Aims: This work aimed at developing an efficient algorithm to help causally predict neurological recovery events for patients from coma after cardiac arrest.
Methods: 18-channel EEG recordings and clinical data were jointly analyzed to build an ensemble classifier for prediction. Specifically, for the time domain, standard descriptive statistics such as interquartile range, and skewness were computed. For the frequency domain, we utilized the periodogram of the signal and made an analysis for each 5-min epoch with Welch's method. The relative spectral power for certain kinds of bands and the power ratios were calculated. Non-linear features like the permutation entropy and fractal dimension were also included. Furthermore, deep neural networks were used to automatically extract deep representations of the EEG recordings. We found in our preliminary experiments that direct training on patient-level data would lead to an over-fitting problem so we carried out an analysis on the hour-level and augmented the patient-level data, which remarkably mitigated the over-fitting and dramatically promoted the feature extraction ability of the deep learning method. We then took all these features extracted from EEG recordings with clinical features to train an ensemble classifier and efficiently predict neurological recovery outcomes.
Results: We applied a 10-fold cross-validation method to evaluate the performance of the ensemble model on the training data, which resulted in an average of 0.67 for the challenge score for 72 hours. Moreover, we scored 0.43 for 72 hours on the validation data in the leaderboard with only the expert features used.
Conclusion: The ensemble model trained with both expert features and deep representations from neural networks is promising to promote the prediction performance for neurological recovery, as is shown in our existent cross-validation experiments.