基于机器学习的糖尿病肾脏病患者疲乏状态诊断预测模型构建

doi:10.3877/cma.j.issn.2095-3216.2026.03.004

目的

构建基于机器学习的糖尿病肾脏病(diabetic kidney disease，DKD)患者疲乏状态诊断预测模型，为早期筛查及干预提供依据。

方法

回顾性分析2023年1月至2025年6月于安徽中医药大学第七附属医院收治的DKD患者，按照疲乏严重程度量表评分分值为疲乏组(评分≥4分)与非疲乏组(评分<4分)，收集患者人口学、临床与实验室指标及用药情况。通过LASSO回归筛选稳健特征，采用重复分层嵌套交叉验证，比较LASSO Logistic回归、随机森林及轻量梯度提升机(light gradient boosting machine, LightGBM)模型的预测性能，并采用决策曲线分析评估模型的临床应用价值。

结果

共纳入DKD患者166例，其中疲乏组78例、非疲乏组88例。DKD患者中疲乏发生率为46.99%(78/166例)。通过LASSO回归筛选得到11项稳健特征，分别为估算的小球滤过率、血红蛋白、白蛋白、尿白蛋白/肌酐比值、年龄、二氧化碳结合力、糖化血红蛋白、利尿剂使用、睡眠障碍、心力衰竭及钠-葡萄糖协同转运蛋白2抑制剂使用。内部验证结果显示，LightGBM模型的预测性能最优，其受试者操作曲线下面积为0.836，Brier评分为0.175，灵敏度为0.756，特异度为0.795；校准曲线表明该模型校准效果良好，决策曲线分析显示该模型在阈值0.05~0.60范围内可获得良好的临床净获益。

结论

基于LightGBM构建的诊断预测模型可有效识别DKD患者疲乏状态，有助于临床早期风险分层及干预决策制定。

关键词: 糖尿病肾脏病, 疲乏, 风险识别, 机器学习, 诊断预测模型

Objective

To construct a diagnostic and prediction model based on machine learning for fatigue status in patients with diabetic kidney disease (DKD), so as to provide a basis for early screening and intervention.

Methods

A retrospective analysis was conducted on patients with DKD admitted to the Seventh Affiliated Hospital of Anhui University of Chinese Medicine from January 2023 to June 2025. The patients were divided into a fatigue group (score ≥ 4 points) and a non-fatigue group (score < 4 points) according to the Fatigue Severity Scale (FSS). Demographic, clinical, laboratory, and medication data of the patients were collected. Robust features were screened via LASSO regression. Repeated stratified nested cross-validation was adopted. The predictive performances of LASSO logistic regression, Random Forest (RF), and Light Gradient Boosting Machine (LightGBM) models were compared. And the clinical application value of the models was evaluated using decision curve analysis.

Results

A total of 166 patients with DKD were enrolled, including 78 cases in the fatigue group and 88 cases in the non-fatigue group. The incidence of fatigue among the DKD patients was 46.99% (78/166). Eleven robust features were screened out via LASSO regression, namely estimated glomerular filtration rate, hemoglobin, albumin, urinary albumin/creatinine ratio, age, carbon dioxide combining power, glycated hemoglobin, diuretic use, sleep disorder, heart failure, and sodium-glucose cotransporter 2 inhibitor use. Internal validation results showed that the LightGBM model achieves the best predictive performance, with an area under the receiver operating characteristic curve of 0.836, a Brier score of 0.175, a sensitivity of 0.756, and a specificity of 0.795. The calibration curve indicated that the model had a good calibration effect. And the decision curve analysis revealed that this model yielded good clinical net benefit within the threshold range of 0.05-0.60.

Conclusion

The diagnostic prediction model built based on LightGBM effectively identified the fatigue status in the DKD patients, facilitating early clinical risk stratification and intervention decision-making.

