In this study, we create a system using an combination of gradient boosted tree and recurrent neural network (RNN) classifiers that processes a 12-lead electrocardiogram (ECG) recording in order to automatically identify normal sinus rhythms as well as cardiac abnormalities including atrial fibrillation, first-degree atrioventricular block, left or right bundle branch block, ST-segment elevation or depression, and premature atrial or ventricular complexes. Our system is an entrant in the PhysioNet/Computing in Cardiology Challenge 2020.
Our system begins by downsampling the data, removing noisy regions and normalizing amplitudes. The Pan-Tompkins algorithm is used to detect QRS events, and using the boundaries created by the identified QRS peaks, P and T waves are calculated by comparing timing and noise levels across leads.
Attributes of these different QRS components, such as their amplitudes and durations, are calculated and summarized over the full duration of the input recording for each of the 12 channels. Additional input features include overall levels of noise in the recording, a patient’s demographic data, and the output of a separate long-short term memory (LSTM) network.
This feature set is then passed to an ensemble of gradient boosted s, each trained to classify data into one of the nine target classes. Ten fold cross validation was applied on the training data with the trained models producing average class-weighted F-score (F_β ) and generalised Jaccard measure (G_β ) scores of 0.616 and 0.388 on the holdout set. When tested on the unseen test set, the model achieved scores of 0.577 and 0.361, respectively.
Future ideas include training CNNs on the de-noised ECG signal, calculating additional clinically relevant features using domain knowledge, and adjusting the RNN architecture to better suit the structure of the data.