Assessment of individualized treatment effect of antidiabetic prescriptions for type 2 diabetes based on uplift model
-
摘要:
目的 应用uplift模型评价现实世界2型糖尿病用药处方个性化疗效，识别受益个体特征。 方法 以2012年1月1日至2017年12月31日山东省胶南市“全人群高血压、糖尿病综合防治项目”中的2型糖尿病管理人群为研究对象。干预组三种用药处方包括二甲双胍、格列吡嗪及二甲双胍与格列吡嗪联合用药；对照组为不用药组；结局为观察期内最后一次FPG测量值是否达标。根据倾向得分匹配法，按照1∶1匹配，模拟随机对照试验，用uplift模型评价个性化疗效，并识别受益个体特征。 结果 队列共纳入5 652人，年龄(64.20±11.48)岁；男性2 239名，占比39.61%；二甲双胍、格列吡嗪及联合用药组人数分别为1 707人、321人及535人，不用药组3 089人。3种降糖处方处理组和对照组的倾向性评分在匹配后基本达到平衡，各组血糖控制率与不用药组差异均无统计学意义。但uplift模型个性化疗效评价显示，三种降糖处方分别对组内68.59%、65.73%及51.89%患者有效，净效益增长较随机干预分别提高8.24%、9.60%和10.53%。 结论 uplift模型有助于评价个性化效应，为2型糖尿病个性化用药的受益个体特征识别提供参考。 Abstract:Objective To evaluate the individualized treatment effect of antidiabetic medications for type 2 diabetes in the real world and to identify the benefited individuals from each prescription as well as their characteristics by uplift modeling. Methods Data was collected from the Comprehensive Intervention Program of Chronic Diseases in Jiaonan, Shandong Province from Jannary 1, 2012 to Delember 31, 2018. Patients with type 2 diabetes were included in this study. The intervention groups were given three prescriptions, including metformin, glipizide and metformin combined with glipizide, while the control group was given no medication. The outcome was whether the last FPG measured during the observation period reached standard. According to the propensity score matching method with a match ratio of 1∶1, the simulated randomized controlled trial was conducted to evaluate the average effect of three prescriptions. The uplift model was used to evaluate the individual treatment effect and to identify the characteristics of the individuals who benefit from treatment. Results A total of 5 652 people were included in the cohort, with an average age of (64.20±11.48) years old, 2 239 males (39.61%). There were 1 707 patients in the metformin group, 321 patients in the glipizide group, 535 patients in the metformin combined with glipizide group, and 3 089 patients in the non-treatment group. The propensity scores of the three hypoglycemic prescriptions groups and the control group were basically balanced after matching. There were no statistical difference in blood glucose control rate between the three groups and the control group. However, the individual treatment effect evaluation based on the uplift model showed that all of the three drug prescriptions were more effective in some patients, with a ratio of 68.59%, 65.37% and 51.89%, respectively. What's more, three hypoglycemic prescriptions had a cumulative incremental effect of over 8.24%, 9.60% and 10.53% compared with random intervention, respectively. Conclusion According to the uplift model, the personalized effect can be evaluated, which is helpful to provide reference for the characteristic identification of personalized medication for type 2 diabetes. -
Key words:
- Type 2 diabetes /
- Real world research /
- Uplift model /
- Individualized treatment effect
-
图 1 二甲双胍、格列吡嗪及联合用药受试对象在处理组和对照组中的血糖控制达标概率分布a
注：aX轴表示受试者进入对照组血糖控制达标的概率，Y轴表示受试者进入处理组血糖控制达标的概率。