Key words: Diabetic kidney disease, Fatigue, Risk identification, Machine learning, Diagnostic prediction model

表1 两组糖尿病肾脏病患者临床特征比较

项目	疲乏组(78例)	非疲乏组(88例)	t/χ²值	P值
年龄(岁)	56.38±8.76	51.70±9.79	3.229^a	0.002
性别(男/女，例)	45/33	42/46	1.646^b	0.199
糖尿病病程(年)	11.93±6.71	10.02±5.44	1.997^a	0.048
DKD病程(年)	4.37±2.45	4.24±2.48	0.357^a	0.722
合并症[例(%)]
高血压	60(76.92)	58(65.91)	1.934^b	0.164
冠心病	19(24.36)	15(17.05)	0.946^b	0.331
心力衰竭	16(20.51)	7(7.95)	4.462^b	0.035
脑卒中史	9(11.54)	9(10.23)	<0.001^b	0.983
糖尿病周围神经病变	27(34.62)	23(26.14)	1.242^b	0.265
糖尿病视网膜病变	35(44.87)	29(32.95)	2.507^b	0.113
睡眠障碍	26(33.33)	15(17.05)	5.602^b	0.018
KDIGO eGFR分期[例(%)]			17.110^b	0.004
G1	1(1.28)	2(2.27)
G2	12(15.38)	35(39.77)
G3a	22(28.21)	26(29.55)
G3b	21(26.92)	13(14.77)
G4	18(23.08)	8(9.09)
G5	4(5.13)	4(4.55)
治疗用药[例(%)]
肾素-血管紧张素系统阻断剂	49(62.82)	60(68.18)	0.528^b	0.468
钠-葡萄糖协同转运蛋白2抑制剂	24(30.77)	46(52.27)	6.983^b	0.008
胰岛素	54(69.23)	54(61.36)	0.759^b	0.384
利尿剂	31(39.74)	17(19.32)	7.428^b	0.006
促红细胞生成素	23(29.49)	14(15.91)	3.658^b	0.056
铁剂	17(21.79)	12(13.64)	1.414^b	0.234

表1 两组糖尿病肾脏病患者临床特征比较

项目	疲乏组(78例)	非疲乏组(88例)	t/χ²值	P值
年龄(岁)	56.38±8.76	51.70±9.79	3.229^a	0.002
性别(男/女，例)	45/33	42/46	1.646^b	0.199
糖尿病病程(年)	11.93±6.71	10.02±5.44	1.997^a	0.048
DKD病程(年)	4.37±2.45	4.24±2.48	0.357^a	0.722
合并症[例(%)]
高血压	60(76.92)	58(65.91)	1.934^b	0.164
冠心病	19(24.36)	15(17.05)	0.946^b	0.331
心力衰竭	16(20.51)	7(7.95)	4.462^b	0.035
脑卒中史	9(11.54)	9(10.23)	<0.001^b	0.983
糖尿病周围神经病变	27(34.62)	23(26.14)	1.242^b	0.265
糖尿病视网膜病变	35(44.87)	29(32.95)	2.507^b	0.113
睡眠障碍	26(33.33)	15(17.05)	5.602^b	0.018
KDIGO eGFR分期[例(%)]			17.110^b	0.004
G1	1(1.28)	2(2.27)
G2	12(15.38)	35(39.77)
G3a	22(28.21)	26(29.55)
G3b	21(26.92)	13(14.77)
G4	18(23.08)	8(9.09)
G5	4(5.13)	4(4.55)
治疗用药[例(%)]
肾素-血管紧张素系统阻断剂	49(62.82)	60(68.18)	0.528^b	0.468
钠-葡萄糖协同转运蛋白2抑制剂	24(30.77)	46(52.27)	6.983^b	0.008
胰岛素	54(69.23)	54(61.36)	0.759^b	0.384
利尿剂	31(39.74)	17(19.32)	7.428^b	0.006
促红细胞生成素	23(29.49)	14(15.91)	3.658^b	0.056
铁剂	17(21.79)	12(13.64)	1.414^b	0.234

