Research Of Key Algorithms For Continuous Glucose Monitoring Signals Based On Data-Driven Methods

Diabetes has gradually become one of the world’s public health problems that endanger human life and affect global economic development.Artificial pancreas has gained more and more attention since it is a novel and promising way for the personalization treatment of diabetes.Continuous glucose monitoring(CGM)devices are one of the key parts of the artificial pancreas.With the improvement of sensor technology and the rapid development of computer technology,it is possible to monitor the blood glucose of patients continuously.CGM signals can provide data for hyper/hypoglycemia alert and blood glucose control.Therefore,the research on continuous glucose monitoring signals based on data-driven methods has become a hot topic for all researchers around the world.For physiological model,it is time-consuming and difficult to develop accurate mathematical models.Besides,the model-based monitoring methods may not work well since the model mismatch is a great challenge for different physiological models.With the development of data storage technology,data-driven monitoring methods play a more and more important role in continuous glucose monitoring.The analysis of CGM signals based on data-driven method has several problems.The CGM sensors are easily affected by random noise and different kinds of faults.The accuracy of the abnormal glycemic event alert based on glucose prediction has to be improved because of the prediction errors.And the assessment of glucose control performance is not comprehensive enough.To solve the above-mentioned problems,a series of signal filtering,fault detection,predictive alert and control performance assessment methods have been developed to help patients to get better glucose management,which are summarized as follows:(1)In order to solve the interindividual variability and the intraindividual variability of diabetes patients,an automatic denoising method with noise level estimation and responsive filter updating for online CGM denoising is proposed.It can quantitatively and accurately determine proper filter parameters by exploring the state transition matrix and noise level from measurement data.It avoids the uncertainty and ambiguity of conventional filter methods which have to set filter parameters artificially for different patients and thus gives more accurate filtering results for specific patients.It can automatically detect the noise variability and judge whether the filter parameters should be adjusted.A proper confidence interval is defined by quantitatively analyzing the power spectral density values of high frequency band signals instead of arbitrarily setting filter parameters.It avoids the blindness of conventional filters that have to update filter parameters continuously with time no matter when the noise level is changing or not.In this way,the accuracy of CGM signals is improved by the proposed method.(2)For distinguishing the transient loss of sensitivity and the unannounced meals accurately,a concurrent fault detection method for continuous glucose monitoring sensor is proposed by a total dynamic analysis including the serial correlation and temporal features of CGM signals.First,the canonical variate analysis is used to extract the time series correlation of normal CGM signals.The monitoring metric can be determined to detect whether the faults and unannounced meals have happened.Further,to distinguish faults and unannounced meals,the slow feature analysis is employed to extract temporal information of CGM signals.The monitoring metric which can reflect the changing speed of the glucose is defined to distinguish the loss of sensitivity from unannounced meals.In this way,the reliability of CGM signals is guaranteed by the proposed method(3)The prediction uncertainty influences the accuracy of hyper/hypoglycemia alert.To address this problem,a probabilistic soft alert method for abnormal glycemic event by quantitative analysis of prediction uncertainty is proposed.It is the first time that the concept of probabilistic soft alert is proposed for online glucose risk prediction.The probability density function of prediction errors is estimated using the Gaussian mixture model to quantify the uncertainty caused by prediction errors.Then,a soft confidence interval can be set and hypo/hyper alert distance is calculated to estimate the probability of the hyper/hypoglycemia alert to quantify the hypo/hyper alert level.In order to capture changes of prediction uncertainty with time,a proper parameter updating strategy is proposed which can automatically update the parameters of the GMM and adjust the distribution of the prediction errors to give a fine-scale alert level.(4)In order to solve the problem that traditional glycemic control metrics ignore the time series variation information of CGM signals,a fine-scale online evaluation of glycemic control performance based on temporal feature analysis is proposed.The glucose control performance within normal glycemic range is divided into three categories by analyzing the temporal information of CGM signals based on slow feature analysis to give more detailed assessment information.The proposed method can give online assessment of control performance to help improve the efficacy of the controller.Finally,the control assessment grid analysis is proposed to give an overall performance assessment by integrating static information and temporal information for both qualitative analysis and quantitative analysis,which can be used for comparison of different diabetes patients over different time period.The efficacy of the proposed method is validated based on in silico subjects,clinical subjects of open source website and clinical datasets provided by University of Cambridge.Finally,the future research work are discussed based on the conclusion of this dissertation.