SwagataDas_PlosOne_202102.zip

We propose a feed-forward Artificial Neural Network (ANN)-based lower limb assessment method using squat and one-leg standing exercise features. An Intel Realsense Depth Camera was used to track and record the 3D skeletal data, from which we extracted nine squat features and four one-leg standing features. We used those features as the input layer to train our ANN. We base our assessment method to precisely identify the progression of \textit{Locomotive Syndrome (LS)} in adults. The Japanese Orthopaedic Association (JOA) first defined LS as a mobility disorder that occurs from the degeneration of locomotive organs and makes nursing care mandatory. A self-test called Short Test Battery Locomotive Syndrome (STBLS) was also introduced to identify LS's progression in adults. Therefore, the output layer for training the ANN was obtained using the STBLS test to detect the progression of LS. Three assessment scores were obtained through this test: stand-up, 2-stride, and Geriatric Locomotive Function Scale (GLFS-25). The ANN consisted of 2 hidden layers with six nodes per layer. We used the Rectified-Linear-Unit (ReLU) activation function. We observed that the correct choice of the input features to train the ANN was essential to achieve an acceptable risk prediction accuracy. The stand-up and 2-stride scores of the STBLS test could be predicted with correlation coefficients of 0.59 and 0.76 between the real and predicted data, respectively.<br>

本研究提出一种基于前馈式人工神经网络（Artificial Neural Network, ANN）的下肢评估方法，该方法利用深蹲与单腿站立动作的特征参数。本研究采用英特尔实感深度相机（Intel Realsense Depth Camera）采集并记录三维骨骼数据，从中提取9项深蹲动作特征与4项单腿站立动作特征。将上述特征作为输入层参数，用于训练所提出的人工神经网络。本评估方法旨在精准识别成人运动综合征（Locomotive Syndrome, LS）的进展程度。日本骨科协会（Japanese Orthopaedic Association, JOA）首次将LS定义为因运动器官退行性病变引发、需依赖护理的运动功能障碍。同时，该协会还推出了用于成人LS进展程度筛查的自我评估工具——运动综合征简评量表（Short Test Battery Locomotive Syndrome, STBLS）。因此，本研究将STBLS测试结果作为人工神经网络训练的输出层标签，用于LS进展程度的检测。该测试共包含三项评估评分：起立评分、两步行走评分与老年运动功能量表（Geriatric Locomotive Function Scale, GLFS-25）。所构建的人工神经网络包含2个隐藏层，每层设置6个神经元节点。本研究采用修正线性单元（Rectified-Linear-Unit, ReLU）作为激活函数。实验结果表明，合理选择训练人工神经网络的输入特征，是获得可接受风险预测精度的关键。STBLS测试的起立评分与两步行走评分的真实值与预测值之间的相关系数分别为0.59和0.76。