AI at the edge: From movement models to neurological outcome

项目目标：-开发最先进的模型，使用人工智能从医院视频流中提取人体姿势-使用循环神经网络和图形神经网络表征人体运动的时间成分-了解运动模型与脑成像特征之间的因果关系-共同使用运动模型，项目描述：人类步态是一个高维系统，针对健康年轻人的准确性、稳定性、速度或能量效率进行了优化。随着大脑老化，慢性脑缺血和脑萎缩，运动特性改变，最终步态和平衡受到损害，导致跌倒的风险。以前的研究是有限的——相对较少的步态模式分析，在非生态环境下的测量，如步态实验室和由于计算限制的尺寸最小化。生物医学工程学院的MoCat（运动表征）团队开发了一种低成本的便携式运动捕捉系统，该系统使用现代机器学习方法，具有机载计算能力，可以实现真实世界的数据捕获和人体运动的机载计算。博士候选人将与MoCat团队合作，开发新的最先进的专用人工智能算法，以表征从临床环境（KHP医院的神经病学和中风单位）招募的任何脑血管疾病患者的全身运动和步态，以开发有关步态和跌倒风险的高维性质的新见解。目前的通用身体跟踪模型针对拥挤环境下的高召回率进行了优化，并且对单摄像机视频的3D身体姿势预测的需求有限。本项目开发的算法将引入单摄像机视图的时间一致性和3D身体姿势预测，多摄像机的姿势立体匹配，以及人/人和人/房间交互。在提取出人体的时间一致向量表示后，循环图神经网络将用于表征和聚类运动，诊断运动障碍，并预测与正常（健康）人体运动的偏差。最后，我们将使用该队列的大量子集（患有短暂性缺血性发作- TIA的患者），这些患者具有医学成像数据（MRI/CT），以了解脑病变如何影响纤维连接并随后导致观察到的运动障碍，从而能够从成像数据中预测长期疾病影响。

Aims of the project:- To develop state-of-the-art models that extract human pose using AI from in-hospital video streams - To characterise the temporal component of human movement using recurrent and graph neural networks - Understand the causal relationships between movement models and brain imaging features - Jointly use movement models, imaging and neurological tests to diagnose and prognose neurological conditions Project Description:Human gait is a high-dimensional system optimised for accuracy, stability, speed or energy efficiency in healthy young adults. As the brain ages through chronic cerebral ischaemia and atrophy, properties of motion changes and eventually gait and balance are compromised resulting in a falls risk. Previous studies have been limited - relatively small number of gait patterns analysed, measurements in non-ecological environments like gait laboratories and dimensional minimisation due to computational limits. The MoCat (motion characterisation) team at the school of biomedical engineering has developed a low-cost portable motion capture system with on-board computing which enables real-world data capture and on-board computation at-scale of human motion, using modern machine learning approaches. The PhD candidate will work with the MoCat team and develop new state-of-the-art purpose-built AI algorithms to characterise the whole body motion and gait of patients with any cerebral small vessel disease recruited from clinical environments (the neurology and stroke units of KHP hospitals) to develop novel insights on the high-dimensional nature of gait and falls risk. Current general-purpose body-tracking models are optimised for high recall in a crowded setting, and have limited need for 3D body-pose prediction from single camera video. The algorithms developed in this project will introduce time-consistency and 3D body pose prediction from single camera views, stereoscopic matching of pose from multiple cameras, and human/human and human/room interaction. After a time-consistent vectorial representation of the human body is extracted, recurrent graph neural networks will be used to characterise and cluster movements, diagnose movement disorders, and predict deviations from normal (healthy) human movement. Lastly, we will use a substantial subset of the cohort (patients suffering from a transient ischaemic attack - TIA) which have medical imaging data (MRI/CT) to understand how brain lesions affect fibre connectivity and subsequently result in observed movement impairments, enabling the prediction of long term disease effects from imaging data.