论文的related work部分怎么写

第 II 节讨论了 BGL 预测领域的最新技术。

Based on knowledge requirement, BGL prediction models could be classified into three main groups of physiological (extensive knowledge), hybrid (intermediate knowledge) and data-driven (black-box approach) models [3]. Data-driven models establish the relationship between the present and past BGL and future values. ML and classical time-series approaches are widely used for building these models [3], [18], [19]. The following section briefly discusses some recent works for BGL prediction.
根据知识要求，BGL 预测模型可以分为三大类：生理（广泛知识）、混合（中间知识）和数据驱动（黑盒方法）模型 [3]。数据驱动的模型建立了现在和过去的 BGL 与未来价值之间的关系。 ML 和经典的时间序列方法被广泛用于构建这些模型 [3]、[18]、[19]。以下部分简要讨论了一些最近的 BGL 预测工作。

In their study, Mirshekarian et al. [20] investigated several experiments for BGL prediction using continuous glucose monitoring (CGM), insulin, meal, and activity data from simulated and real T1DM datasets in the prediction horizons of up to one hour. They used the data of two diabetes simulators (i.e., AIDA and UVa/Padova) as synthetic datasets and the Ohio T1DM dataset as the real one and developed a new memory-augmented LSTM for the time series forecasting task. They also considered an autoregressive integrated moving average model as a baseline and observed that the LSTM model meaningfully outperformed the baseline model. Based on the comparison results of the experiments for in-silico and real data, they found that the designed neural attention module improved prediction performance in synthetic data, although it failed to improve it in the real data. Contrarily, using day time as an extra input of the LSTM model enhanced BGL prediction performance only in real data. They concluded that the attitude of synthetic and real data is not always the same. Finally, by examining the LSTM on real data, they found that adding skin conductance and heart rate to BG, insulin, meal, and time could improve prediction performance.
在他们的研究中，Mirshekarian 等人。 [20] 在长达一小时的预测范围内，使用来自模拟和真实 T1DM 数据集的连续血糖监测 (CGM)、胰岛素、膳食和活动数据研究了几个 BGL 预测实验。他们使用两个糖尿病模拟器（即 AIDA 和 UVa/Padova）的数据作为合成数据集，将俄亥俄 T1DM 数据集作为真实数据集，并为时间序列预测任务开发了一种新的记忆增强 LSTM。他们还将自回归集成移动平均模型作为基线，并观察到 LSTM 模型明显优于基线模型。基于模拟数据和真实数据的实验比较结果，他们发现设计的神经注意力模块提高了合成数据中的预测性能，尽管它未能在真实数据中提高预测性能。相反，使用白天时间作为 LSTM 模型的额外输入仅在真实数据中增强了 BGL 预测性能。他们得出的结论是，合成数据和真实数据的态度并不总是相同的。最后，通过在真实数据上检查 LSTM，他们发现将皮肤电导和心率添加到 BG、胰岛素、膳食和时间可以提高预测性能。

Similarly, Martinsson et al. [21] presented an end-to-end system for predicting BGL in the prediction horizons of 30 and 60 minutes. To develop and evaluate their system, they used the Ohio T1DM dataset by considering the history of BGL as input and proposed a recurrent neural network (RNN) model for the regression task. They also estimated certainty for the predicted values, and uncertainty was the standard deviation (SD) of the prediction achieved by a parameterised univariate Gaussian distribution over the output. The mean and SD of the root mean square error (RMSE) over six T1DM patients using their proposed model was 18.867 ± 1.794 mg/dl and 31.403 ± 2.078 mg/dl for the 30- and 60-minute prediction horizons, respectively.
同样，Martinsson 等人。 [21] 提出了一个端到端系统，用于在 30 和 60 分钟的预测范围内预测 BGL。为了开发和评估他们的系统，他们使用俄亥俄州 T1DM 数据集，将 BGL 的历史作为输入，并提出了用于回归任务的递归神经网络 (RNN) 模型。他们还估计了预测值的确定性，不确定性是通过输出上的参数化单变量高斯分布实现的预测的标准偏差 (SD)。在 30 分钟和 60 分钟预测范围内，使用他们提出的模型对 6 名 T1DM 患者的均方根误差 (RMSE) 的平均值和 SD 分别为 18.867 ± 1.794 mg/dl 和 31.403 ± 2.078 mg/dl。

