焦坤1，赵广高1，付近梅2，苏利强3，孙顺利2，何梓豪1，马欣欣1，覃雅睿1，刘晓虎1，张晓雨1，董志鑫1，王家昌1，韦有迪1，贺嘉一1
(1.南昌大学 体育学院，江西 南昌 330031；2.江西省体育局 体育科学医疗中心，江西 南昌 330006；3.江西师范大学 体育学院，江西 南昌 330022)
摘 要：运用极端梯度提升(XGB)等14种机器学习(ML)技术，分别构建包含静坐行为(SB)以及各类幼儿身体活动(PA)强度的分类模型，探索高准确率的幼儿PA强度识别办法。研究设计遵循COSMIN标准，招募50名3~6岁幼儿，在园内佩戴ActiGraph GT3X-BT并采集不同强度PA的视频数据。借助儿童PA强度评分量表(CARS)与ELAN工具将PA划分为SB、低强度PA(LPA)、中高强度PA(MVPA)3类，或SB、LPA、中强度PA(MPA)和高强度PA(VPA)4类。以加速度计原始数据为输入特征，采用5 s、10 s、15 s滑动窗口切分并进行SMOTE处理，基于XGB、支持向量机(SVM)、随机森林(RF)、卷积神经网络(CNN)、深度学习集成网络(DLEN)等14种ML技术分别构建SB-LPA-MVPA三分类和SB-LPA-MPA-VPA四分类模型。采用5折交叉验证评估各模型的表现，并对最优模型进行等效性测试。结果显示：在三分类任务中，14种模型的F1分数与总体准确率分别为80.3%~97.5%与80.6%~97.5%，XGB、SVM、RF、CNN、DLEN 5种模型的F1分数与总体准确率均高于97.0%。其中，XGB的F1分数、总体准确率以及SB(99.06%±0.31%)、LPA(95.43%±1.61%)、MVPA(97.99%±0.72%)各强度类别的准确率均位居首位。在四分类任务中，14种模型的F1分数与总体准确率分别为71.9%~97.2%与73.2%~97.2%，XGB、SVM、RF、CNN、DLEN 5种模型的F1分数与总体准确率均高于96.0%。其中，XGB的F1分数、总体准确率及SB(98.89%±0.77%)、LPA(95.51%±1.16%)、MPA(96.00%±0.72%)、VPA(98.40%±0.73%)各强度类别的准确率均居首位。等效性测试表明，XGB的预测值与实际值之间的平均偏差为－0.009 min。研究表明，XGB、SVM、RF、CNN、DLEN 5种ML技术均能够精确识别幼儿SB-LPA-MVPA与SB-LPA-MPA-VPA强度类别，其中XGB的表现最佳。
关 键 词：运动生理学；幼儿身体活动；活动强度；机器学习；加速度计
中图分类号：G804.2 文献标志码：A 文章编号：1006-7116(2026)03-0139-10
Identification of physical activity intensity in young children:Based on 14 machine learning technologies
JIAO Kun1，ZHAO Guanggao1，FU Jinmei2，SU Liqiang3，SUN Shunli2，HE Zihao1，MA Xinxin1，QIN Yarui1，LIU Xiaohu1，ZHANG Xiaoyu1，DONG Zhixin1，WANG Jiachang1，WEI Youdi1，HE Jiayi1
(1.School of Physical Education，Nanchang University，Nanchang 330031，China；2.Sports Science and Medical Center，Jiangxi Provincial Sports Bureau，Nanchang 330006，China；3.School of Physical Education，Jiangxi Normal University，Nanchang 330022，China)
Abstract: By using 14 machine learning (ML) techniques, from extreme gradient enhancement (XGB) to other techniques, classification models were respectively developed to cover sedentary behavior (SB) in young children and various physical activity (PA) intensity, which has designed to explore high-precision identification methods for PA intensity for this population. The study adhered to the COSMIN guidelines, and fifty children aged 3–6 years were recruited and wore ActiGraph GT3X-BT accelerometers in their preschool setting, while their PA was video-recorded. Using the Children’s Activity Rating Scale (CARS) and the ELAN tool, PA was categorized into either three groups (SB, light-intensity PA [LPA], and moderate-to-vigorous PA [MVPA]) or four groups (SB, LPA, moderate PA [MPA], and vigorous PA [VPA]). Using raw accelerometer data as input features, 5 s, 10 s, and 15 s sliding windows were applied for data segmentation, followed by Synthetic Minority Over-sampling Technique (SMOTE) processing. 14 ML techniques, including XGB, support vector machine (SVM), random forest (RF), convolutional neural network (CNN), and deep learning ensemble network (DLEN), were used to construct three-class (SB-LPA-MVPA) and four-class (SB-LPA-MPA-VPA) classification models. Model performance was evaluated using 5-fold cross-validation, and an equivalence test was conducted on the optimal model. Results show that for three-class models respectively, the F1-scores of fourteen models change from 80.3% to 97.5%, and their total accuracy changes from 80.6% to 97.5%. Five models including XGB, SVM, RF, CNN, DLEN, have both F1-scores and total accuracy over 97.0%. Among them, XGB gets the first position in F1-score, total accuracy, and accuracy for SB, LPA, MVPA intensity classes, which are (99.06% ± 0.31%), (95.43% ± 1.61%) and (97.99% ± 0.72%) in sequence. For four-class models, F1-scores are in the range from 71.9% to 97.2%, and total accuracy is in the range from 73.2% to 97.2%. Therefore, the same five models, XGB, SVM, RF, CNN, DLEN, have both F1-scores and total accuracy over 96.0%. Hence, XGB again ranks first in F1-score, total accuracy, and accuracy for SB, LPA, MPA, VPA intensity classes, which are (98.89% ± 0.77%), (95.51% ± 1.16%), (96.00% ± 0.72%), (98.40% ± 0.73%), and the equivalence test results show that the average bias between XGB’s predictive values and real values is –0.009 min. The conclusion holds that five machine learning techniques including XGB, SVM, RF, CNN, DLEN could carry out effective identification of SB-LPA-MVPA and SB-LPA-MPA-VPA intensity categories in young children, and among them, the XGB showing the highest performance level.
Keywords: sports physiology；physical activity of young children；activity intensity；machine learning；accelerometer