【摘要】 目的：采用随机森林、支持向量机、k-近邻算法、决策树与高斯朴素贝叶斯5种机器学习算法构建覆盖全面、分类精细与准确率高的学龄前儿童身体活动（PA）类型识别模型，为PA的精准评估及运动干预提供多维度、多样化的方法参考。方法：招募31名4～6岁儿童，在其右侧髂嵴处佩戴ActiGraph GT3X-BT加速度计，连续3 d在幼儿园日常活动，同步采集视频。通过ELAN软件对视频进行标注，识别坐姿不动、坐姿活动、站姿不动、站姿活动、爬行、步行、跑步和跳跃共8类PA类型，并通过时间戳匹配加速度数据。数据按5 s、10 s、15 s窗口分割，提取22项时频特征。经数据增强与SMOTE处理后，构建5种机器学习模型。采用网格搜索法优化参数，并通过五折交叉验证评估模型的准确率、F1分数及混淆矩阵表现。结果：随机森林、决策树、支持向量机、k-近邻算法与高斯朴素贝叶斯的总体F1分数最高分别为93.3%、85.5%、82.5%、68.2%与67.0%。其中，随机森林模型在10 s 10%的重叠窗口时性能最佳，8类PA类型准确率均超过80%，且优于其余4种分类模型。随机森林模型准确率依次为站姿不动（99%）、步行（99%）、跑步（93%）、跳跃（90%）、坐姿活动（90%）、站姿活动（88%）、坐姿不动（88%）、爬行（87%）。结论：基于自由生活环境数据训练的随机森林分类模型能有效预测学龄前儿童的PA类型，为幼儿园体育课程设计、家庭体育健康干预方案以及相关体育政策的制定提供科学依据。更多还原

【Abstract】 Objective: Five machine learning techniques, including Random Forest, Support Vector Machine, KNearest Neighbor Algorithm, Decision Tree and Gaussian Plain Bayes, were used to construct a model for identifying the types of physical activity(PA) of preschool children with comprehensive coverage, fine classification and high accuracy.By this way, the study aimed to provide multidimensional and diversified tools and ideas for precise PA assessment and exercise intervention.Methods: A total of 31 children aged 4-6 years were recruited to wear ActiGraph GT3X-BT accelerometers on their right iliac crest for three consecutive days during routine activities in kindergarten, with video recorded simultaneously.The videos were annotated using the ELAN tool to identify eight PA types: sitting motionless, sitting active, standing motionless, standing active, crawling, walking, running, and jumping.The accelerometer data were matched with video annotations via timestamps.The data were segmented into 5-second, 10-second, and 15-second windows, from which 22 time-frequency features were extracted.After data augmentation and SMOTE processing, five machine learning models were constructed.Hyperparameter optimization was performed using grid search, and five-fold cross-validation was used to evaluate model accuracy, F1-scores, and confusion matrix performance.Results: The highest overall F1-scores for Random Forest, Decision Tree, Support Vector, K-Nearest Neighbor Algorithm, and Gaussian Plain Bayes were 93.3%、85.5%、82.5%、68.2% and 67.0%, respectively.The Random Forest model performed best at a 10% overlap window of 10 seconds, with over 80% accuracy for all eight PA types, and outperformed the other four classification models.The classification accuracies of the Random Forest Model were: standing motionless(99%), walking(99%), running(93%), jumping(90%), sitting active(90%), standing active(88%), sitting motionless(88%) and crawling(87%).Conclusion: The Random Forest Classification model trained on free-living environment data can effectively predict the PA types of preschool children, laying a scientific foundation for kindergarten physical education curriculum design, family-based sports and health intervention programs, and future sports policy development.更多还原