This is sollution for solving Physionet 2017 Challenge by Residual Convolution Recurrent Neural Network with 0.86 accuracy and 0.83 F1 score.

The task of competition is the classification of ECG records into 4 classes (normal, abnormal, other, noisy), the quality of classification was measured as an average F1 of three classes (N, A, O).

You can read more here.

This model combines residual convolution and recurrent layers.

Top1 result achived 0.83 F1 score on test set (still not published) and 0.91, 0.79 and 0.77 F1 scores on 5 fold cross-validation.

This solution achived 0.92, 0.8, 0.78 on same cross-validation which is slightly higher and at least comparable to the first place.

Google Colab Notebook.

Gihub repo with nice description of competition and it's data.