Moreover, Xie and Wang [9] evaluated a set of well-known ML approaches for predicting the BGL of people with T1DM using the data of the BGL, insulin injected, carbohydrate intakes,and exercises as inputs measured in the Ohio dataset. Further-more, a classical autoregression with exogenous inputs (ARX) model was benchmarked against 10 different ML models. These models included Elastic-Net, Lasso, Huber, Random-Forest, Gradient-Boosting-Trees, Ridge, and support vector (with both linear and radial basis kernels) regressions along with two deep learning models (i.e., vanilla LSTM and temporal convolution networks). Their results showed that the ARX model and Ridge regression had the lowest average RMSE (19.48 ± 2.91 mg/dl) in the prediction horizon of 30 minutes for BGL prediction. However, the ARX model had worse robustness compared to the NNs. It over-predicted peaks while under-predicting valleys.
此外，Xie 和 Wang [9] 使用俄亥俄州数据集中测量的 BGL、胰岛素注射、碳水化合物摄入和锻炼数据，评估了一组著名的 ML 方法，用于预测 T1DM 患者的 BGL。此外，具有外源输入的经典自回归 (ARX) 模型针对 10 种不同的 ML 模型进行了基准测试。这些模型包括 Elastic-Net、Lasso、Huber、Random-Forest、Gradient-Boosting-Trees、Ridge 和支持向量（具有线性和径向基核）回归以及两个深度学习模型（即 vanilla LSTM 和时间卷积网络）。他们的结果表明，ARX 模型和 Ridge 回归在 BGL 预测的 30 分钟预测范围内具有最低的平均 RMSE (19.48 ± 2.91 mg/dl)。然而，与 NN 相比，ARX 模型的鲁棒性更差。它高估了峰值，而低估了谷值。

Jeon et al. [22] performed another investigation for predicting BGL in the prediction horizon of 30 minutes using the Ohio dataset. In their previous work [23], it was postulated that a gradient-boosted regression tree model outperformed a random forest regression and an LSTM model in predicting BGLs. They further found that the missing data of the sensors had been a challenging factor in BGL prediction. Furthermore, they explored the impact of 19 physiological and monitoring variables provided in the Ohio dataset. By grouping the variables into four classes and creating 15 combinations of these groups, they concluded that using all feature classes could benefit BGL prediction by evading probably lost information. They also examined 11 different imputation techniques and validated their methodology using two traditional train-test and online settings. They then selected five missing data imputation approaches to apply, including linear, spline, Stineman, Kalman, and the last-observed-carried-forward interpolations. They finally combined the predictions to generate an ensemble model and demonstrated that the ensemble model made better BGL predictions in both settings compared to the individual predictive models.
全等人。 [22] 使用俄亥俄数据集进行了另一项研究，用于在 30 分钟的预测范围内预测 BGL。在他们之前的工作 [23] 中，假设梯度增强回归树模型在预测 BGL 方面优于随机森林回归和 LSTM 模型。他们进一步发现，传感器的缺失数据一直是 BGL 预测的一个挑战因素。此外，他们探索了俄亥俄数据集中提供的 19 个生理和监测变量的影响。通过将变量分为四个类并创建这些组的 15 个组合，他们得出结论，使用所有要素类可以避免可能丢失的信息，从而有利于 BGL 预测。他们还检查了 11 种不同的插补技术，并使用两种传统的训练测试和在线设置验证了他们的方法。然后，他们选择了五种缺失数据插补方法来应用，包括线性、样条、斯廷曼、卡尔曼和最后观察到的前向插值。他们最终将这些预测结合起来生成一个集成模型，并证明与单个预测模型相比，集成模型在两种设置中都能做出更好的 BGL 预测。