表2 两组糖尿病肾脏病患者实验室指标比较

项目	疲乏组(78例)	非疲乏组(88例)	t/χ²值	P值
血红蛋白(g/L)	103.76±15.02	114.22±14.67	－4.989^a	<0.001
白细胞计数(×10⁹/L)	7.42±2.13	7.05±1.91	1.180^a	0.240
中性粒细胞比例(%)	69.03±10.28	67.26±8.26	1.216^a	0.226
白蛋白(g/L)	34.57±4.27	38.09±4.31	－5.516^a	<0.001
血钠(mmol/L)	137.27±3.51	138.02±3.20	－1.447^a	0.150
血钾(mmol/L)	4.74±0.59	4.51±0.49	2.453^a	0.016
血钙(mmol/L)	2.14±0.12	2.18±0.11	－1.707^a	0.090
血磷(mmol/L)	1.36±0.29	1.25±0.25	2.709^a	0.008
二氧化碳结合力(mmol/L)	20.46±3.77	22.31±3.59	－2.691^a	0.008
糖化血红蛋白(%)	8.52±1.57	7.95±1.45	3.900^a	<0.001
总胆固醇(mmol/L)	4.48±1.02	4.36±0.98	0.804^a	0.423
甘油三酯(mmol/L)	1.72(1.20，2.56)	1.65(1.15，2.40)	0.540^b	0.590
低密度脂蛋白胆固醇(mmol/L)	2.55±0.82	2.49±0.79	0.454^a	0.650
高密度脂蛋白胆固醇(mmol/L)	1.05±0.26	1.12±0.27	－2.271^a	0.024
血清肌酐(μmol/L)	176.41(122.82，260.78)	129.83(93.19，201.35)	3.358^b	0.001
血尿素氮(mmol/L)	10.84±3.99	9.19±3.66	4.231^a	<0.001
eGFR [ml/(min·1.73 m²)]	41.56±17.82	54.18±19.76	－4.063^a	<0.001
UACR(mg/g)	545.37(239.31，955.83)	248.69(130.29，637.52)	2.970^b	0.003

表2 两组糖尿病肾脏病患者实验室指标比较

项目	疲乏组(78例)	非疲乏组(88例)	t/χ²值	P值
血红蛋白(g/L)	103.76±15.02	114.22±14.67	－4.989^a	<0.001
白细胞计数(×10⁹/L)	7.42±2.13	7.05±1.91	1.180^a	0.240
中性粒细胞比例(%)	69.03±10.28	67.26±8.26	1.216^a	0.226
白蛋白(g/L)	34.57±4.27	38.09±4.31	－5.516^a	<0.001
血钠(mmol/L)	137.27±3.51	138.02±3.20	－1.447^a	0.150
血钾(mmol/L)	4.74±0.59	4.51±0.49	2.453^a	0.016
血钙(mmol/L)	2.14±0.12	2.18±0.11	－1.707^a	0.090
血磷(mmol/L)	1.36±0.29	1.25±0.25	2.709^a	0.008
二氧化碳结合力(mmol/L)	20.46±3.77	22.31±3.59	－2.691^a	0.008
糖化血红蛋白(%)	8.52±1.57	7.95±1.45	3.900^a	<0.001
总胆固醇(mmol/L)	4.48±1.02	4.36±0.98	0.804^a	0.423
甘油三酯(mmol/L)	1.72(1.20，2.56)	1.65(1.15，2.40)	0.540^b	0.590
低密度脂蛋白胆固醇(mmol/L)	2.55±0.82	2.49±0.79	0.454^a	0.650
高密度脂蛋白胆固醇(mmol/L)	1.05±0.26	1.12±0.27	－2.271^a	0.024
血清肌酐(μmol/L)	176.41(122.82，260.78)	129.83(93.19，201.35)	3.358^b	0.001
血尿素氮(mmol/L)	10.84±3.99	9.19±3.66	4.231^a	<0.001
eGFR [ml/(min·1.73 m²)]	41.56±17.82	54.18±19.76	－4.063^a	<0.001
UACR(mg/g)	545.37(239.31，955.83)	248.69(130.29，637.52)	2.970^b	0.003

图1 LASSO Logistic回归的系数路径图(A)和交叉验证曲线(B)

图2 3种模型受试者工作特征曲线比较注：LASSO Logistic：least absolute shrinkage and selection operator logistic，最小绝对收缩和选择算子逻辑回归模型；RF：Random Forest,随机森林模型；LightGBM：Light Gradient Boosting Machine，轻量梯度提升机模型

表3 不同模型的内部验证性能比较

模型	曲线下面积	Brier评分	敏感度	特异度
最小绝对收缩和选择算子逻辑回归模型	0.737	0.208	0.667	0.727
随机森林模型	0.790	0.191	0.679	0.795
轻量梯度提升机模型	0.836	0.175	0.756	0.795

表3 不同模型的内部验证性能比较

模型	曲线下面积	Brier评分	敏感度	特异度
最小绝对收缩和选择算子逻辑回归模型	0.737	0.208	0.667	0.727
随机森林模型	0.790	0.191	0.679	0.795
轻量梯度提升机模型	0.836	0.175	0.756	0.795

图3 最终轻量梯度提升机模型平滑校准曲线

图4 不同模型的决策曲线分析比较注：LASSO Logistic：least absolute shrinkage and selection operator logistic，最小绝对收缩和选择算子逻辑回归模型；RF：Random Forest,随机森林模型；LightGBM：Light Gradient Boosting Machine，轻量梯度提升机模型

