【作者】 马宁；

【导师】 李鸿儒； 于霞；

【作者基本信息】 东北大学 ， 控制科学与工程， 2022， 硕士

【摘要】 糖尿病是全球面临的严重健康问题之一,表现为血糖值偏离正常范围,长时间的血糖异常可导致严重的短期或长期并发症。夜间低血糖的多发性以及运动的无感知性提高临床风险,阻碍了人工胰脏的推广。通过对低血糖事件进行预测性防控,及时对无症状或潜在低血糖进行预判与干预。同时,血糖受外界和自身多种因素影响,呈现个体和自身差异,低血糖频率和严重程度存在较大不同。针对不同人群的低血糖预警方法研究对于临床血糖控制具有重要的现实意义。实际中,血糖的控制效果不只是降血糖,而是尽量减少血糖波动。临床血糖评判指标血糖变异系数(Coefficient of Variation,CV)能够反映血糖波动情况,波动越剧烈出现低血糖的频率越大。同时,二型糖尿病患者占比高,个体特异性大,对于血糖波动的预判更加困难,危害性更大。因此,本文将二型患者划分为[0,0.26)、[0.26,0.36)、[0.36,0.46)和[0.46,+∞)四种不同群体,设计适应不同群体的低血糖动态预警方法。其中,[0,0.26)群体发生低血糖概率相对较小、易于预测不进行研究。本文的主要研究工作和成果如下:(1)针对特定人群CV∈[0.26,0.36)表现为患者数量多、波动性较小且不可忽视,提出基于多尺度卷积神经网络的低血糖概率预警。此方法引入卷积神经网络旨在克服血糖数据特征不易手工提取和选择的问题。同时,通过构造不同尺度的卷积通道进行多尺度特征提取,实现血糖长期变化和短时波动的学习。采用输出概率的表现形式避免标签判别中阈值设定导致的偏差过大的问题,以提供更详细的辅助决策信息。最后,通过临床数据验证了方法的有效性。(2)特定人群[0.36,0.46)和[0.46,+∞)患者数据少且波动明显,深度学习方法和直接预测难以适用,将两群体预警作为分类问题研究时需要重点考虑模型的输入。目前,血糖预测和预警模型采用原始血糖值及其他生理信息等作为输入,忽略临床专家知识的可用性。在特征提取时采用单一长度存在信息不全面的问题,本文提出结合专家知识的多尺度特征提取。针对多尺度特征矩阵冗余问题,提出Relief-SVM-RFE两步特征选择方法,旨在获得具有更高可靠性和解释性的特征及尺度组合。通过实验分析,此方法在时间复杂度具有优势,且获得的特征组合具有很好的临床参考价值。(3)针对特定人群CV∈[0.36,0.46)具有较大的血糖波动、数据量少且血糖预测模型效果在可接受范围内,提出基于特征工程的低血糖动态预警方法。首先,针对血糖波动性采用多尺度特征提取来增加特征信息含量。其次,使用改进的支持向量机回归进行血糖值预测,与多尺度特征融合并作为分类模型输入。除此以外,针对患者夜间或运动前后血糖波动性差异明显的问题,在状态可判别下引入多模态划分的思想提取更精确的特征表示。通过临床数据验证了此方法在群体的预警效果。(4)群体CV∈[0.46,+∞)波动剧烈且患者数据少,其预测及模态划分的准确性难以保证,本文引入三支决策思想提出基于稀疏WGPR的低血糖动态预警方法。旨在对难以预判的低血糖事件引入更多信息获得最终决策。通过扭曲高斯过程回归获得血糖预测值和均值,实现置信区间的构造并分为预测阈值区域和两个边界区域。对边界区域分别采用基于特征工程的低血糖动态预警方法,采取的边界划分思想适用于临床决策的判定过程。其中,通过相似性度量、模型输入长度及训练集更新机制的设计实现预测模型稀疏化和个性化。通过验证,此方法具有较低的时间复杂度和人群的适应性。综上,针对不同人群特点进行低血糖预警方法设计,能够提高整体的预警效果并加强临床依赖性。低血糖预警方法通过对即将发生的低血糖事件进行预报,有助于患者和医护人员提前采取措施,对于糖尿病患者治疗的效果以及避免低血糖事件的发生起到了积极的作用。更多还原

