Feature Selection and Adaptive Connectionist Classification Models and a System for Biological Time Series Analysis on the case study data of Heart Rate Variability

biological time series;data mining;exploratory data analysis;artificial neural networks;evolving connectionist systems;adaptive resonance theory;fractal analysis;ARTdECOS;rule extraction;heart rate variability

Swope, Jay Arthur ; Kasabov, Nikola K. ; Professor George Benwell ; Dr Robert Kozma ; Geoffrey Kennedy

University of Otago

关键词: biological time series;    data mining;    exploratory data analysis;    artificial neural networks;    evolving connectionist systems;    adaptive resonance theory;    fractal analysis;    ARTdECOS;    rule extraction;    heart rate variability;   

Others  :  https://ourarchive.otago.ac.nz/bitstream/10523/2155/1/SwopeJayA2012PhD.pdf

美国|英语

来源: Otago University Research Archive

【 摘 要 】

Statement of Problem:Biological systems are constantly evolving and multi-dimensional.They have subsystems that are coupled to each other with nonlinear interactions that are time dependent.Data measured from biological systems over time are nonstationary with changing mean and variance.In order to characterise, analyse and extract information from time dependent biological data, a model must be capable of evolving, be capable of categorising dynamic information and provide a mechanism for extraction of ongoing knowledge.In this thesis we examine existing artificial neural network (ANN) models and their capabilities in the application to real time, biological time series data.We investigate existing features extracted from biological time series data.We develop ANN techniques further incorporating extracted features in an ongoing basis and providing real time extraction of knowledge.Explanation of method and procedures:We study the human biological system by examining the time series constructed from the time differences between heart beats.Measures derived from this time series are known as heart rate variability (HRV).We extract time, frequency and fractal domain HRV features.The data was collected as part of this study from 31 post myocardial subjects and 31 age and sex matched healthy subjects.The heart beat interval time series for each subject was constructed from ECG records of twenty to thirty minute duration.Existing models are explored for data modelling including fuzzy c-means clustering, fuzzy neural networks and fuzzy adaptive resonance networks (fuzzy-ART).A new ANN model ARTdECOS is constructed, which incorporates aspects of fuzzy-ART and evolving connectionist systems (ECOS).ARTdECOS is implemented on a portable data capture device to show its viability in handling real time data and to reveal issues requiring further development.Summary of results and conclusions:Category nodes generated by fuzzy ART reach expansion limits, and multiple nodes are generated to represent a single classification state.A category amalgamation procedure in ARTdECOS allows consolidation of these multiple nodes into a single node.As a consequence, meaningful rule extraction is made possible.A graphical representation of feature boundary limits allows a quick and convenient way to extract knowledge from classification results in ARTdECOS.State switching dynamics are evident in HRV time series data through segmenting of data from individual subjects.Real time scaling of features is necessary to implement ARTdECOS in a real time environment.This is accomplished in ARTdECOS by rescaling weight vectors when input features are rescaled.ANN models are a useful tool in understanding and extracting knowledge from biological time series data.These tools may be applied to biofeedback applications in real time, ongoing environments.Fractal features provide a representation of the complexity of biological time series data, as part of multiple feature extraction across feature domains.Future research includes constructing ANN models that incorporate results generated over short time intervals into temporally global space.The global model would also incorporate anomaly information, for instance ectopic detection in HRV applications.Additional integrative ANN modelling is needed to provide a supervisory system to incorporate the addition of expert knowledge.