Abstract:
ECG Classification or Heartbeat Classification is an important task in cardiology. Deep
Learning based methods assist human experts in the diagnosis of cardiac diseases and
help save important lives. In this research, a deep learning framework is proposed for
classification of cardiac diseases using the heartbeats from ECG data. Multiple models
are proposed including CNN, LSTM and Autoencoders. The models are trained on
a digitized format of CPEIC Cardiac dataset. First, the ECG images are digitized,
segmenting the Lead-II heartbeats and then the digitized signals are passed to the deep
learning models. The proposed CNN model achieves highest accuracy of 97%, equivalent
to the best accuracy (98%) of image based models with all 12-leads data. The proposed
model is highly accurate and fast for inference for real-time and direct monitoring of
ECG signals. The self-supervised learning (SSL) based model is experimented on MIT BIH dataset as well and it achieves state-of-the-art results (∼96% accuracy). In the SSL
based model, after training on un-labeled data, fine-tuning on only 50% of the training
data gives 94.8% accuracy, which cuts the training time by 50% and accuracy is almost
the same, making this model computationally efficient.
