Table 2.
Description of the content of working subdirectories in the repository attached to the paper.
| Directory name | Description |
|---|---|
| acc_loss | Stores training and validation loss curves side by side, as well as the training and validation accuracy curves |
| best_models | Stores the best CNN models obtained during training (in terms of model weights, i.e. trainable parameters). The data is saved in the binary HDF5 format. Best models can be loaded later and thus there is no need to train the neural network every time when you want to run a classifier for test data. An example of how to load a best model is shown in the enclosed Jupyter notebook (the load_weights function) |
| hists | Stores models’ training and validation accuracy and loss values. This data allows you to prepare visualizations of network training, similarly to those depicted in Figs. 13 and 14. The data is saved in the PCKL format (implemented in the Python’s pickle module) and as CSV text files. An example of how to use these files is shown in the enclosed Jupyter notebook (the pickle.load function) |
| inputs | Stores HDF5 files which are inputs for our CNN model. These files are created in R (EEG_neonatal.R script) using the raw EDF files which are stored in edf directory. To find out exactly which fragments of the original raw EDF files were used in HDF5 files (i.e. the exact samples numbers), files with names beginning with non_seizures_ and seizures_ are additionally generated |
| logs | Stores log files to be parsed by TensorBoard (TensorBoard is a tool for providing the measurements and visualizations needed during the machine learning workflow). |
| results | Stores CNN classification results of the validation and test datasets (given in %). The classification results presented in Tables 3, 4 and 5 are the average of the validation scores obtained using K-fold validation scheme. Additionally, execution times for every fold and GPU card types are given |
| ROC | Stores ROC curves along with the AUC metrics |
| waveforms | Stores all EEG seizure waveforms annotated by 3 experts. There are 1379 waveforms in total, as depicted in Table 6. The lengths of the waveforms were arbitrarily set at 10 seconds. However, the user can generate waveforms with different lengths by running the generate_eeg_waveforms function in R. See the Supplementary Information files for details how to do this |