Guzmán-Bárcenas José, Hernández José Alfredo, Arias-Martínez Joel, Baptista-González Héctor, Ceballos-Reyes Guillermo, Irles Claudine
Department of Physiology and Cellular Development, Instituto Nacional de Perinatología Isidro Espinoza de los Reyes (INPerIER), Montes Urales 800, Lomas de Virreyes, Mexico city, C.P. 11000, Mexico.
Centro de Investigación en Ingeniería y Ciencias Aplicadas (CIICAp)-Universidad Autónoma del Estado de Morelos (UAEM), Cuernavaca, Morelos, Mexico.
BMC Pregnancy Childbirth. 2016 Jul 21;16(1):179. doi: 10.1186/s12884-016-0967-z.
Leptin and insulin levels are key factors regulating fetal and neonatal energy homeostasis, development and growth. Both biomarkers are used as predictors of weight gain and obesity during infancy. There are currently no prediction algorithms for cord blood (UCB) hormone levels using Artificial Neural Networks (ANN) that have been directly trained with anthropometric maternal and neonatal data, from neonates exposed to distinct metabolic environments during pregnancy (obese with or without gestational diabetes mellitus or lean women). The aims were: 1) to develop ANN models that simulate leptin and insulin concentrations in UCB based on maternal and neonatal data (ANN perinatal model) or from only maternal data during early gestation (ANN prenatal model); 2) To evaluate the biological relevance of each parameter (maternal and neonatal anthropometric variables).
We collected maternal and neonatal anthropometric data (n = 49) in normoglycemic healthy lean, obese or obese with gestational diabetes mellitus women, as well as determined UCB leptin and insulin concentrations by ELISA. The ANN perinatal model consisted of an input layer of 12 variables (maternal and neonatal anthropometric and biochemical data from early gestation and at term) while the ANN prenatal model used only 6 variables (maternal anthropometric from early gestation) in the input layer. For both networks, the output layer contained 1 variable to UCB leptin or to UCB insulin concentration.
The best architectures for the ANN perinatal models estimating leptin and insulin were 12-5-1 while for the ANN prenatal models, 6-5-1 and 6-4-1 were found for leptin and insulin, respectively. ANN models presented an excellent agreement between experimental and simulated values. Interestingly, the use of only prenatal maternal anthropometric data was sufficient to estimate UCB leptin and insulin values. Maternal BMI, weight and age as well as neonatal birth were the most influential parameters for leptin while maternal morbidity was the most significant factor for insulin prediction.
Low error percentage and short computing time makes these ANN models interesting in a translational research setting, to be applied for the prediction of neonatal leptin and insulin values from maternal anthropometric data, and possibly the on-line estimation during pregnancy.
瘦素和胰岛素水平是调节胎儿及新生儿能量稳态、发育和生长的关键因素。这两种生物标志物都被用作婴儿期体重增加和肥胖的预测指标。目前还没有使用人工神经网络（ANN）的脐血（UCB）激素水平预测算法，该算法直接使用孕期暴露于不同代谢环境（肥胖伴或不伴妊娠期糖尿病或瘦女性）的新生儿的母体和新生儿人体测量数据进行训练。研究目的如下：1）开发基于母体和新生儿数据模拟UCB中瘦素和胰岛素浓度的人工神经网络模型（围产期人工神经网络模型），或仅基于妊娠早期母体数据的模型（产前人工神经网络模型）；2）评估每个参数（母体和新生儿人体测量变量）的生物学相关性。
我们收集了血糖正常的健康瘦女性、肥胖女性或肥胖合并妊娠期糖尿病女性的母体和新生儿人体测量数据（n = 49），并通过酶联免疫吸附测定法测定了UCB中瘦素和胰岛素的浓度。围产期人工神经网络模型的输入层由12个变量组成（妊娠早期和足月时的母体和新生儿人体测量及生化数据），而产前人工神经网络模型在输入层仅使用6个变量（妊娠早期的母体人体测量数据）。对于这两种网络，输出层均包含1个变量，分别对应UCB瘦素或UCB胰岛素浓度。
估计瘦素和胰岛素的围产期人工神经网络模型的最佳架构为12 - 5 - 1，而产前人工神经网络模型估计瘦素和胰岛素的最佳架构分别为6 - 5 - 1和6 - 4 - 1。人工神经网络模型在实验值和模拟值之间表现出极好的一致性。有趣的是，仅使用产前母体人体测量数据就足以估计UCB瘦素和胰岛素值。母体BMI、体重和年龄以及新生儿出生体重是影响瘦素的最主要参数，而母体发病率是胰岛素预测的最重要因素。
低误差率和短计算时间使得这些人工神经网络模型在转化研究中具有吸引力，可用于根据母体人体测量数据预测新生儿瘦素和胰岛素值，并可能用于孕期的在线估计。
