Time Series Data Mining In Biomedical Research

With the development of information technology and ultra large space storage technology and their applications in various fields, the accumulated data show the exponential growth. The massive data have a large number of hidden information with important value. How to make good use of the massive data and mine for more valuable information has been a hot topic in the field of data mining.Time-Series Data have the characteristics of both time and spatial sequences, through which can get useful information closely related to time or spatial and achieve the extraction of related knowledge. In-depth research on time-series data mining is very challenging but necessary because time-series data have the characteristic of noise interference, signal fluctuation, multi-dimensional and complex. There are various aspects on time-series data mining such as similarity queries, anomaly detection, and mode represent, association rules, classification and clustering, trend segmentation on time series data, as well as the frame work on time-series data mining.The current paper focused on four aspects of time-series data mining: framework, anomaly detection, trend segmentation and mapping function within the framework, based on domestic and international research and the practical need of time-series data mining and from the perspective of intracranial pressure, arterial blood pressure in the biological and medical area. Proposed a number of algorithms and solutions, and got several positive results. The major contribution and innovation is embodied in the following aspects:1) Framework of time-series data miningThe framework of time-series data mining is a hot topic for its application. The current research focuses on how to design a mining framework for a sequence of data and accurately estimate the unknown time series, based on the existing time series and relevant theories and techniques. In this paper, which designed one time-series data mining framework and built knowledge learning model, which is able to do fast training on related time-series, using techniques such as selecting data, extracting feature vectors, model identification, etc, and then estimate the target time-series. The experimental results show that using relevant time series this data mining frame work is able to make better estimation on target time series.2) Anomaly detection of time-series data miningThe research on anomaly detection of time-series data--how to design the corresponding algorithm to quickly and efficiently identify the abnormal data, exclude useless data and noise data, and extract valuable time-series data--is another hot topic. Since the normal search algorithm of signal abnormalities is not able to fulfill the need of other time-series data, the current research proposed an expansion of search algorithm of signal abnormalities and applied it to detect SPO2, CBFV signals in biomedical area. The experiments showed that this expansion is sensitive to noise signal, interference signal and un-expectable false signal, and is able to quickly and accurately detect abnormal data.3) Trend segmentation of time-series data miningResearch on segmentation of time series trend is another hot issue. Identification of different trends is a common problem in the various fields of time-series applications. To predict the occurrence of key events, trend analysis has been successfully used as data processing steps, and further to predict some key events working with other data analysis modules. At the mean time, in many application fields, a lot of methods have been developed to solve related problems in trend detection and time series segmentations. Based on the accumulation of residual error and some problems of time series segmentation, an adaptive segmental algorithm for time series is proposed. The experiments showed that this algorithm is better to reflect the trend of time series with small samples, and therefore has better guidance in estimation of cardiac arrest occurrence.4) Linear mapping function in time-series data mining frameworkMapping function is one of the most critical modules in time-series data mining. proposed to construct linear mapping function using total least squares(TLS), truncated singular value decomposition (TSVD), and Standard Tikhonov Regularization(STR), according to the characteristic of the existing signals. The results showed that the linear mapping function constructed by TSVD and STR has better estimation than that constructed by linear least square method and ratio of total least squares method.5) Non-linear mapping function in time-series data mining framework Proposed a nonlinear mapping function using support vector machine regression (SVR). The experiments on using it to estimate intracranial pressure signals showed that using nonlinear SVR method to predict waveforms was more consistent with actual waveforms and improved the accuracy of the prediction significantly, compared with the linear mapping function method using TLS and STR.