Reduced-Lead Electrocardiogram Classification Using Wavelet Analysis and Deep Learning

For the 2021 PhysioNet/Computing in Cardiology Challenge, we developed a deep learning model for automated ECG classification for both the standard 12-lead as well as various reduced lead sets. SqueezeNet was leveraged in order to construct this deep learning model. Training inputs consisted of scalograms, which are time-frequency representations of the absolute value of the continuous wavelet transform coefficients. Multiple preprocessing steps were necessary in order to accommodate the timing and resource constraints of the Challenge. These steps included limiting the training samples to only single diagnosis samples, truncating the signal to only the first five seconds, and only using leads I, II and V2. Our algorithm obtained a score of 0.30 on 12-lead, 6-lead, 4-lead, 3-lead and 2-lead test data and ranked 27th (among 39 teams) on 12-lead, 6-lead and 4-lead and ranked 25th on 3-lead and 26th on 2-lead test data. The mainfindingfrom this work is that though wavelet analysis and deep learning can produce high accuracy in a variety of applications, the method applied to this classification task is too computationally intensive and requires substantial improvement to make it clinically viable.