Study of Biomedical Signals using Signal Processing Techniques

Abstract

Electrocardiogram (ECG) has been used by the doctors for a long time to monitor the condition of the heart and is popular due to its ease of use and non-invasive nature. Electrocardiogram (ECG) is a record of electrical activity of a heart. Detection of heart diseases at an early stage can prolong life through appropriate treatment, and the minute variations in the signal are not always easily discernible to a naked eye. Thus, our objective regarding ECG signals is to use different signal processing and pattern recognition techniques to extract useful information from them and to try to increase the classification accuracy, even in noisy signals. An attempt has been made to detect the R-peaks present in the QRS complex. This is done because R-peaks are easier to detect and further help us, to extract features which will help in classification. The signals used in this work, have been downloaded from a very popular dataset MIT-BIH database, which has been extensively used in prior scientific research. When the algorithm was applied on the dataset, it was found that almost all the peaks were being detected in the signal. Further the techniques of noise reduction and baseline-wander reduction were also applied on the signals. It was achieved by converting the signal from time domain to frequency domain. After noise filtering, the wavelets transform was used to extract features from the signal. Wavelet transform was used because, it works in both time domain and frequency domain, which allows to extract more information from the signal. First, the signal was filtered to remove noise and then different features were extracted from the wavelet coefficients like mean, mode, variance, standard deviation, Shannon entropy, Spectral energy etc. Then these features were used to see, which ones could help in classification of the signal. It was found out that for most of the feature set could not classify the ‘Disease’ signal from ‘Non-Disease’ signal. Only mean vs log energy graph showed some non-overlapping dataset, and thus could be used for further classification. Finally, ANN (Artificial Neural Network) and SVM (Support Vector Machine), was attempted to classify the signals. Both are Machine Learning algorithms, but work on different principles. At first a model was created using SVM, to classify the signals. However, the model could provide only 50% accuracy. The reason for less accuracy was due to the dataset, which was used for training, was highly overlapped. Next, a model was created using ANN, which 2 provided better results. The training and testing accuracy was about 95%-96%, and the algorithm was tested on 9 disease and 9 non-disease signals, in which all the signals were successfully classified.