Intelligent Methods For Anomaly Detection Of Health Time Series Data

With the acceleration of the aging process,the frequent occurrence of various chronic diseases has brought a heavy medical burden to families and society,and has become a major strategic problem restricting social development.Although traditional medical model based on professional medical experts and sophisticated medical equipment can accurately detect and assess various health diseases,its human and material costs are very huge,and hence cannot meet the people’s increasing health service needs.Meanwhile,the continuous development of health perception technology makes it possible to continuously collect various health data in daily environments such as home and office.On this basis,the continuous analysis of massive health data using artificial intelligence technology to achieve accurate detection and real-time warning of various diseases(i.e.,anomalies of data)will greatly reduce the cost of health care,and hence become an inevitable choice in dealing with the aging population in the future.In essence,disease is the specific manifestation of various abnormalities that occur in the course of human life.In addition,most of the health data collected in daily environment exists in the form of time series data(so it is often called "healthy time series data").Hence,using artificial intelligence technology to build anomaly detection models for health time series data is the key to achieving “smart health model”.After analyzing the characteristics of various types of health time series and the anomalies they contain,the anomalies of health time series data can be roughly abstracted and classified into three types: abnormal period,abnormal event and abnormal state.This classification method can cover most health problems,and has a high degree of abstraction,which helps to design universal anomaly detection models.However,in addition to the high dimensionality,complexity,dynamics and differences of ordinary time series data,the health time series data of the thre above three types of anomalies also have their unique characteristics,which makes it hard to build anomaly detection models.To solve this problem,the thesis focuses on the unique characteristics of each type of anomaly,and designing highly adaptable and efficient anomaly detection methods by emplying pattern recognition and deep learning technology.The main contributions are summarized as follows:1)Propose a hybrid attentional LSTM-CNN based abnormal period detection method for health time series dataThe health time series data to which the abnormal period belongs has pseudo-periodicity(such as electrocardiogram,etc.),that is,the data waveform exhibits a certain periodicity,but the duration and fluctuation pattern of each period are not the same,which makes it difficult for existing methods to precisely divide the health time series into successive period segments and to accurately characterize their fluctuation pattern.In view of this,this study proposes a two-level clustering-based automatic segmentation algorithm to automatically split periodic health time series data into successive period segments.This method does not need to set any parameters in advance according to the characteristics of the data and can eliminate the adverse effect of outliers on the segmentation accuracy,and has a strong adaptability.In addition,when modeling the fluctuation pattern of the periodic segment,this study designed a hybrid neural network model based on LSTM and CNN,which can simultaneously mine the long-term fluctuation trend and local fluctuation characteristics of the data to characterize its fluctuation pattern.Furthermore,this study also embeds three attention mechanisms in LSTM and CNN,so that the model can fine-tune the output according to the type and location of the aforementioned trends and features,and further improve the model’s fitting ability.Experimental results on 4 public datasets show that the accuracy of the proposed method reaches 99.3%,which is about 4.8%～7.7% higher than existing methods.2)Propose a fine-grained signal fluctuation parameters based abnormal event detection method for health time series dataAbnormal events are sudden changes of the physiological state of the human body within a short period of time.They are sporadic and the length of them are not the same,hence,it is difficult to accurately locate them by using existing methods.In addition,abnormal events usually have a certain internal structure,and different parts often contain different aspects of information about the abnormal events.Existing methods fail to make full use of the above structural information,resulting in limited detection accuracy.To sovle the above problems,this study proposes an abnormal event detection method based on fine-grained fluctuation parameters.First,to solve the problem that abnormal events happen unpredictably and their duration is unfixed,an automatic potential event positioning method is designed,which can obtain high positioning accuracy and guarantee the integrity of the event.Afterwards,to deal with the differences between multiple abnormal events,a sliding window based adaptive potential event segmentation method is designed,which does not rely on any fixed threshold and hence has good adaptability.Then extract some highly robust features from different parts of the potential event to accurately characterize the signal fluctuation pattern,based on which an Adaboost-based ensemble abnormal event detection mechanism is designed to determine whether the potential event segment contains abnormal events.Experimental results on real dataset show that the accuracy,recall and AUC obtained by the proposed method in detecting obstructive sleep apnea events reach 98.4%,97.6% and 0.983 respectively,which outperforms the baseline by 12.7%,14.8% and 0.152 respectively.3)Propose a multiple weighted class association rules based abnormal state detection method for health time series dataAbnormal state refers to the long-term and complex physiological state changes that occur in the human body,and different internal change patterns may cause the same external abnormal appearance.In order to guarantee the effectiveness of health services,the abnormal state detection model not only needs to achieve higher detection accuracy,but also generate interpretable detection results so as to explain the internal causes of anomalies to a certain degree.To guarantee the effectiveness of health services,the abnormal state detection model should not only accurately distinguish the abnormal state from the normal state,but its test results should also have certain interpretability to explain the internal causes of the abnormal occurrence.However,the traditional machine learning classification model is difficult to meet the requirements of high accuracy and interpretability at the same time.To this end,this study transforms the inherent association relationships among multi-dimensional features into class association rules(CARs)that can be used for classification,and constructs an abnormal state detection method based on multiple weighted class association rules.This method can not only accurately detect abnormal states,but also obtain a set of CARs that can explain the internal causes of abnormal occurrence.First,a dictionary order based CAR mining algorithm is designed,which optimizes the most time-consuming steps of the mining process and hence is able to quickly extract the complete set of CARs.Then,by investigating the support,confidence,coverage,and redundancy of CARs,a branch-based CAR selection algorithm is proposed to select a concise and representative CAR subset.Finally,based on the CAR subset,a multiple weighted CARs based universal classification model is built which solves the "no matching CAR problem" and "conflicting CAR problem" of ordinary associative classifiers.Applying the proposed method to real hypertension detection,the achieved accuracy reaches92.7%,which is about 5.3%～17.7% higher than existing methods.In summary,this thesis first abstracts and categories all kinds of anomalies occurring in health time series data into three types: abnormal period,abnormal event and abnormal state.Then,it uses advanced technology such as data mining,machine learning to design innovative methods to detect different types of anomalies.The effectiveness of the proposed methods are verified by extensive experiments,which provides theoretical methods and technical support for accurately detecting various health problems in daily living environments and constructing smart health model.It is believed that the in-depth application of the new generation of information technology in the field of health computing will make contributions to improving human health,and especially for elderly to live a healthy life and enjoy their old age.