[1]	van Raalte DH, Bjornstad P, Cherney DZI, et al. Combination therapy for kidney disease in people with diabetes mellitus [J]. Nat Rev Nephrol, 2024, 20(7): 433-446.
[2]	Zhang H, Wang K, Zhao H, et al. Diabetic kidney disease: from pathogenesis to multimodal therapy-current evidence and future directions [J]. Front Med (Lausanne), 2025, 12: 1631053.
[3]	Yi TW, Sridhar VS, Scott J, et al. Next-generation therapeutics for diabetic kidney disease [J]. Nat Rev Nephrol, 2026, 22(3): 145-162.
[4]	Morales J, Handelsman Y. Cardiovascular outcomes in patients with diabetes and kidney disease: JACC review topic of the week [J]. J Am Coll Cardiol, 2023, 82(2): 161-170.
[5]	刘艳琴，芦园月，李旺鑫，等. 基于不平衡数据的非肾病水平蛋白尿的膜性肾病预后模型建立[J]. 中华医学杂志，2023, 103(18): 1386-1392.
[6]	胡嘉桐，刘明晴，李红旗，等. 基于机器学习的2型糖尿病与骨代谢的关联性研究[J]. 中国科学技术大学学报，2023, 53(12): 1205.
[7]	Wallisch C, Dunkler D, Heinze G. Selection of variables for multivariable models: opportunities and limitations in quantifying model stability by resampling [J]. Stat Med, 2020, 40(2): 369-381.
[8]	Parvandeh S, Yeh HW, Paulus MP, et al. Consensus features nested cross-validation [J]. Bioinformatics, 2020, 36(10): 3093-3098.
[9]	Hughes A, Ju A, Cazzolli R, et al. Patient-reported outcome measures for fatigue in patients with chronic kidney disease: a systematic review [J]. BMJ Open, 2025, 15(7): e099592.
[10]	Schade van Westrum E, Hoogeveen EK, Broekman BFP, et al. Fatigue across different chronic kidney disease populations: experiences and needs of patients [J]. Clin Kidney J, 2025, 18(5): sfaf118.
[11]	Zhang Y, Guo J, Yang N, et al. Machine learning to predict postdialysis fatigue in patients undergoing hemodialysis [J]. Ren Fail, 2025, 47(1): 2529452.
[12]	Nakoui N, Ilbeigi S, Ahmadi MM, et al. Comparison of the effect of aerobic and resistance training on fatigue, quality of life and biochemical factors in hemodialysis patients [J]. Sci Rep, 2025, 15(1): 10052.
[13]	石洁，张瑶，林佳睿，等. 基于列线图的肥胖2型糖尿病患者糖尿病肾病风险评估模型构建与验证[J]. 新医学，2025, 56(10): 986-994.
[14]	Fu S, Huang J, Feng Z, et al. Inflammatory indexes and anemia in chronic kidney disease: correlation and survival analysis of the National Health and Nutrition Examination Survey 2005-2018 [J]. Ren Fail, 2024, 46(2): 2399314.
[15]	Massini G, Caldiroli L, Molinari P, et al. Nutritional strategies to prevent muscle loss and sarcopenia in chronic kidney disease: what do we currently know? [J]. J Ren Nutr, 2024, 34(3): 387-396.
[16]	Visser WJ, van de Braak EEM, de Mik-van Egmond AME, et al. Effects of correcting metabolic acidosis on muscle mass and functionality in chronic kidney disease: a systematic review and meta-analysis [J]. J Cachexia Sarcopenia Muscle, 2023, 14(6): 2498-2508.
[17]	Navab F, Rouhani MH, Moeinzadeh F, et al. The effects of oral sodium bicarbonate supplementation on anthropometric measures in patients with chronic kidney disease: a systematic review and meta-analysis of randomized clinical trials [J]. Food Sci Nutr, 2023, 11(11): 6749-6760.
[18]	Gajewska A, Wasiak J, Sapeda N, et al. SGLT2 inhibitors in kidney diseases-a narrative review [J]. Int J Mol Sci, 2024, 25(9): 4959.
[19]	Preda A, Montecucco F, Carbone F, et al. SGLT2 inhibitors: from glucose-lowering to cardiovascular benefits [J]. Cardiovasc Res, 2024, 120(5): 443-460.
[20]	Girardi ACC, Polidoro JZ, Castro PC, et al. Mechanisms of heart failure and chronic kidney disease protection by SGLT2 inhibitors in nondiabetic conditions [J]. Am J Physiol Cell Physiol, 2024, 327(3): C525-C544.

