Introduction: Time series classification is a challenging problem particularly when the underlying dynamics of the system exhibit chaotic behavior. We will explore the simple idea that the graph of the spectrogram of a time series is sufficient for classifying the time series. Incze et al. [2018] show that it is possible to classify different types of bird songs by feeding images of the spectrogram of time series data into a MobileNet model. The authors achieved an impressive binary classification accuracy of 82% given the simplicity of their methods. We will adapt these methods to classify three different types of cardiac arrhythmias (normal sinus, simple arrhythmias & malignant ventricular ectopy) from ECG data with a pretrained image classification network such as ResNet.
Challenges: Although the project is going smoothly thus far, the size of the data and complexity of the preprocessing has slowed down our timeline. There is roughly 150 hours of ECG data between the three datasets. Note that to create each spectrogram we must 1. resample all of the data to the same target frequency, 2. apply a baseline wander algorithm, 3. apply a lowpass butterworth filter, 4. apply a highpass butterworth filter, 5. normalize the frequency bounds, 6. split the ECG signals into partially overlapping windows of 20 seconds and 7. compute the discrete Fourier transform of each window and apply a jet color mapping. We have just completed the data collection and have created spectrograms that correspond to 20 seconds of preprocessed ECG data. Our full dataset contains approximately 25,000 spectrogram images.
Insights: Given that we are using a pre-trained image classifier we should be getting preliminary results soon now that all of the images have been created. Upon 1 initial inspection, the spectrogram images appear to be significantly different depending on the type of arrhythmia present. We believe that we should easily meet our target accuracy of 60%.
Plans: Initially we were planning on using a spectral color mapping, however Incze et al. [2018] achieved their best results using a jet color map. One area that we can explore is how changing the color maps used to generate the spectrograms relates to classification accuracy. Additionally, it appears the the use of common ECG preprocessing methods such as applying lowpass and highpass butterworth filters has a significant impact on the look of the spectrogram. If time permits, we can experiment with whether applying these standard signal processing filters influences classification accuracy. We hoped to have some preliminary results at this point in the project, however now that our spectrogram images have been create we should be able to get results soon.
Log in or sign up for Devpost to join the conversation.