STTR Phase I: Development of a Safety System for Individuals with Alzheimer's Disease and Related Dementias

这个小企业技术转移（STTR）第一阶段项目的更广泛影响/商业潜力是一个安全系统，用于提高痴呆症护理的质量和降低成本。阿尔茨海默病在美国影响了540万人，其中65岁以上的9人中有1人，85岁以上的3人中有1人，占所有痴呆症患者的三分之二。尽管阿尔茨海默病是美国最昂贵的疾病，跌倒是阿尔茨海默病治疗住院的主要原因，但目前的工具提供的支持很少。尽管四分之三的老年人会重复跌倒，但床上报警器和可穿戴式跌倒检测系统等解决方案无法看到跌倒是如何发生的。护理人员无法从第一次跌倒中吸取教训，以减少第二次跌倒的可能性，因此必须实施痛苦而昂贵的政策，例如将每一次无人目击的跌倒都送到急诊室，以防头部受到撞击。拟议中的项目通过基于摄像机视频检测跌倒，解决了阿尔茨海默氏症护理中的这一关键空白，在这种情况下，跌倒可以由人类助手实时和事后进行检查。如果发生头部撞击，实时审查允许即时通知，并且事后审查允许确定坠落的原因，以查看是否可以更改房间布局和/或政策。该项目的主要目的是收集真实跌倒的视频数据，1)应用和扩展最先进的深度学习方法来执行高精度检测，2)验证受影响的个人、家庭和护理人员是否接受基于摄像头的解决方案。摔倒检测将通过扩展基于区域的卷积神经网络（RCNN）算法来执行，该算法使用域自适应技术开发，以强大地处理夜视相机操作，遮挡和非标准人体姿势。技术成功将由1%的漏检率和50%的假阳性率来衡量。第一个准确性阈值将定义一个下限，正如深度学习范式中反复证明的那样，随着收集到的数据越来越多，准确性将继续提高。

The broader impact/commercial potential of this Small Business Technology Transfer (STTR) Phase I project is a safety system for improving the quality and reducing the cost of dementia care. Alzheimer's disease affects 5.4M in the US, including 1 in 9 over 65 and 1 in 3 over 85, and represents two thirds of all those affected by dementia. Despite that Alzheimer's disease is the single most expensive disease in the US and falls are the leading cause of hospitalization in Alzheimer's care, current tools offer little support. Although 3/4 of elderly fallers will experience a repeat fall, solutions like bed alarms and wearable fall detection systems offer no way to see how falls occur. Care staff have no way of learning from the first fall to reduce the likelihood of the second and must implement painful and expensive policies such as sending every unwitnessed fall to the emergency room in case a hit to the head occurred. The proposed project addresses this critical gap in Alzheimer's care by detecting falls based on camera video where falls can be reviewed by a human assistant in real-time and after the fact. Real-time review allows for instant notification if a hit to the head occurred, and review after the fact allows for determining the cause of the fall to see if changes in room layout and/or policy could be made. The primary aim of this project is to collect video data of real falls 1) to apply and extend state-of-the-art deep learning methods to perform high accuracy detection and 2) to validate that affected individuals, family, and care staff are accepting of a camera-based solution. Fall detection will be performed by extending the Region-Based Convolutional Neural Network (RCNN) algorithm using domain adaptation techniques developed to robustly handle night-vision camera operation, occlusion, and non-standard human pose. Technical success will be measured by 1% missed detection and 50% false positive rate from this feasibility study. This first accuracy threshold will define a lower bound where, as has been demonstrated repeatedly in the deep-learning paradigm, accuracy will continue to improve as more data is collected.