Our pipeline consists of a preprocessing and a training step. The first stage applies linear and nonlinear transformations on the physiologic signals in order to gain features. The second stage is a lightweight neural network classifier, which classifies measurements. The proposed algorithm was trained on 6877 annotated records of ECG signals with demographic meta-information. After R-wave detection the evenly resampled R-R intervals were averaged. Moreover, the patient age and sex data were forwarded to the second step. All the features were normalized. This averaged ECG signal is fed into a convolutional neural layer, which detects tendencies and significant levels (eg. ST). Personal meta-information is added before the last fully-connected layer, which significantly helped the classification, while the average length of R-R intervals and the standard deviation of these lengths were not useful. Rate conditions were enriched during training, as they consist of a small part of the data.
Currently the method is constrained to output the most probable category, which will be fixed before the final deadline.
Our submitted results for the entire test set were evaluated and published on the leaderboard as follows: F2 score: 0.446, G2 score: 0.230, geometric mean: 0.320. Evaluation took 11 minutes, which is among the fastest ones.
10-fold cross-validation resulted in F2: [0.485, 0.452, 0.429, 0.448, 0.458, 0.464, 0.488, 0.443, 0.488, 0.439], averaging 0.459 and G2: [0.258, 0.234, 0.220, 0.233, 0.248, 0.243, 0.197, 0.232, 0.263, 0.227], averaging 0.235.