Prenatal Cortisol Levels Estimation Using Heart Rate and Heart Rate Variability: A Weak Supervised Learning Based Approach.
Prenatal Cortisol Levels Estimation Using Heart Rate and Heart Rate Variability: A Weak Supervised Learning Based Approach.
复制标题
使用心率和心率变异性估计产前皮质醇水平：一种基于弱监督学习的方法。
DOI：
10.1109/embc48229.2022.9871718
复制
发表时间：
2022
期刊：
影响因子：
--
通讯作者：
Lindsay,Karen
中科院分区：
文献类型：
--
作者：
Cao,Rui;Huang,Yong;Rahmani,AmirM;Lindsay,Karen
皮质醇是一种类固醇激素，调节全身范围内的各种生命体征。然而，目前的皮质醇监测方法不方便日常设置。心率（HR）和心率变异性（HRV）是很容易收集的生物学参数，其波动与皮质醇高度相关，然而，目前还没有一项工作试图利用这些信号来估计皮质醇水平。在本文中，据我们所知，我们首次提出了一种基于机器学习的唾液皮质醇水平估计方法，该方法使用了佩戴心电图胸带的孕妇的HR和HRV。我们首先从心电图信号的间歇数据中提取HR和HRV参数。然后，我们应用特征选择算法来选择贡献最大的特征，并引入基于机器学习的弱监督方法来解决在实际设置中收集的标签数量不平衡的问题。实施了五种机器学习算法来执行基线皮质醇水平（BL）与两种不同皮质醇水平（CL1和CL2）的二元分类。一个深度神经网络用于在所有三个级别上执行分类。作为一项先驱研究，我们获得的预测准确率高达69% (BL VS. CL1), 71% （BL VS. CL2）和60% （BL VS. CL1 VS. CL2）。
Cortisol is a steroid hormone that regulates a wide range of vital signs throughout the body. However, current cortisol monitoring methods are inconvenient for everyday settings. Heart Rate (HR) and Heart Rate Variability (HRV) are easily collected biological parameters whose fluctuations highly correlate with cortisol, however, there does not exist a work attempting to estimate cortisol levels using these signals. In this paper, to the best of our knowledge, for the first time, we propose a machine learning-based salivary cortisol level estimation method using HR and HRV collected from pregnant women wearing an ECG chest strap. We first extract HR and HRV parameters from inter-beat-interval data derived from electrocardiogram signals. Then, we apply a feature selection algorithm to select the most contributing features and introduce a machine learning-based weak supervision method to address the unbalanced number of labels collected in real settings. Five machine learning algorithms are implemented to perform binary classification of baseline cortisol level (BL) versus two distinct cortisol levels (CL1 and CL2). One deep neural network is used to perform the classification across all three levels. As a pioneer study, we obtain prediction accuracy of up to 69% (BL VS. CL1), 71% (BL VS. CL2), and 60% (BL VS. CL1 VS. CL2).