Figure 1. The probability distribution of achieving blood glucose control among subjects with metformin, glipizide, or combination in the treatment and control groups a
图 2 二甲双胍、格列吡嗪及联合用药的uplift随机森林模型Qini曲线
Figure 2. Qini curves of the uplift random forest model of metformin, glipizide, and combination
表 1 三种降糖处方人群基线特征a [n(%)]
Table 1. Characteristics of study participants according to 3 hypoglycemic prescriptions at baseline a [n(%)]
变量 不用药(n=3 089) b 二甲双胍 格列吡嗪 联合用药 用药(n=1 707) c t/χ2值 P值 用药(n=321) t/χ2值 P值 用药(n=535) t/χ2值 P值 年龄(x±s, 岁) 64.68±11.74 64.61±11.50 -0.216 0.829 61.88±9.91 -4.721 ＜0.001 61.58±10.24 -6.328 ＜0.001 性别 13.534 ＜0.001 1.060 0.303 1.958 0.162 男 1 271(41.1) 608(35.7) 122(38.0) 238(44.5) 女 1 818(58.9) 1 096(64.3) 199(62.0) 297(55.5) BMI(x±s, kg/m2) 25.07±3.36 25.27±3.11 2.141 0.032 24.81±2.88 -1.514 0.131 25.02±2.88 -0.308 0.758 吸烟 1.285 0.257 0.571 0.450 10.285 0.001 有 506(16.4) 258(15.1) 47(14.6) 119(22.2) 无 2 570(83.6) 1 445(84.9) 274(85.4) 416(77.8) 饮酒 7.863 0.005 7.001 0.008 0.426 0.514 有 368(12.0) 158(9.3) 22(6.9) 70(13.1) 无 2 705(88.0) 1 545(90.7) 299(93.1) 465(86.9) 运动 1.494 0.222 21.187 ＜0.001 64.684 ＜0.001 有 1 762(57.2) 942(55.3) 140(43.6) 205(38.3) 无 1 319(42.8) 761(44.7) 181(56.4) 330(61.7) 职业 64.226 ＜0.001 63.037 ＜0.001 95.475 ＜0.001 农民 1 990(64.5) 1 290(75.7) 278(86.6) 460(86.0) 非农民 1 097(35.5) 413(24.3) 43(13.4) 75(14.0) 高血压 7.032 0.008 41.106 ＜0.001 94.509 ＜0.001 有 1 548(50.1) 785(46.1) 100(31.2) 146(27.3) 无 1 541(49.9) 919(53.9) 221(68.80) 389(72.7) 基线FPG(x±s, mmol/L) 8.70±2.14 8.90±2.12 3.140 0.002 8.53±2.40 -1.202 0.230 9.28±2.90 4.464 ＜0.001 注：a表格统计量为各用药组与不用药组比较所得；b部分个体变量信息缺失，因此相应变量合计不为3 089；c部分个体变量信息缺失，因此相应变量合计不为1 707。 
下载: 导出CSV
表 2 三种降糖处方倾向得分匹配后的基线特征[n(%)]
Table 2. Characteristics of study participants according to 3 antidiabetic prescriptions after PSM at baseline [n(%)]
变量 二甲双胍 格列吡嗪 联合用药 不用药(n=1 697) 用药(n=1 697) t/χ2值 P值 不用药(n=321) 用药(n=321) t/χ2值 P值 不用药(n=530) 用药(n=530) t/χ2值 P值 年龄(x±s, 岁) 64.79±11.96 64.61±11.44 0.431 0.666 61.09±11.98 61.88±9.91 -0.912 0.362 62.08±12.17 61.75±10.10 0.472 0.637 性别 0.218 0.641 0.107 0.744 3.734 0.053 男性 593(34.9) 607(35.8) 117(36.4) 122(38.0) 204(38.5) 236(44.5) 女性 1 104(65.1) 1 090(64.2) 204(63.6) 199(62.0) 326(61.5) 294(55.5) BMI(x±s, kg/m2) 25.32±3.44 25.26±3.11 0.544 0.587 25.20±3.50 24.81±2.88 1.548 0.122 25.02±3.34 25.02±2.89 0.034 0.973 吸烟 ＜0.001 >0.999 0.868 0.352 0.453 0.501 有 258(15.2) 257(15.1) 38(11.8) 47(14.6) 109(20.6) 119(22.5) 无 1 439(84.8) 1 440 (84.9) 283(88.2) 274(85.4) 421(79.4) 411(77.5) 饮酒 0.089 0.765 0.092 0.762 ＜0.001 >0.999 有 151(8.9) 157(9.3) 25(7.8) 22(6.9) 69(13.0) 70(13.2) 无 1 546(81.1) 1 540(81.7) 296(92.2) 299(93.1) 461(87.0) 460(86.8) 运动 0.001 0.972 0.923 0.337 0.325 0.568 有 938(55.3) 940(55.4) 127(39.6) 140(43.6) 195(36.8) 205(38.7) 无 759(44.7) 757(44.6) 194(60.4) 181(56.4) 335(63.2) 325(61.3) 职业 52.840 ＜0.001 44.362 ＜0.001 55.850 ＜0.001 农民 1 091(64.3) 1 286(75.8) 204(63.6) 278(86.6) 350(66.0) 455(85.8) 非农民 606(35.7) 411(24.2) 117(36.4) 43(13.4) 180(34.0) 75(14.2) 高血压 0.019 0.891 0.454 0.500 0.117 0.733 有 787(46.4) 782(46.1) 109(34.0) 100(31.2) 152(28.7) 146(27.5) 无 910(53.6) 915(53.9) 212(66.0) 221(68.8) 378(71.3) 384(72.5) 基线FPG(x±s, mmol/L) 8.82±2.20 8.90±2.12 -1.003 0.316 8.49±1.72 8.53±2.40 -0.242 0.809 9.25±2.75 9.19±2.74 0.366 0.715