[1]	梅森, 江涛. 人工智能机器学习模型预测关节置换术后效果的进展[J/OL]. 中华关节外科杂志(电子版), 2025, 19(06): 735-741.
[2]	吕思怡, 王琰琪, 仇珺, 陈宇江, 高洁. 人工智能图像处理技术在口腔医学中的应用[J/OL]. 中华口腔医学研究杂志(电子版), 2026, 20(01): 40-46.
[3]	徐宏博, 胡玉良, 魏雪栋, 金李晨, 武克风, 陈逸伦, 陆兵, 周守军, 侯建全. 基于临床和CT影像组学特征的机器学习模型对经皮肾镜术后脓毒症的预测价值[J/OL]. 中华腔镜泌尿外科杂志(电子版), 2026, 20(03): 297-306.
[4]	黄楚曦, 吴卓. 人工智能在膀胱癌影像中的应用进展[J/OL]. 中华腔镜泌尿外科杂志(电子版), 2026, 20(03): 241-247.
[5]	刘紫芋, 陈方梅, 高盼, 宋溢, 田宇燕, 魏静. 肺癌患者同步放化疗癌因性疲乏潜在剖面分析及影响因素研究[J/OL]. 中华肺部疾病杂志(电子版), 2026, 19(01): 49-55.
[6]	宋佳耕, 袁文翰, 郑莹. 人工智能在妇科微创手术中的应用与展望[J/OL]. 中华腔镜外科杂志(电子版), 2026, 19(01): 60-64.
[7]	王小振, 陈灿辉, 唐善华, 代浩嘉, 丰扬舸, 王恺, 李清平, 李川江. 基于不同机器学习技术构建肝移植术后早期严重并发症预测模型和效能比较[J/OL]. 中华肝脏外科手术学电子杂志, 2026, 15(02): 197-204.
[8]	邢颖, 王峰. 基于机器学习构建肝切除术后肝衰竭预测列线图模型及其预测价值[J/OL]. 中华肝脏外科手术学电子杂志, 2026, 15(02): 190-196.
[9]	戴宗伯, 张城硕, 郭庭维, 何知远, 赵昊宇, 张宇慈, 张佳林. 基于MRI影像组学机器学习构建肝细胞癌微血管侵犯预测模型[J/OL]. 中华肝脏外科手术学电子杂志, 2026, 15(01): 36-44.
[10]	邓美, 代德琴, 颜晓勇, 苏朝江. 姜黄素对糖尿病肾脏病大鼠肾组织坏死性凋亡的影响[J/OL]. 中华肾病研究电子杂志, 2026, 15(02): 85-92.
[11]	王珍. 基于"肠-肾"轴探讨"内热致癥"病机在糖尿病肾病中的应用[J/OL]. 中华肾病研究电子杂志, 2026, 15(01): 60-60.
[12]	刘娜, 盛华, 高超. 机械通气与急性肾损伤：从机制到临床实践[J/OL]. 中华肾病研究电子杂志, 2025, 14(06): 345-350.
[13]	荣锦, 骆明星, 王禹, 刘婷婷, 张宏斌. 全膝关节置换术后慢性疼痛影响因素的多种模型预测性能比较[J/OL]. 中华老年骨科与康复电子杂志, 2025, 11(06): 337-344.
[14]	储丹丹, 何培, 万浩宇, 虎子单, 闻玉婷, 王佳, 匡野, 金腾川, 冯磊. 基于循证医学和机器学习的自身免疫性疾病检测项目组合新策略构建[J/OL]. 中华临床实验室管理电子杂志, 2026, 14(02): 166-173.
[15]	梁怡凡, 牟婧宇, 吴雅婷, 邳靖陶, 陈乐, 武剑. 不同人工智能模型预测脑卒中不良预后诊断效能的荟萃分析[J/OL]. 中华脑血管病杂志(电子版), 2026, 20(03): 308-319.

阅读次数

全文

摘要

选择文件类型/文献管理软件名称

选择包含的内容

Construction of a diagnostic and prediction model based on machine learning for fatigue status in patients with diabetic kidney disease

模态框（Modal）标题

选择文件类型/文献管理软件名称

选择包含的内容

Construction of a diagnostic and prediction model based on machine learning for fatigue status in patients with diabetic kidney disease