Zhu et al. [24] proposed a model using dilated RNNs for predicting BGL in the prediction horizon of 30 minutes. After investigating vanilla RNN, LSTM, and GRU cells, they selected a vanilla RNN cell to build the final model. The model was trained by BGL history data, bolus, and meal intake of the Ohio T1DM dataset and data from the UVa/Padova simulator. Overall, they observed that the performance of the proposed model for BGL prediction in the synthetic dataset was better compared to the Ohio dataset. In addition, their results showed that preprocessing steps such as interpolation and extrapolation could decrease the average of RMSE by 0.3 mg/dl. Applying transfer learning to exploit other subjects’ data was useful for one subject with various missing data. Their model had a smaller RMSE compared to autoregressive, support vector regression, and conventional NNs. Hence, they expressed that the dilated RNN model could improve the performance of BGL prediction and suggested adding the exercise data to the input for future investigation.
朱等人。 [24] 提出了一个使用扩张 RNN 的模型，用于在 30 分钟的预测范围内预测 BGL。在研究了 vanilla RNN、LSTM 和 GRU 单元之后，他们选择了一个 vanilla RNN 单元来构建最终模型。该模型由俄亥俄州 T1DM 数据集的 BGL 历史数据、推注和膳食摄入量以及来自 UVa/Padova 模拟器的数据进行训练。总体而言，他们观察到，与俄亥俄数据集相比，所提出的 BGL 预测模型在合成数据集中的性能更好。此外，他们的结果表明，插值和外推等预处理步骤可以将 RMSE 的平均值降低 0.3 mg/dl。应用迁移学习来利用其他受试者的数据对于一个具有各种缺失数据的受试者很有用。与自回归、支持向量回归和传统 NN 相比，他们的模型具有更小的 RMSE。因此，他们表示扩张的 RNN 模型可以提高 BGL 预测的性能，并建议将运动数据添加到输入中以供将来研究。

Guemes et al. [25] introduced a data-driven approach for predicting nocturnal adverse glycaemia to alarm people with T1DM to take precautionary actions. To generate and evaluate their methodology, they used the Ohio dataset by considering CGM data, carbohydrate intake, and bolus during day time as inputs for the models. Accordingly, they developed three classification methodologies for predicting the occurrence of hypoglycaemia, normoglycaemia, and hyperglycaemia during bedtime by investigating several well-known binary classification algorithms and then presented the feasibility of the overnight glycaemia prediction. Based on their report, the extended tree classifier and support vector machine performed better at nocturnal normoglycaemia and hypoglycaemia prediction, while the random forest classifier predicted better hyperglycaemia. They further suggested applying state-of-the-art classification approaches such as LSTM networks using a larger dataset as future work.
格梅斯等人。 [25] 引入了一种数据驱动的方法来预测夜间不良血糖，以提醒 T1DM 患者采取预防措施。为了生成和评估他们的方法，他们使用俄亥俄州数据集，将白天的 CGM 数据、碳水化合物摄入量和推注量作为模型的输入。因此，他们通过研究几种著名的二元分类算法，开发了三种分类方法来预测睡前低血糖、正常血糖和高血糖的发生，然后提出了过夜血糖预测的可行性。根据他们的报告，扩展树分类器和支持向量机在夜间正常血糖和低血糖预测方面表现更好，而随机森林分类器预测高血糖更好。他们进一步建议应用最先进的分类方法，例如使用更大数据集的 LSTM 网络作为未来的工作。

Rodriguez et al. [26] to enhance the management of T1DM, analysed extensive glycemia-related data of 25 people with T1DM collected from a monitoring period of 14 days within the context of the Internet of Things. To model BGL through patterns’ identification, glycaemia, insulin, meal, steps count, heart rate, and sleep data were collected via various biosensors. The authors, to model glycaemia, used and compared four techniques; including Gaussian processes with radial basis function kernels, multi-layer perceptron, support vector machines, and bayesian regularised neural networks (BRNN). Their results showed that BRNN offered the best performance on R-squared and RMSE criteria and hence was the most capable technique for BGL modelling.
罗德里格斯等人。 [26] 为加强对 T1DM 的管理，分析了物联网背景下从为期 14 天的监测期收集的 25 名 T1DM 患者的大量血糖相关数据。为了通过模式识别来模拟 BGL，通过各种生物传感器收集血糖、胰岛素、膳食、步数、心率和睡眠数据。作者为了模拟血糖，使用并比较了四种技术；包括具有径向基函数内核的高斯过程、多层感知器、支持向量机和贝叶斯正则化神经网络 (BRNN)。他们的结果表明，BRNN 在 R 平方和 RMSE 标准上提供了最佳性能，因此是 BGL 建模最有效的技术。