下载: 导出CSV
表 3 二甲双胍、格列吡嗪及联合用药组净效应值排序前10%人群特征[n(%)]
Table 3. Characteristics of the top 10% population according to the incremental effect value of metformin, glipizide, and combination [n(%)]
特征 二甲双胍
(n=170)格列吡嗪
(n=33)联合用药
(n=53)年龄(x±s, 岁) 65.39±12.41 75.06±5.90 59.21±8.98 性别 男 68(40.0) 9(27.3) 34(64.2) 女 102(60.0) 24(72.7) 19(35.8) BMI(x±s, kg/m2) 22.15±1.65 24.25±2.92 24.82±3.42 基线FPG(x±s, mmol/L) 10.67±2.24 7.48±1.62 12.32±3.53 高血压 有 77(45.3) 3(9.1) 19(35.8) 无 93(54.7) 30(90.9) 34(64.2) 吸烟 有 30(17.6) 1(3.0) 5(9.4) 无 140(82.4) 32(97.0) 48(90.6) 饮酒 有 17(10.0) 1(3.0) 13(24.5) 无 153(90.0) 32(97.0) 40(75.5) 运动 有 84(49.4) 15(45.5) 11(20.8) 无 86(50.6) 18(54.5) 42(79.2)
下载: 导出CSV
-
[1] Chen L, Magliano DJ, Zimmet PZ. The worldwide epidemiology of type 2 diabetes mellitus: present and future perspectives[J]. Nat Rev Endocrinol, 2011, 8(4): 228-236. DOI: 10.1038/nrendo.2011.183. [2] NCD Risk Factor Collaboration (NCD-RisC). Worldwide trends in diabetes since 1980: a pooled analysis of 751 population-based studies with 4.4 million participants[J]. Lancet, 2016, 387(10027): 1513-1530. DOI: 10.1016/S0140-6736(16)00618-8. [3] Shaw JE, Sicree RA, Zimmet PZ. Global estimates of the prevalence of diabetes for 2010 and 2030[J]. Diabetes Res Clin Pract, 2010, 87(1): 4-14. DOI: 10.1016/j.diabres.2009.10.007. [4] Maffi P, Secchi A. The burden of diabetes: emerging data[J]. Dev Ophthalmol, 2017, 60: 1-5. DOI: 10.1159/000459641. [5] Navarro-Pérez J, Orozco-Beltran D, Gil-Guillen V, et al. Mortality and cardiovascular disease burden of uncontrolled diabetes in a registry-based cohort: the ESCARVAL-risk study[J]. BMC Cardiovasc Disord, 2018, 18(1): 180. DOI: 10.1186/s12872-018-0914-1. [6] 中华医学会糖尿病学分会. 中国2型糖尿病防治指南(2017年版)[J]. 中华糖尿病杂志, 2018, 10(1): 4-67. DOI: 10.3760/cma.j.issn.1674-5809.2018.01.003.Diabetes Society of Chinese Medical Association. Guidelines for the prevention and treatment of type 2 diabetes in China (2017 edition)[J]. Chin J Diabetes Mellitus, 2018, 10(1): 4-67. DOI: 10.3760/cma.j.issn.1674-5809.2018.01.003. [7] Ji LN, Lu JM, Guo XH, et al. Glycemic control among patients in China with type 2 diabetes mellitus receiving oral drugs or injectables[J]. BMC Public Health, 2013, 13: 602. DOI: 10.1186/1471-2458-13-602. [8] Sherifali D, Nerenberg K, Pullenayegum E, et al. The effect of oral antidiabetic agents on A1C levels: a systematic review and meta-analysis[J]. Diabetes Care, 2010, 33(8): 1859-1864. DOI: 10.2337/dc09-1727. [9] American Diabetes Association. Diabetes advocacy: standards of medical care in diabetes-2018[J]. Diabetes Care, 2018, 41(Suppl 1): S152-S153. DOI: 10.2337/dc18-S015. [10] Blonde L, Khunti K, Harris SB, et al. Interpretation and impact of real-world clinical data for the practicing clinician[J]. Adv Ther, 2018, 35(11): 1763-1774. DOI: 10.1007/s12325-018-0805-y. [11] Barnish MS, Turner S. The value of pragmatic and observational studies in health care and public health[J]. Pragmat Obs Res, 2017, 8: 49-55. DOI: 10.2147/POR.S137701. [12] 宋捷, 林海, 金春林, 等. 