الفهرس 
introductionOnly 14 pages are availabe for public view
 |  | from | 122 |  |  |
| | | from | 122 | | |
Abstract
This thesis has been concerned with the investigation of time and frequency domain methods that are used in the analysis of heart rate variability signals. The main focus was to select the best feature extraction and classification techniques that can be used in the development of an automatic system that detects and characterizes the transient episodes in the heart rate variability signals (HRV). The HRV signals were derived from records of the American Heart Association (AHA) arrhythmia database. It consists of 80 recordings digitized at 250 Hz for a period of 30 minutes. They represent eight classes of ECG signals: Normal, Isolated uniform premature ventricular couplet (PVC), Isolated multiform PVCS, Bigeminy, R-on-T beats, Couples, Ventricular rhythms, and Ventricular fibrillation or ventricular flutter beat. 440 segments have been selected from the available records. 220 segments are transient episodes and 220 are non-transient episodes.
Four feature extraction techniques were used. They include time-domain, frequency domain and time-frequency approaches. These are: (1) Auto-regressive modeling. (2) Discrete Wavelet Transform (DWT). (3) Wavelet Packet Transform (WP). (4) Hidden Markov Model (HMM).
The features derived from the four methods were used as feature vectors to characterize the transient episodes in different HRV records using two types of neural network. These are a multi-layer feed forward neural network and competitive neural network. The Hold-Out method was used for evaluation.
It can be concluded that the highest percentage for correct classification is obtained using the set of features extracted from the autoregressive model. The multi-layer feed forward classifier performs better than the competitive neural network. It gives 97.24% correct classification for transient episodes and 92.66% for non-transient episodes. This makes an average detection rate of 94.95%.
Two schemes of data fusion were adopted in order to compensate the individual weakness of the utilized techniques and to preserve their individual strength. These are fusion at the matching score level and fusion at the decision level.
Three normalization methods were used for fusion at the matching score level. These are: Min-Max (MM), Z-Score (ZS) and Tanh (TH) normalization method. The best results obtained were obtained when using (TH) normalization method, the classification rate has been improved and reaches 99.62%.
Fusion at the decision level was applied using two methods: Dempster’s rule and Bayesian formalism. Fusion at the decision level gives better results than fusion at the matching score. The highest classification rate is 99.81% when using Bayesian formalism.
In conclusion using autoregressive model for feature extraction of different heart rate variability signals, multi-layer feed forward neural network and making fusion at the decision level by using Bayesian formalism in order to detect transient episodes in different heart rate variability signals, the developed system shows promise considering the number of cases studied. The system could possibly be tested further by increasing the database of HRV records.