Muhammad Farhan Safdar
supervisor: Robert Marek Nowak, Piotr Palka
The non-invasive Electrocardiogram (ECG) signals are useful in heart condition assessment and are found helpful in diagnosing cardiac diseases. However, traditional ways require effort, knowledge, and time to interpret the ECG signals due to large data size. Neural networks have shown to be efficient recently and can play an essential role in interpretation. The purpose of this work was to increase the classification accuracy and to reduce the data size by retaining the essential information through Fourier Transformation.
In this study, we adopted the diverse approach by acquiring spectrograms as an input to convolutional neural network model. A large publicly available PTB-XL dataset was utilized, from which two datasets were prepared i.e., spectrograms and raw signals to classify the signals as myocardial infarction. The signal denoising, unnecessary frequency filtration and Short Time Fourier Transformation were applied to generate the spectrograms. Further, we performed up and down sampling of the signals at various points and accuracies attained. The classification model was assessed on spectrograms and raw signal datasets separately. Study results revealed that the spectrograms achieved high accuracy of 99.06% with 100% precision and 0.04 minimum loss. On the other hand, 75.93% accuracy was obtained on raw signal dataset. The conversion of raw signals into spectrograms can achieve better classification results with early convergence and holds less physical memory.