Abstract :
A new class of prediction-based independent training (PBIT) networks for temporal patterns classification is proposed. The authors´ approach combines the universal neural network (NN) approximator and TDNN. The input vectors of PBIT are consecutively created by the time-delayed segment of a pattern. To demonstrate the feasibility of PBIT, extensive simulations on electrocardiogram (ECG) classification are conducted. They have respectable training performance and relatively good generalization accuracy. Infinite impulse response (IIR) filters can be adopted as preprocessors to extract information out of the time sequence so that smaller networks will suffice. For a comparative study, another independent training classifier-the hidden Markov model (HMM)-is included for temporal patterns, and some mutually training (MT) models, e.g., the (static) decision-based neural network, are also included. Hybrid IT and MT techniques are proposed to further improve classification accuracy
Keywords :
digital filters; electrocardiography; medical diagnostic computing; medical signal processing; neural nets; pattern recognition; ECG classification; IIR filters; PBIT networks; TDNN; decision-based neural network; hidden Markov model; hybrid techniques; infinite impulse response filters; mutually training models; prediction-based independent training networks; preprocessors; temporal patterns classification; time-delayed segment; universal neural network approximator; Electrocardiography; Hidden Markov models; IIR filters; Neural networks; Pattern classification; Performance analysis; Predictive models; Robustness; Transient analysis; Vectors;
