Analysis and classification of biomedical signals by time-frequency methods

This thesis seeks the classification problem of cardiac arrhythmias with a high accuracy by using a small number of electrocardiogram (ECG) signals. To find a reasonable solution to this problem, we advise using a common matrix approach (CMA)-based classifier model fused with time-frequency analysis methods. Fourier Decomposition Method (FDM) and Short Time Fourier Transform (STFT) methods were used to generate time-frequency (T-F) representations of ECG data. The feature matrices used in classification were obtained by applying two-dimensional principal component analysis to data matrices generated from time domain, frequency domain and T-F domain representations of ECG signals. Since the data matrices are generated separately with two different T-F representations, the proposed methods for ECG arrhythmia classification are introduced under two different topics as FDM-CMA and STFT-CMA. The effectiveness of these methods is evaluated over classification assessments on ECG signals from three different publicly available databases: MIT-BIH, Chapman and PTB-XL. With the experiments conducted using the MIT-BIH database, FDM-CMA achieved 99.87% and 99.84% overall accuracy scores for the classification of ECG signals of four arrhythmia types under intra-patient and inter-patient classification scenarios, while STFT-CMA achieved 99.97% accuracy rates, respectively. With both methods, Classification of the V and S type arrhythmias reach to the rates exceeding almost 99% on the five-performance metrics was observed as a better performance than the most recent methods. To test the classification performance of both methods in the presence of noisy ECG signals, experiments were performed on noisy and noise-free ECG signals from the Chapman database. FDM-CMA method trained and tested with noisy and noise-free ECG recordings related to four arrhythmia classes achieved overall classification accuracies of 99.58% and 99.05%, respectively, while these scores were obtained 99.92% and 99.90% with STFT-CMA. The FDM-CMA and STFT-CMA methods achieved overall accuracy rates of 98.49% and 99.97%, respectively for the five different arrhythmia types taken from PTB-XL database. The obtained results show that the proposed methods resulted in high classification performance with a small number of training sets regardless of the database.