基于纵向体检数据的肌少症预测研究

岳益兵; 于颖; 沈磊; 王燕; 王莹莹; 吕伟波; 刘闯

doi:10.16462/j.cnki.zhjbkz.2023.11.009

基于纵向体检数据的肌少症预测研究

doi: 10.16462/j.cnki.zhjbkz.2023.11.009

岳益兵¹,
于颖²,
沈磊³,
王燕³,
王莹莹⁴,
吕伟波^2, ,,
刘闯^1, ,

1.
杭州师范大学阿里巴巴商学院，杭州 311121
2.
上海中医药大学护理学院，上海 201203
3.
上海市浦东新区唐镇社区卫生服务中心，上海 201210
4.
安徽医科大学公共卫生学院，合肥 230032

基金项目:

国家自然科学基金 61873080

详细信息

通讯作者:
刘闯，E-mail: liuchuang@hznu.edu.cn

吕伟波，E-mail: Lyu_weibo@shutcm.edu.cn

中图分类号: R685;TP181
计量
- 文章访问数: 223
- HTML全文浏览量: 73
- PDF下载量: 35
- 被引次数: 0
出版历程
- 收稿日期: 2021-10-26
- 修回日期: 2022-03-22
- 网络出版日期: 2023-11-20
- 刊出日期: 2023-11-10

Prediction of sarcopenia based on longitudinal physical examination data

YUE Yibing¹,
YU Ying²,
SHEN Lei³,
WANG Yan³,
WANG Yingying⁴,
LYU Weibo^{2
, ,},
LIU Chuang^{1
, ,}

1.
Alibaba Business School, Hangzhou Normal University, Hangzhou 311121, China
2.
School of Nursing, Shanghai University of Traditional Chinese Medicine, Shanghai 201203, China
3.
Tangzhen Community Health Care Center, Pudong New Area Shanghai, Shanghai 201210, China
4.
School of Public Health, Anhui Medical University, Hefei 230032, China

Funds:

National Natural Science Foundation of China 61873080

More Information

Corresponding author: LYU Weibo, E-mail: Lyu_weibo@shutcm.edu.cn; LIU Chuang, E-mail: liuchuang@hznu.edu.cn

摘要

摘要: 目的肌少症是老年群体的常发疾病，高效地识别肌少症是保障老年群体健康的重要手段。方法基于上海市某医院共计2 544例被试者，根据2013―2019年选取纵向体检数据，结合年度指标的差值作为特征，构建不同的机器学习模型对老年人肌少症风险进行预测，并使用决策曲线分析，为临床决策者提供参考。结果预测模型结果显示，基于Light Gradient Boosting Machine(LightGBM)模型的预测准确率较高，受试者工作特征(receiver operating characteristic, ROC)曲线下面积(area under curve, AUC)达到0.913 4。决策曲线分析(decision curve analysis, DCA)表明，阈概率(判定肌少症的阈值)在0.01~0.42和0.84~0.92内，LightGBM模型的净收益较大；阈概率在0.42~0.50和0.60~0.67内，随机森林(random forest, RF)模型净收益较大；阈概率在0.50~0.60和0.67~0.84内，逻辑回归(logistic regression, LR)模型净收益较大。结论基于纵向体检数据和机器学习方法建立的预测模型可以有效地预测老年人未来肌少症的患病风险，对肌少症的早期诊断和干预具有重要价值。
- 肌少症 /
- 机器学习 /
- 纵向体检数据 /
- 疾病预测
Abstract: Objective In the elderly, sarcopenia is a common disease and efficient identification of sarcopenia is very important to keep health. Methods Based on the longitudinal physical examination data, a total of 2 544 subjects in a hospital from Shanghai during 2013-2019 were included. Considering the difference across annual indices, various machine learning models were constructed to predict the risk of sarcopenia in the elderly, and the decision curve analysis was applied to provide reference for clinical decision makers. Results The prediction models results showed that the prediction accuracy based on the Light Gradient Boosting Machine (LightGBM) model was relatively high (AUC=0.913 4). Decision curve analysis indicated that the net profit of the LightGBM model became larger when threshold probability (threshold for judging sarcopenia) ranged from 0.01 to 0.42 and from 0.84 to 0.92. And the net profit of the Random Forest model was larger when threshold probability ranged from 0.42 to 0.50 and from 0.60 to 0.67, while in the case of logistic regression model, the range was located in 0.50-0.60 and 0.67-0.84. Conclusions The prediction model established based on longitudinal physical examination data and machine learning methods can effectively predict the future risk of sarcopenia in the elderly, and is of great value for the early diagnosis and intervention of sarcopenia.
- Sarcopenia /
- Machine learning /
- Longitudinal physical examination data /
- Disease prediction

HTML全文

图 1 3种机器学习算法ROC曲线

1. LR，逻辑回归算法；RF，随机森林算法；LightGBM，LightGBM; AUC, 曲线下面积。
2. 实线为加入差值特征后预测效果，虚线为不引入差值特征的预测结果。

Figure 1. ROC curves of three machine learning algorithms

1. LR, logistic regression; RF, random forest; LightGBM, light gradient boosting machine; AUC, area under curve.
2. The solid line is the prediction effect after adding the difference feature, and the dashed line is the prediction result without adding the difference feature.

