A new approach of electrocardiography (ECG) analysis system is developed to process noisy ECG signals leading for improved arrhythmia detection. The system employs two processing units comprising a novel noise reduction unit and a novel pattern recognition unit. Each unit incorporates numbers of processing techniques. In the noise reduction unit, the ECG signal is denoised before proceeding to the next stage. Four main noises contaminating the ECG signal are considered that includes baseline wander (BW), powerline interference (PLI), electromyogram (EMG) and motion artifact (MA). For BW and MA noise reduction, a novel Sqtwolog Threshold and High/Low pass (STHL) wavelet based filter is used. An Improved Proportionate Normalised Least Mean Square (IPNLMS) adaptive filter is used to the effects of EMG noise and a bandstop notch filter is used to cope with PLI.The pattern recognition unit comprises feature extraction process and classification process. In this research, some features from ECG signals are extracted to be used as the input vector for classification stage. A new Rectangular Pulse Domain (RPD), feature extraction technique is proposed that operates by taking the amplitude of intersection between filtered ECG signal and rectangular pulses. The signal is divided according to R to R peak interval (RRI) before superimposing with the rectangular pulses. The research also investigates the use of the P to T peak interval (PTI) as the signal limiter as an alternative to RRI approach and has shown encouraging performance. The classification process is performed to identify the signal whether it belongs to Atrial Fibrillation (AF) or vice-versa.The extracted features are used as the input vectors to the classifier.Two novel classifiers are designed which are the Cascade Hybrid Multilayer Perceptron (CHMLP) and the Multi-Classify Hybrid Multilayer Perceptron (MCHMLP) networks which improved version of Hybrid Multilayer Perceptron (HMLP) neural network.The MCHMLP network performs a multiple classification by doing the second classification after the first classification has achieved the optimal point. Compare with the CHMLP network, however, performs the second classification process after the first classification has been done at each iteration and stops after it reaches the optimal structure. Both networks provide better results than the conventional HMLP network but the CHMLP network gives better accuracy and standard deviation results than the MCHMLP network.The combination of both the novel noise reduction and the novel pattern recognition units are used to develop the new approach of ECG analysis system in identifying the AF signal. The performance of the new ECG analysis has been tested using the MIT-BIH ECG database.
|Date of Award||27 Jan 2017|
- University Of Strathclyde
|Supervisor||John Soraghan (Supervisor) & Lykourgos Petropoulakis (Supervisor)|