Research On Data Processing And Intelligent Analysis In Mobile Health

Through providing medical services and information by the mobile communication technology,the mobile health(m-Health)has changed the mode of medical treatments.Such health management service as the electrocardiogram(ECG)monitoring becomes the leading trend of m-Health which meets the market demand.However,the ECG monitoring system encounters challenges on the mobility of monitoring and the precision and individualization of intelligent analysis.Although the mobile communication technology has enhanced the mobility of ECG monitoring system at the level of subjects' activity range,which allows the subject to enjoy the ECG monitoring anywhere and anytime,the mobility is inadequate at the level of motion freedom.During accessing the ECG monitoring,the activity of subject is restricted,so as to prevent motion artifacts affecting arrhythmia detections.Besides,because of the individual differences in cardiac disease data,the arrhythmia detection algorithm designed on the standard database loses its accuracy on the personalized data.Enabling the personalized and precise detection has great influences on the reliability of intelligent analysis in ECG monitoring.This thesis is aimed at resolving above mentioned questions,integrating the impulse radio ultra wide-band(IR-UWB)radar to assist the ECG monitoring system.This thesis proposed a personalized ECG monitoring scheme,which makes improvements as followings.First point is fusing multi-source data based on a cascade convolutional neural network(CNN),achieving a robust arrhythmia detection during a tester moveing slightly.This thesis has proposed a cascade convolutional neural network,which adopts one dimensional kernels for feature extractions and applies two dimensional kernels for integrating ECG signal with the heartbeat signal extracted from radar data in the feature level.Due to detecting arrhythmia by the comprehensive analysis,the cascade CNN enhances the arrhythmia classification in resisting affects of random body movements.Besides,to overcome the data imbalance problem,this thesis has designed an extended deep convolutional generative adversarial network,enlarging the dataset of the minority class for training neural networks.On the dataset with random movement interferences,the cascade convolutional neural network achives a detection accuracy 88.6%,greater than the accuracy 75%of ECG machines on the market.Second point is proposing an optimizing solution for the personalized ECG monitoring system based on the deep domain adaptation,which enhances the robustness of arrhythmia detection against individual difference interferences.This thesis has proposed a semi-supervised self-adjustable domain adaptation(SADA)approach.SADA predicts labels by self-organizing map based clustering,which is modified by transfer learning for the combination of ECG and radar data.Then,SADA adopts the one class support vector machine to overcome the issue that deep domain adaptation algorithms cannot work with unconsistent classes across domains.SADA is applied with three existing domain adaptation algorithms,increasing the sensitivity of existing algorithms as 8.8%on the large-scale dataset and 14.4%on the small-scale dataset.The third point is designing a large body movement compensation based method,which enables the vital sign detection during a target walking randomly.Given that the traditional method cannot extract vital signs of moving targets,this thesis proposes a heartbeat estimation and recovery(HEAR)approach,including the large body movement compensation and the separartion of heartbeats and respirations.HEAR builds up the mapping from echo amplitude attenuations to variations of echo reflection delay,to cancel the large body movement directly.HEAR modifies the variation nonlinear Chirp mode decomposition to extract the instantaneous frequency and instantaneous waveforms of heartbeats,and evaluates the reconstructed heartbeat signal by the heart rate variability analysis and the Bland-Altman consistency analysis.During a target walking randomly,the traditional methods cannot measure heart rates,while our proposal achives an average error rate of heart rate estimation smaller than 15%.The personalized monitoring system designed in this thesis achives heart rate detections for a subject who walks around randomly,and personalized arrhythmia detections,offering a feasible solution for implementing the m-Health in smart home scenarios.