国内外糖尿病用药结构比较与分析[J]. 中国药业, 2020, 29(22): 7-10. DOI: 10.3969/j.issn.1006-4931.2020.22.002.Song J, Lin H, Jin CL, et al. Comparison and analysis of the structure of antidiabetic drugs usage at home and abroad[J]. China Pharm, 2020, 29(22): 7-10. DOI: 10.3969/j.issn.1006-4931.2020.22.002. [13] Guelman L, Guillén M, Pérez-Marín AM. Uplift random forests[J]. Cybern Syst, 2015, 46(3-4): 230-248. DOI: 10.1080/01969722.2015.1012892. [14] Austin PC. An introduction to propensity score methods for reducing the effects of confounding in observational studies[J]. Multivariate Behav Res, 2011, 46(3): 399-424. DOI: 10.1080/00273171.2011.568786. [15] Drake C, Fisher L. Prognostic models and the propensity score[J]. Int J Epidemiol, 1995, 24(1): 183-187. DOI: 10.1093/ije/24.1.183. [16] Guelman L, Guillén M, Pérez-Marín AM. Random forests for uplift modeling: an insurance customer retention case[J]. LNBIP, 2012, 115: 123-133. DOI: 10.1007/978-3-642-30433-0_13. [17] Rzepakowski P, Jaroszewicz S. Decision trees for uplift modeling with single and multiple treatments[J]. Knowl Inf Syst, 2012, 32(2): 303-327. DOI: 10.1007/s10115-011-0434-0. [18] Edwards S. Thomas M. Cover Joy A. Thomas elements of information theory 2nd ed. 2006 John Wiley & Sons, Inc[J]. Inf Process Manag, 2008, 44(1): 400-401. DOI: 10.1016/j.ipm.2007.02.009. [19] Rzepakowski P, Jaroszewicz S. Decision trees for uplift modeling with single and multiple treatments[J]. Knowl Inf Syst, 2012, 32(2): 303-327. DOI: 10.1007/s10115-011-0434-0. [20] 王国强, 刘云霞. 不同药物联合治疗2型糖尿病的效果对比[J]. 中国卫生标准管理, 2021, 12(4): 107-109. DOI: 10.3969/j.issn.1674-9316.2021.04.040.Wang GQ, Liu YX. Comparison of the effect of different drugs in the treatment of type 2 diabetes mellitus[J]. China Heal Stand Manag, 2021, 12(4): 107-109. DOI: 10.3969/j.issn.1674-9316.2021.04.040. [21] 高蕾莉, 纪立农, 陆菊明, 等. 2009~2012年我国2型糖尿病患者药物治疗与血糖控制状况调查[J]. 中国糖尿病杂志, 2014, 22(7): 594-598. DOI: 10.3969/j.issn.1006-6187.2014.07.005.Gao LL, Ji LN, Lu JM, et al. Current status of blood glucose control and treatment of type 2 diabetes in China 2009-2012[J]. Chin J Diabetes, 2014, 22(7): 594-598. DOI: 10.3969/j.issn.1006-6187.2014.07.005. [22] Holland PW, Rubin DB. Causal inference in retrospective studies[J]. ETS Res Rep Ser, 1987(1): 203-231. DOI: 10.1002/j.2330-8516.1987.tb00211.x. [23] Anyanwagu U, Mamza J, Gordon J, et al. Premixed vs basal-bolus insulin regimen in type 2 diabetes: comparison of clinical outcomes from randomized controlled trials and real-world data[J]. Diabet Med, 2017, 34(12): 1728-1736. DOI: 10.1111/dme.13518. [24] Rosenbaum PR, Rubin DB. The central role of the propensity score in observational studies for causal effects[J]. Biometrika, 1983, 70(1): 41-55. DOI: 10.1093/biomet/70.1.41. [25] Guelman L, Guillén M, Pérez-Marín AM. A decision support framework to implement optimal personalized marketing interventions[J]. Decis Support Syst, 2015, 72: 24-32. DOI: 10.1016/j.dss.2015.01.010. -
点击查看大图
图(2) / 表(3)
计量
- 文章访问数: 603
- HTML全文浏览量: 328
- PDF下载量: 66
- 被引次数: 0