基于Group Lasso的Logistic回归模型构建绝经后骨质疏松性骨折初发风险评估工具

Establishment of risk assessment tool for postmenopausal osteoporotic fractures based on Group Lasso's logistic regression model

目的构建符合北京、上海两地40~65岁女性人口学特征的危险因素和中医症状相结合的骨质疏松性骨折早期风险预测工具。方法本研究采用注册登记式研究的方法,对2009年3月-8月在北京市东城区及上海市徐汇区收集的1129例40~65岁女性骨质疏松症高危人群的危险因素及中医症状信息,进行连续3年的登记观察。采用SMOTE过抽样算法平衡数据,基于Group Lasso的Logistic回归模型筛选与骨质疏松症骨折有关的危险因素及中医症状,建立骨质疏松性骨折风险评估工具。结果基于R 3.3.3软件的Grplasso包,在不同λ水平上,我们进行对绝经后骨质疏松性骨折的危险因素与中医证候要素的学习。最终结合数理与医理,认为λ=0.0235时遴选出的变量最佳。具体组变量包括:骨密度(bone mineral density,BMD)、年龄、食物类、身高、月经情况、孕产次数和肝肾阴虚。进而基于Logistic回归模型得出骨质疏松性骨折预测工具:P=-1.88+0.437*BMD+0.289*年龄+0.023*大米面条-0.007*奶制品-0.096*豆制品-0.128*肉类-0.084*鱼类-0.007*新鲜蔬菜-0.018*蛋类+0.047*海藻类+0.048*身高-0.035*是否变矮-0.081*初潮年龄+0.171*是否绝经+0.121*绝经年限+0.039*怀孕次数+0.192*生产次数-0.056*子宫卵巢是否切除+0.05*手足烦热-0.094*盗汗+0.008*腿软+0.15*目眩-0.048*视物模糊-0.045*目睛干涩-0.089*恶热+0.08*脱发+0.034*齿摇-0.101*口苦+0.004*易怒+0.054*午后潮热-0.056*失眠+0.019*多梦易惊-0.02*胸胁苦满+0.137*下肢转筋。对该预测模型预测概率绘制受试者工作特征曲线,结果显示曲线下面积为0.8775(95%CI=0.8412~0.9138)。结论初步建立了基于北京、上海人口学特征40~65岁女性骨质疏松性骨折早期风险预测工具。

Objective To establish an early risk prediction tool for osteoporotic fractures that combines demographic risk factors and TCM symptoms in females aged between 40 and 65 years in Beijing and Shanghai. Methods In this study,risk factors and TCM symptoms of 1129 women aged 40 to 65 years at high risk of osteoporosis were collected in Dongcheng District of Beijing andXuhui District of Shanghai from March to August 2009 using a registered study method. Observational registration continued for three consecutive years. The SMOTE oversampling algorithm was used to balance the data. Group Lasso＇s logistic regression model was used to screen for risk factors associated with osteoporosis and TCM symptoms,and a risk assessment tool for osteoporotic fractures was established. Results Based on the grplasso package of R 3. 3. 3 software,we studied risk factors of postmenopausal osteoporotic fractures and TCM syndrome factors at different λ levels. Finally,combining mathematics and medical science,we believe that the variables selected are the best when λ = 0. 0235. The specific group variables include： bone mineral density,age,food type,height,menstruation,number of births,and liver and kidney yin deficiency. Further,based on the logistic regression model,osteoporotic fracture prediction tools were derived： P =-1. 88 ＋ 0. 437＊ BMD ＋ 0. 289＊ age ＋ 0. 023＊ rice noodles-0. 007＊ dairy products-0. 096＊ soy products-0. 128＊ meat-0. 084＊ fish-0. 007＊ fresh vegetables-0. 018＊ eggs ＋ 0. 047＊ seaweeds ＋ 0. 048＊ height-0. 035 ＊ whether became shorter-0. 081 ＊ menarche age ＋ 0. 171 ＊ whether menopause ＋0. 121＊ menopause years ＋ 0. 039＊ number of pregnancy ＋ 0. 192＊ number of birth-0. 056＊ whether the uterus and ovary were resected ＋ 0. 05＊ fever of hands and feet-0. 094＊ night sweat ＋ 0. 008＊ soft legs ＋ 0. 15＊ dizziness-0. 048＊ blurred vision-0. 045＊ dry eyes-0. 089 ＊ bad heat ＋ 0. 08 ＊ hair loss ＋ 0. 034 ＊ loose teeth-0. 101 ＊ bitter mouth ＋ 0. 004 ＊irritable ＋ 0. 054＊ afternoon hot flash-0. 056＊ insomnia ＋ 0. 019＊ dreaminess and easily frightened-0. 02＊ chest thrush ＋0. 137＊ lower limb spasm. The receiver operating characteristic curve was plotted against the predictive probability of the predictive model. The result showed that the area under the curve was 0. 8775（95% CI = 0. 8412-0. 9138）. Conclusion A preliminary risk prediction tool for osteoporotic fractures in 40 to 65-year-old females based on the demographic characteristics of Beijing and Shanghai was established.