Sakane Naoki, Yamauchi Ken, Kutsuna Ippei, Suganuma Akiko, Domichi Masayuki, Hirano Kei, Wada Kengo, Ishimaru Masashi, Hosokawa Mitsuharu, Izawa Yosuke, Matsumura Yoshihiro, Hozumi Junichi
Division of Preventive Medicine, Clinical Research Institute, National Hospital Organization Kyoto Medical Center, 1-1 Mukaihata-cho, Fukakusa, Fushimi-ku, Kyoto 612-8555, Japan.
Institute of Physical Education, Keio University, 4-1-1 Hiyoshi, Kohoku-ku, Yokohama, Kanagawa 223-8521, Japan.
J Occup Health. 2025 Jan 7;67(1). doi: 10.1093/joccuh/uiae075.
Falls are among the most prevalent workplace accidents, necessitating thorough screening for susceptibility to falls and customization of individualized fall prevention programs. The aim of this study was to develop and validate a high fall risk prediction model using machine learning (ML) and video-based first 3 steps in middle-aged workers.
Participants to provide training data (n = 190, mean [SD] age = 54.5 [7.7] years, 48.9% male) and validation data (n = 28, age = 52.3 [6.0] years, 53.6% male) were enrolled in this study. Pose estimation was performed using a marker-free deep pose estimation method called MediaPipe Pose. The first 3 steps, including the movements of the arms, legs, trunk, and pelvis, were recorded using an RGB camera, and the gait features were identified. Using these gait features and fall histories, a stratified k-fold cross-validation method was used to ensure balanced training and test data, and the area under the curve (AUC) and 95% CI were calculated.
Of 77 gait features in the first 3 steps, we found 3 gait features in men with an AUC of 0.909 (95% CI, 0.879-0.939) for fall risk, indicating an "excellent" (0.9-1.0) classification, whereas we determined 5 gait features in women with an AUC of 0.670 (95% CI, 0.621-0.719), indicating a "sufficient" (0.6-0.7) classification.
These findings suggest that fall risk prediction can be developed based on ML and the first 3 steps in men; however, the accuracy was only "sufficient" in women. Further development of the formula for women is required to improve its accuracy in the middle-aged working population.
跌倒属于最常见的工作场所事故之一,因此有必要对跌倒易感性进行全面筛查,并定制个性化的跌倒预防方案。本研究的目的是使用机器学习(ML)和基于视频的中年工人前三步动作来开发并验证一个高跌倒风险预测模型。
本研究纳入了提供训练数据的参与者(n = 190,平均[标准差]年龄 = 54.5 [7.7]岁,男性占48.9%)和验证数据的参与者(n = 28,年龄 = 52.3 [6.0]岁,男性占53.6%)。使用一种名为MediaPipe Pose的无标记深度姿态估计方法进行姿态估计。使用RGB相机记录包括手臂、腿部、躯干和骨盆运动的前三步动作,并识别步态特征。利用这些步态特征和跌倒史,采用分层k折交叉验证方法确保训练和测试数据的平衡,并计算曲线下面积(AUC)和95%置信区间。
在前三步的77个步态特征中,我们发现男性有3个步态特征,其跌倒风险的AUC为0.909(95%置信区间,0.879 - 0.939),表明分类为“优秀”(0.9 - 1.0),而我们确定女性有5个步态特征,AUC为0.670(95%置信区间,0.621 - 0.719),表明分类为“充分”(0.6 - 0.7)。
这些发现表明,可以基于机器学习和男性的前三步动作来开发跌倒风险预测模型;然而,在女性中其准确性仅为“充分”。需要进一步改进针对女性的公式,以提高其在中年工作人群中的准确性。