标3：确定协调困难是否可以分为不同的亚组，使用深度聚类方法。接近自闭症和非自闭症的成年人会执行不同的动作(例如，抓杯子、走路、平衡)，而运动传感器会跟踪他们的动作。将开发算法来提取每个任务的相关运动参数。首先，使用特征提取和选择的方法，为每个动作任务识别用于分类的组均值和区分特征。其次，将研究监督学习、分类和深度神经学习方法，以生成稳健的分类器，从而能够从对照组中检测自闭症患者。我们将通过比较基于变异性的经典统计方法和深度学习方法，继续我们在可解释深度学习(理解网络做出特定决定的原因)方面的最新发展。然后，新的自闭症和非自闭症参与者将被用于测试模型的稳健性，以及患有PD和DCD的参与者。K近邻等聚类技术将被用来识别自闭症数据中与非自闭症个体不同的任何子组。新颖性/潜在结果结果将是一个原型ML/人工智能工具，用于识别那些有自闭症风险的人，支持临床决策，并导致更早和更快的诊断。此外，该项目将开发新的分析科学工具，使用ML和深度神经学习来创建健壮的模型。这项研究与医疗保健技术主题中的重大挑战一致：“优化治疗：通过有效的诊断、针对患者的预测和基于证据的干预来优化护理。”

Autism is a life-long developmental condition that affects how a person communicates and interacts with people. In addition to these social aspects, ~80% of autistic individuals have coordination difficulties such as poor eye-hand coordination, unstable balance and unusual gait. The healthcare aim of this project is to uncover whether these coordination difficulties can be used to diagnose autistic adults. Currently, diagnosis of autistic adults is difficult and time consuming and autistic adults have placed the need for earlier and improved diagnosis in their top 10 research priorities. This is because existing diagnostic criteria have not been validated in an adult population, autistic adults have developed compensatory strategies and the subjective nature of the observational inventories mean that diagnosis can vary between clinicians. Consequently, access to valuable support is delayed. This project will combine motion tracking data collected during movement tasks with data science methods to investigate whether an automated test based on coordination skills could provide added value for diagnostic precision, when used in combination with current observational inventories. Using movement tasks to diagnose adults is advantageous over current methods as coordination difficulties occur throughout the lifespan and movement can be measured quantitatively and objectively, providing a rich dataset to identify discriminating features. Our recent published EPSRC-funded work demonstrates the potential of this approach: Machine Learning (ML) techniques were successfully applied to motion tracking data from 44 autistic and non-autistic participants. We now need to create new models on a wider repertoire of movements and larger sample size to enable identification of consistent motor patterns that will increase classification accuracy.ObjectivesObjective 1: To collect data on a wider range of motor tasks and develop robust classification tools Objective 2: To collect data and test the robustness of the models on a larger group of autistic individuals, as well as those with motor disorders (Parkinson's Disease (PD), Developmental Coordination Disorder (DCD)).Objective 3: To identify whether coordination difficulties can be divided into different subgroups, using deep clustering approaches.Approach Autistic and non-autistic adults will perform different actions (e.g. grasping a cup, walking, balancing) while motion sensors track their movements. Algorithms will be developed to extract the relevant movement parameters for each task. First, group means and discriminative features for classification will be identified for each movement task using feature extraction and selection methods. Second, supervised learning, classification and deep neural learning methods will be investigated to generate robust classifiers, which enable the detection of autistics from controls. We will continue our recent developments on interpretable deep learning (understanding why the network made a particular decision) by comparing variability based classical statistics with deep learning approaches. New autistic and non-autistic participants will then be used to test the robustness of models along with participants with PD and DCD. Clustering techniques such as K-nearest neighbours will be used to identify any subgroups within the autistic data, which differ from non-autistic individuals. Novelty/Potential outcomes The result will be a prototype ML/artificial intelligence tool to identify those at risk from having autism, supporting clinical decision making and leading to earlier and quicker diagnosis. In addition, the project will develop novel analytical science tools using ML and deep neural learning to create robust models. The research aligns with the grand challenge within the Healthcare Technologies theme: "Optimising Treatment: Optimising care through effective diagnosis, patient-specific prediction and evidence-based intervention"