The standard twelve-lead electrocardiogram (ECG) is a well-established diagnostic tool for detecting cardiac arrhythmias and abnormalities; however, not all hospitals and clinics worldwide have access to this equipment. The 2021 PhysioNet/Computing in Cardiology Challenge focuses on developing automated classification algorithms for twenty-seven arrhythmias using twelve-lead, six-lead, three-lead and two-lead ECGs on a large, diverse dataset. Developing algorithms for these different lead configurations will help to determine if robust, accurate classification is possible with reduced-lead ECGs.
Our approach utilizes wavelet analysis and transfer learning to create individual deep learning models for each arrhythmia per lead. We use a subset of the available leads to reduce processing time; one subset we used consists of leads I, II, V2 and V5, and another was lead II only. We convert the ECG signals to scalograms, since the deep learning network we leverage, SqueezeNet, is designed for image classification. To assign a particular diagnosis, we require a minimum one-thirds vote of the available leads. This design allows for the assignment of multiple concurrent diagnoses and also easily allows for additional arrhythmias to be included at a later time without any modification required to the existing classification system.
In the Unofficial Phase, our best performing entry for team Eagles received scores of 0.344 for the 12-lead, 0.352 for the 6-lead, 0.352 for the 3-lead, and 0.351 for the 2-lead. Validation accuracy ranged from 82-96% during training. In the Official Phase, we plan to further investigate different lead combinations in order to leverage their varied anatomical positions, since certain arrhythmias are better observed in particular leads rather than being equally observable across all leads. We also plan to use this information to set variable weights for voting, in order to favor predictions from leads that should be superior at detecting particular arrhythmias due to their physical locations.