【Abstract】 Diabetes is one of the serious health problems in the world,and prolonged abnormal blood glucose can lead to short-or long-term complications.The nocturnal frequency of hypoglycemia and the insensitivity of motion increase clinical risk,leading to hindering the promotion of artificial pancreas.Through predictive of hypoglycemia,timely prediction and intervention of asymptomatic hypoglycemia.However,blood glucose is affected by many factors,showing individual and self-differences,and the frequency and severity of hypoglycemia are quite different.The research on hypoglycemia alarm methods for different populations has important practical significance for clinical blood glucose control.The effect of blood glucose control is not only to reduce blood glucose but to minimize fluctuations.The Coefficient of Variation(CV),a clinical evaluation index of blood glucose,can reflect the fluctuation of blood glucose.The more severe the fluctuation,the greater the frequency of hypoglycemia.The proportion is high for type 2 diabetes and the individual specificity is large.It is more difficult to predict blood glucose fluctuations.Therefore,the type 2 patients were divided into[0,0.26),[0.26,0.36),[0.36,0.46)and[0.46,+∞)groups,and a dynamic early alarm method of hypoglycemia adapted to different groups was designed.The relatively low probability of hypoglycemia will not be discussed for the group[0,0.26).The main research contents of this paper are as follows:(1)The group CV ∈[0.26,0.36)has a large number of patients,and small volatility but cannot be ignored.A multi-scale convolutional neural network-based early alarm of hypoglycemia probability is proposed.This method introduces a convolutional neural network to overcome the problem that blood glucose data features are not easy to be manually extracted and selected.By constructing convolution channels of different scales for multi-scale feature extraction,the learning of long-term changes and short-term fluctuations of blood glucose is realized.The expression of output probability is adopted to avoid the problem of excessive deviation caused by threshold setting in label discrimination and provide more detailed decision information.Finally,the effectiveness of the method is verified by clinical data.(2)The groups[0.36,0.46)、[0.46,+∞)have fewer patient data and obvious fluctuations,so deep learning and prediction are difficult to apply.Taking the early alarm of the two groups as a classification problem needs to focus on the input of the model.Currently,blood glucose prediction and early alarm models use raw blood glucose and other physiological information as input,ignoring the availability of clinical expert knowledge.There is a problem of incomplete information when using a single length in feature extraction.This paper proposes a multi-scale feature extraction combined with expert knowledge.Aiming at the multi-scale feature matrix redundancy problem,a Relief-SVM-RFE two-step feature selection method is proposed,aiming to obtain feature and scale combinations with higher reliability and interpretability.Through experimental analysis,this method has advantages in time complexity,and the obtained feature combination has a good clinical reference value.(3)Aiming at the group[0.36,0.46)with large blood glucose fluctuation,a small amount of data.The effect of the blood glucose prediction model within an acceptable range and a dynamic early alarm method of hypoglycemia based on feature engineering is proposed.First,multi-scale feature extraction is used to increase the feature information content for blood glucose volatility.Second,the blood glucose prediction is performed using an improved support vector machine and fused with multi-scale features as input to the classification model.In addition,for the problem of obvious differences in blood glucose fluctuations between patients at night or exercise,the multi-modal division is introduced.It can extract more accurate feature representations when the state can be discriminated against.The early alarm effect of this method in the population was verified by clinical data.(4)The severe fluctuations in the population CV ∈[0.46,+∞)and the lack of patient data,the accuracy of prediction and modal division is difficult to guarantee.Three decision-making ideas are introduced to propose a dynamic early alarm for hypoglycemia based on sparse WGPR.This method is designed to bring more information to final decision-making in difficult-topredict hypoglycemic events.The predicted value and mean of blood glucose were obtained by distorted Gaussian process regression,and the confidence interval was constructed and divided into a prediction threshold region and two boundary regions.For the boundary area,the dynamic early alarm method of hypoglycemia based on feature engineering is adopted,and the boundary division idea adopted is suitable for the judgment process of clinical decision-making.Among them,the prediction model optimization is realized through the design of similarity measure,model input length,and training set update mechanism.Through verification and analysis,this method has low time complexity and population adaptability.To sum up,the design of early alarm methods for different groups can improve the overall early alarm effect and strengthen the dependence.Hypoglycemia alarm helps patients and medical staff to take measures in advance by predicting upcoming hypoglycemia events.It plays a positive role in the treatment effect of diabetic patients and the avoidance of hypoglycemia events.更多还原