Mean value of CKD and non-CKD patients.

Chronic kidney disease (CKD) affects over 13% of the population, totaling more than 800 million individuals worldwide. Timely identification and intervention are crucial to delay CKD progression and improve patient outcomes. This research focuses on developing a predictive model to classify diabetic patients showing signs of kidney function impairment based on their CKD development risk. Our model utilizes electronic medical record (EMR) data, specifically by incorporating patient demographics, laboratory results, chronic conditions, risk factors, and medication codes to predict the onset of CKD in diabetic patients six months in advance, achieving an average Area Under the Curve (AUC) of 0.88. We leverage aggregated EMR data to effectively capture relevant information within the observation year instead of using temporal EMR data. Furthermore, we identify the most significant features for predicting CKD onset, including mean, minimum, and first quartile of estimated glomerular filtration rate (eGFR) during the observation year, along with variables such as diagnosis age and duration of hypertension, osteoarthritis, and diabetes, as well as levels of hemoglobin and fasting blood glucose (FBG). We also explored a refined model utilizing only these most significant features, which yields a slightly lower AUC of 0.86. These variables are typically available in primary data, empowering physicians for real-time risk assessment. The proposed model’s ability to identify higher-risk patients is essential for timely intervention, personalized care, risk stratification, patient education, and potential cost savings. This research contributes valuable insights for healthcare practitioners seeking efficient tools for early CKD detection in diabetic populations.

慢性肾脏病（Chronic kidney disease, CKD）影响全球超过13%的人口，总患病人数逾8亿。及时识别与干预对于延缓慢性肾脏病进展、改善患者预后至关重要。本研究旨在开发一款预测模型，基于糖尿病患者的慢性肾脏病发病风险，对出现肾功能损伤迹象的糖尿病患者进行分类。本模型采用电子病历（electronic medical record, EMR）数据，通过纳入患者人口统计学信息、实验室检验结果、慢性病史、危险因素及药物编码，可提前6个月预测糖尿病患者的慢性肾脏病发病风险，平均曲线下面积（Area Under the Curve, AUC）达0.88。我们采用观测年内的聚合电子病历数据，而非时序电子病历数据，以有效捕获相关关联信息。此外，我们识别出预测慢性肾脏病发病的核心特征，包括观测年内估算肾小球滤过率（estimated glomerular filtration rate, eGFR）的均值、最小值与第一四分位数，以及确诊年龄、高血压、骨关节炎与糖尿病的病程，还有血红蛋白水平与空腹血糖（fasting blood glucose, FBG）等变量。我们还探索了仅使用上述核心特征的精简模型，其曲线下面积略降至0.86。上述变量通常可从原始数据中获取，能够辅助医生开展实时风险评估。本模型识别高危患者的能力，对于及时干预、个性化诊疗、风险分层、患者教育及潜在的医疗成本节约均具有重要意义。本研究可为寻求高效工具以早期筛查糖尿病群体慢性肾脏病的医疗从业者提供极具价值的参考依据。