Although many studies have focused on this area of research, researchers are still exploring various ML approaches for predicting BGL. Moreover, it is worth mentioning that the used Ohio dataset in these works had only six T1DM patients, then some studies used an in-silico dataset along with the Ohio dataset. The current dataset used in this work now includes data collected from 12 T1DM patients, providing a more extensive dataset for developing and evaluating different models.
尽管许多研究都集中在这一研究领域，但研究人员仍在探索用于预测 BGL 的各种 ML 方法。此外，值得一提的是，这些作品中使用的俄亥俄州数据集只有六名 T1DM 患者，然后一些研究使用了计算机数据集和俄亥俄州数据集。这项工作中使用的当前数据集现在包括从 12 名 T1DM 患者收集的数据，为开发和评估不同模型提供了更广泛的数据集。

论文的related work部分怎么写相关推荐

科技论文新在哪里？-如何写好科技论文之我见(三）
科技论文新在哪里? -----如何写好科技论文之我见(三) 闵应骅全世界在计算机.通信.电力.电机等方面前10名的杂志有60-90%是IEEE出版的.可见,一般来说,IEEE杂志上发表的文章水平是比 ...
论文研究范围从什么角度怎么写？
小伙伴们写论文时,研究范围是需要较早确定的项目,如果想不好研究范围,就会眉毛胡子一把抓,论文堆砌的乱七八糟,被导师批的七荤八时.接下来我从三个方面来为小伙伴们介绍论文研究范围,希望小伙伴们看完之后能有 ...
计算机论文中的技术路线,论文中的技术路线图怎么写
论文中的技术路线图怎么写一.技术路线图论文选题基于政策工具的政策-技术路线图(P-TRM)框架构建与实证分析以中国风机制造业和光伏产业为例基于SAO分析的技术路线图构建研究一种新型技术路线图构 ...
c语言报告对老师的致谢,论文致谢对老师的感谢-论文老师的致谢词怎么写？
急求论文致谢,主要是夸奖老师的就好了,不要太恶心的--诚恳点的? 首先应该感谢师大.百年师大,像一位安详智慧的老人,以其厚重和宽博,毫无保留地给了我历史的底蕴和时代的视角,让我在百年积淀中汲取师者的精 ...
计算机论文摘要200字模板,论文的摘要范文(论文摘要是摘抄还是自己写)
论文摘要200字模板情报汇编依依朗读最近,一个国际团队开发了一种人工智能算法,该算法在社交技能上超越人类,并且可以与人或机器合作玩各种两人游戏. 论文摘要是摘抄还是自己写该算法称为S(" ...
会计论文选题什么方面比较好写且重复率较低？
会计一度被称为"最适合"女生学的专业,那么,大家身边是不是都有会计专业的同学或者朋友呢,或者说你自己就是一个"会计人"!如果你的专业也是会计,那你是不是将要面临 ...
专业论文翻译，英文摘要如何写比较好
据了解,大多数作者在写英文摘要时,直接把中文论文摘要翻译成英文.导致一部分英文读者不能对论文有一个较为完整的了解.那么,针对专业论文翻译,翻译公司如何写英文摘要呢? 首先,强调英文论文摘要的完整性如 ...
期刊论文发表时作者简介怎么写
期刊论文发表时作者简介怎么写?作者简介主要是对作者个人信息的标注,发表论文应当对自己的文章进行详细的署名标注,除了标注基本信息,有时候还需要写作一段简短的作者简介,这是发表论文的基本要求,是毋庸置疑的 ...
闵帆老师论文写作课心得体会——如何写好一篇论文
目录前言一.学术论文二.Latex排版及数学公式三.关于英文四.题目--论文的第一印象五.摘要及关键词六.引言--完整的故事七.文献综述要重视八.伪代码--论文的核心之一九.实验- ...

论文的related work部分怎么写

论文的related work部分怎么写相关推荐

最新文章

热门文章