文献速递流行病学与生物统计组学、生物样本库与精准医学

在五个超过22000人的独立芬兰队列中开发2型糖尿病风险预测模型并评估循环代谢生物标志物的作用

Developing risk prediction models for type 2 diabetes and assessing the role of circulating metabolic biomarkers in five independent Finnish cohorts with over 22,000 individuals.
PubMed 文献 · J Clin Epidemiol · 2025-09-18

研究利用五个芬兰队列共493例病例和21755名对照,开发并验证2型糖尿病风险预测模型。先以15个临床变量构建基线模型,再评估154个代谢变量通过多种统计方法带来的增量改善。基线模型验证队列AUC为0.832至0.913,加入代谢变量后最高AUC为0.840至0.915,改善小且无临床相关性,多数模型校准不良。

要点

  • 基线临床变量模型在验证队列中AUC达0.832至0.913,判别力很好。
  • 加入154个代谢变量后最高AUC为0.840至0.915,改善小且无临床相关性。
  • 多数构建的模型在验证队列中校准不良,判别力泛化较好但校准泛化较差。

核心结论

  • 含临床变量的基线模型在预测2型糖尿病方面判别力很好,代谢变量模型未能有意义地改善其性能,因基线设定较高留有的改善空间很小。
  • 构建的模型在判别力方面对新的独立数据泛化较好,但在校准方面泛化相对较差。

研究设计与样本

  • 使用五个队列共493例病例和21755名对照开发并验证2型糖尿病风险预测模型。
  • 先以15个临床变量构建基线模型,再使用多种不同统计方法评估154个代谢变量对预测的增量改善。
  • 在多个验证队列中以判别力和校准评估模型性能。

主要发现

  • 基线模型在验证队列中受试者工作特征曲线下面积值为0.832至0.913。
  • 加入代谢指标后,验证队列中达到的最高受试者工作特征曲线下面积值为0.840至0.915,较基线模型改善小且无临床相关性。
  • 多数构建的模型在验证队列中校准不良。

局限性

  • 加入代谢变量后的改善小且无临床相关性。
  • 多数构建的模型在验证队列中校准不良。
  • 构建的模型在校准方面对新的独立数据泛化相对较差。

以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。

原文

OBJECTIVES: Type 2 diabetes (T2D) is a serious public health risk. It is not well known whether large scale omics data could help improve the risk prediction. We combined the use of metabolic variables with modern model selection methods to develop risk prediction models for T2D. We then assessed the incremental improvement in model performance produced by the metabolic variables, when compared to a clinical baseline. METHODS: We used five cohorts with a total of 493 cases and 21,755 controls to develop and validate risk prediction models for T2D. We first constructed a baseline model using 15 clinical variables, and then assessed the incremental improvement provided by a set of 154 metabolic variables to the prediction, using multiple different statistical methods. Model performance was assessed in terms of discrimination and calibration in multiple validation cohorts. RESULTS: The baseline model reached area under the receiver operating characteristic curve values between 0.832 and 0.913 in the validation cohorts. On adding the metabolic measures, the highest area under the receiver operating characteristic curve values reached in the validation cohorts were between 0.840 and 0.915, a small but not clinically relevant improvement over the baseline model. Most of the constructed models were miscalibrated in the validation cohorts. CONCLUSION: The baseline model with clinical variables achieved a very good discrimination in predicting T2D. Models with metabolic variables were not able to meaningfully improve on it, as the high level set by the baseline left little room for improvement. The constructed models generalize to new independent data quite well in terms of discrimination, but relatively poorly in terms of calibration.

来源:J Clin Epidemiol,查看原文。版权归原作者所有。

引用

  1. Kiviniemi E, Karhunen V, Perola M, et al. Developing risk prediction models for type 2 diabetes and assessing the role of circulating metabolic biomarkers in five independent Finnish cohorts with over 22,000 individuals. J Clin Epidemiol. 2025. PMID: 40975243. doi:10.1016/j.jclinepi.2025.111978

供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。