Introduction: We describe the creation of a novel data preprocessing method to enhance the learning and predicting performance of the convolutional neural network (CNN) on classification works of predicting neurological recovery outcome from coma after a cardiac arrest on a 72 hours time scale multichannel electroencephalography (EEG) dataset. This dataset was used for the PhysioNet/Computing in Cardiology Challenge 2023, where the maximum data size for one sample can reach 148 megabytes.
Methodology: The 18-channel EEG data with a sampling rate of 100 Hz were first sliced into segments of 10-second length after discarding all the missing data, each segment was labelled with the outcome of the sample to which it belonged. During training, these segments were shuffled and fed into a neural network as independent inputs. The network we used is a simple CNN consisting of 4 convolution blocks with 2-3 convolutional layers and 1 dense block used to map the embedding to the output. During testing, the program averages the network's predictions for all valid 10s slices of the input sample and uses this as the final prediction result.
Results: Using the metrics of 'maximum true positive rate when the false positive rate < 0.05', we achieved challenge scores of 0.60, 0.54, 0.55, and 0.16 for 72, 48, 24, and 12 hours of input data respectively and ranked 18 out of the 91 teams that participated in the unofficial phase of the challenge. Our novel data pre-processing approach enhanced the robustness of the model on truncated data, hence we ranked 6th on the 24-hour data.
Conclusion: The proposed prediction model with preprocessing methods has a fair performance on the hidden test data and was able to maintain a relatively good score even on truncated data. This approach can potentially be extended to other long-time-period signal prediction tasks.