Department of Biostatistics and Translational Medicine, Medical University of Lodz, Lodz, Poland.
Department of Pediatrics, Oncology, Hematology and Diabetology, Medical University of Lodz, Lodz, Poland.
Pediatr Diabetes. 2018 Mar;19(2):231-235. doi: 10.1111/pedi.12551. Epub 2017 Jun 19.
Direct measurement of insulin sensitivity in children with type 1 diabetes is cumbersome and time consuming.
The aim of our study was to develop novel, accurate machine learning-based methods of insulin resistance estimation in children with type 1 diabetes.
A hyperinsulinemic hyperglycemic clamp study was performed to evaluate the glucose disposal rate (GDR) in a study group consisting of 315 patients aged 7.6 to 19.7 years. The group was randomly divided into a training and independent testing set for model performance assessment. GDR was estimated on the basis of simple clinical variables using 2 non-linear methods: artificial neural networks (ANN) and multivariate adaptive regression splines (MARSplines). The results were compared against the most frequently used predictive model, based on waist circumference, triglyceride (TG), and HbA1c levels.
The reference model showed moderate performance ( R = 0.26) with a median absolute percentage error of 49.1%, and with the worst fit observed in young (7-12 years) children ( R = 0.17). Predictions of the MARSplines model were significantly more accurate than those of the reference model (median error 3.6%, R = 0.44 P < .0001). The predictions of the ANN, however, showed significantly lower error than those of the reference model (P < .0001) and MARSplines (P < .0001) and better fit regardless of patient age. ANN-estimated GDRs were within a ±20% error range in 75% of cases with a median error of 0.6% and an R = 0.66. The predictive tool is available at http://link.konsta.com.pl/gdr.
The developed GDR estimation model reliant on ANN allows for an optimized prediction of GDR for research and clinical purposes.
直接测量 1 型糖尿病儿童的胰岛素敏感性既繁琐又耗时。
我们的研究旨在开发新的、基于准确机器学习的 1 型糖尿病儿童胰岛素抵抗估计方法。
对 315 名年龄为 7.6 至 19.7 岁的患者进行高胰岛素-高血糖钳夹试验，以评估葡萄糖处置率（GDR）。该组随机分为训练集和独立测试集，用于评估模型性能。基于简单的临床变量，使用两种非线性方法：人工神经网络（ANN）和多元自适应回归样条（MARSplines）来估计 GDR。结果与最常用的预测模型进行了比较，该模型基于腰围、三酰甘油（TG）和糖化血红蛋白（HbA1c）水平。
参考模型的性能中等（R=0.26），中位数绝对百分比误差为 49.1%，在年龄较小（7-12 岁）的儿童中拟合效果最差（R=0.17）。MARSplines 模型的预测明显比参考模型更准确（中位数误差 3.6%，R=0.44，P<.0001）。然而，ANN 的预测误差明显低于参考模型（P<.0001）和 MARSplines（P<.0001），并且无论患者年龄如何，拟合效果都更好。ANN 估计的 GDR 在 75%的情况下误差在±20%范围内，中位数误差为 0.6%，R=0.66。预测工具可在 http://link.konsta.com.pl/gdr 上获取。
基于 ANN 的开发的 GDR 估计模型能够优化研究和临床目的的 GDR 预测。
