Aim: We aim to develop a predictive model using EEG recordings and clinical data to predict the outcome of comatose patients after cardiac arrest within the first 72 hours.
Methods: The dataset consisted of 5-min, 18-bipolar channel, EEG recordings obtained hourly for up to 72 hours from 607 comatose patients. We divided the dataset into 80% training / 20% test sets (486/121 patients). The developed model was based on the Discrete Wavelet Transform (DWT), as it can adapt to the frequency content of non-stationary signals such as EEG, resulting in optimal time-frequency resolution over all frequencies. We used the most recent recording with the highest quality index, band-pass filtered between 0.56 and 40 Hz and implemented the DWT with four detail levels. For each DWT level, the mean absolute value (MAV), standard deviation, average power, entropy, and the ratio between the MAVs of adjacent bands were calculated. The 24 DWT features and patient metadata were inputted into a random forest (RF) model. A second model was developed using brain functional connectivity (FC), computed for six frequency bands, based on the corrected imaginary phase locking value. To reduce the dimensionality we implemented an autoencoder (AE) model, compressing the FC matrices using the trained encoder. We combined these features with patient metadata and non-coupling features (e.g: psd) and trained a RF model.
Results: The DWT model achieved a challenge score of 0.55 in the internal validation and 0.31 in the external validation (Team DEIB_POLIMI) at 72h. For the FC approach, the model obtained a score of 0.67 in the internal validation.
Conclusion: The FC features extracted by AE improved the internal validation score. In the official phase, we will explore the predictive power of additional FC features extracted from DWT signals while using an AE for dimensionality reduction.