Health Label and Behavioral Feature Prediction Using Bayesian Hierarchical Vector Autoregression Models
Health Label and Behavioral Feature Prediction Using Bayesian Hierarchical Vector Autoregression Models
复制标题
使用贝叶斯分层向量自回归模型进行健康标签和行为特征预测
DOI：
10.1109/embc46164.2021.9630732
复制
发表时间：
2021
期刊：
影响因子：
--
通讯作者：
Sano, Akane
中科院分区：
文献类型：
--
作者：
Lyon, Ethan N.;Victor, Luis H.;Sano, Akane
来自可穿戴设备和无处不在的传感器的数据的可用性和可访问性不断提高，允许利用计算方法来解决人类健康和行为挑战。特别地，最近的工作已经创建了用于从多维数据预测患者健康护理结果的时间序列、可解释和可推广的模型，所述多维数据包括昂贵的自我报告的患者数据、临床数据以及来自移动的和可穿戴设备的数据。在这项工作中，我们使用贝叶斯分层向量自回归（BHVAR）模型来预测大学生参与者的行为和自我报告的健康结果，这些数据来自他们的智能手机，可穿戴设备和环境以及他们的自我报告。我们还评估了模型在3、7、11和13个不同特征上的训练效果，包括一些可操作和可修改的行为特征。然后，我们展示了用许多不同类型的数据来增强自我报告数据集的价值，证明了与仅使用自我报告特征相比，可以进行额外的推断，而不会对准确性造成重大影响。我们的模型被证明是强大的，尽管大大增加了变量的数量，减少的均方误差（RMSE）的BHVAR超过患者特异性，最大似然估计（MLE）模型分别为10.5%，14.9%，26.6%，39.6%，在3，7，11，和13个变量的模型。我们还从患者水平系数的聚类分析中获得了患者水平的见解。
The rising availability and accessibility of data from wearable devices and ubiquitous sensors allow the leveraging of computational methods to address human health and behavioral challenges. In particular, recent works have created time series, interpretable, and generalizable models for predicting patient healthcare outcomes from multidimensional data including expensive self-reported patient data, clinical data, and data from mobile and wearable devices. In this work, we used a Bayesian Hierarchical Vector Autoregression (BHVAR) model to predict behavioral and self-reported health outcomes on college student participants from passively collected data from their smartphones, wearable devices, and environment, as well as their self-reports. We also evaluated how the model performed being trained on 3, 7, 11, and 13 different features including some actionable and modifiable behavioral features. Then, we showed the value of augmenting self-reported datasets with many different types of data by demonstrating that additional inferences can be made with no significant toll on accuracy in comparison to using only self-reported features. Our models proved to be robust despite the greatly increased variable count as the reduced mean squared error (RMSE) of BHVAR over the patient-specific, maximum likelihood estimate (MLE) model was 10.5%, 14.9%, 26.6%, 39.6% in the 3, 7, 11, and 13 variable models respectively. We also obtained patient-level insights from clustering analysis of patient-level coefficients.
登录
查看更多内容
影响因子：
3.2
作者：
Moustafa AA;Diallo TMO;Amoroso N;Zaki N;Hassan M;Alashwal H
通讯作者：
Alashwal H
影响因子：
5.8
作者：
Friedman, Jerome;Hastie, Trevor;Tibshirani, Rob
通讯作者：
Tibshirani, Rob
用于患者水平预测建模的贝叶斯分层向量自回归模型
影响因子：
3.7
作者：
Feihan Lu;Yao Zheng;H. Cleveland;C. Burton;D. Madigan
通讯作者：
D. Madigan
使用关系随机森林从高维纵向健康史预测健康结果
影响因子：
1.3
作者：
Z. Shahn;P. Ryan;D. Madigan
通讯作者：
D. Madigan