下载: 全尺寸图片幻灯片

图 2 3种机器学习算法决策分析曲线

LR, 逻辑回归; RF, 随机森林; LightGBM, LightGBM; ALL, 干预所有人的净收益; NONE, 不干预任何人的净收益。

Figure 2. The decision curve analysis curves of three machine learning

LR, logistic regression; RF, random forest; LightGBM, light gradient boosting machine; ALL, net benefits of intervening for all; NONE, net benefits of intervening for none.

下载: 全尺寸图片幻灯片

表 1 重复测量方差分析结果

Table 1. The results of repeated measures analysis of variance

特征Characteristics	组间效应Between-group effects		组内效应Within-group effects		交互效应Interaction effects
特征Characteristics	F值value	P值value	F值value	P值value	F值value	P值value
体重/kg Weight/kg	631.946	＜0.001	85.868	＜0.001	9.118	＜0.001
身高/cm Height/cm	79.533	＜0.001	490.441	＜0.001	1.639	0.132
BMI/(kg·m^-2)	606.510	＜0.001	39.789	＜0.001	7.937	＜0.001
腰围/cm Waist circumference/cm	388.855	＜0.001	48.777	＜0.001	4.539	＜0.001
臀围/cm Hip circumference/cm	470.593	＜0.001	90.444	＜0.001	9.861	＜0.001
心率/bpm Heart rate/bpm	17.853	＜0.001	96.120	＜0.001	2.105	0.049
SBP/mmHg	13.678	＜0.001	19.885	＜0.001	0.987	0.433
DBP/mmHg	56.915	＜0.001	131.226	＜0.001	1.713	0.114

下载: 导出CSV

表 2 3种预测模型在测试集上分类结果

Table 2. Three prediction models classify results on the test set

预测结果Predicted results	实际结果 ^① Actual results ^①		合计 ^① Total ^①
预测结果Predicted results	2019年罹患肌少症 Sarcopenia in 2019	2019年未患肌少症 Non-sarcopenia in 2019	合计 ^① Total ^①
LR-罹患肌少症LR-sarcopenia	52(61.90)	22(5.19)	74(14.57)
LR-未患肌少症LR-non-sarcopenia	32(38.10)	402(94.81)	434(85.43)
RF-罹患肌少症RF-sarcopenia	51(60.71)	20(4.72)	71(13.98)
RF-未患肌少症RF-non-sarcopenia	33(39.29)	404(95.28)	437(86.02)
LightGBM-罹患肌少症LightGBM-sarcopenia	48(57.14)	17(4.01)	65(12.80)
LightGBM-未患肌少症LightGBM-non-sarcopenia	36(42.86)	407(95.99)	443(87.20)
合计Total	84(100.00)	424(100.00)	508(100.00)
注：1. LR，逻辑回归；RF，随机森林；LightGBM，LightGBM。 2. 算法预测是否患肌少症的人数，实际是否患肌少症的人数。 ①以[人数(占比/%)] 表示。 Note: 1. LR, logistic regression; RF, random forest; LightGBM, light gradient boosting machine. 2. The algorithm predicts whether the number of people with sarcopenia，The actual number of people with sarcopenia. ① [Number of people (proportion/%)].

下载: 导出CSV

参考文献(9)

[1]	Wiedmer P, Jung T, Castro JP, et al. Sarcopenia- Molecular mechanisms and open questions[J]. Ageing Res Rev, 2021, 65: 101200. DOI: 10.1016/j.arr.2020.101200.
[2]	Beaudart C, Zaaria M, Pasleau F, et al. Health outcomes of sarcopenia: a systematic review and meta-analysis[J]. PLoS One, 2017, 12(1): e0169548. DOI: 10.1371/journal.pone.0169548.
[3]	Bruyère O, Beaudart C, Ethgen O, et al. The health economics burden of sarcopenia: a systematic review[J]. Maturitas, 2019, 119: 61-69. DOI: 10.1016/j.maturitas.2018.11.003.
[4]	Rajkomar A, Dean J, Kohane I. Machine learning in medicine[J]. N Engl J Med, 2019, 380(14): 1347-1358. DOI: 10.1056/NEJMra1814259.
[5]	欧阳平, 李小溪, 冷芬, 等. 机器学习算法在体检人群糖尿病风险预测中的应用[J]. 中华疾病控制杂志, 2021, 25(7): 849-853, 868. DOI: 10.16462/j.cnki.zhjbkz.2021.07.020. Ou YP, Li XX, Leng F, et al. Application of machi-ne learning algorithm in diabetes risk prediction of physical examination population[J]. Chin J Dis Control Prev, 2021, 25(7): 849-853, 868. DOI: 10.16462/j.cnki.zhjbkz.2021.07.020.
[6]	Lavalley MP. Logistic regression[J]. Circulation, 2008, 117(18): 2395-2399. DOI: 10.1161/circulationaha.106.682658.
[7]	Liu ZS, Siu WC, Chan YL. Features guided face super-resolution via hybrid model of deep learning and random forests[J]. IEEE Trans Image Process, 2021, 30: 4157-4170. DOI: 10.1109/TIP.2021.3069554.
[8]	Ke GL, Meng Q, Finley T, et al. LightGBM: a highly efficient gradient boosting decision tree[J]. NeurIPS Curran Associates Inc, 2017: 3149-3157.
[9]	Vickers AJ, Elkin EB. Decision curve analysis: a novel method for evaluating prediction models[J]. Med Decis Making, 2006, 26(6): 565-574. DOI: 10.1177/0272989X06295361.