Aim: EEG signals of comatose patients post cardiac arrests show promise in predicting the CPC scores and the possibility of recovery (outcome). We aim to compare and study several DL models to recommend an accurate prediction model.
Methods: Data cleaning and pre-processing involved a two-pronged approach, separately for patient metadata and corresponding EEG signals. Missing metadata values were imputed using KNN algorithm. EEG signals were resampled and common-average referenced. Further, noise was filtered by band-pass-filter (0.5-30 Hz) and wavelet-based denoising (using DB2 wavelets). Missing signal data segments were interpolated. Feature engineering: Short Fourier Transforms and Hamming Windows were used for extracting power spectral densities of the EEG signals. The statistical data therein was used to compute the alpha-to-delta ratio of power spectral densities which are indicative of the balance of brainwave activity. Higuchi's method was deployed to obtain the fractal dimensions and complexity of each EEG channel. Model architecture: The deep learning model developed is a multi-input-multi-output (MIMO) framework. Sequential input sequences to the model were deployed as follows: Firstly, 12-hour-window EEG-signal features were sent to a Bi-LSTM network to factor temporal dependencies. The corresponding quality scores were used to classify class weights. Secondly, patient metadata was input to the second network layer (dense layer) for patient-data associations. Both layers were concatenated and fed to a fully connected layer to the output to predict CPC and Outcomes.
Results and conclusions: At this stage, the development and deployment of the models has finished. Due to computational limitations, the authors are still in the test and validation phases of the research. Preliminary results suggest that prediction accuracy through using deep learning as opposed to other classification models can yield a more accurate CPC and outcome. The authors intend to implement other models and perform hyper-parameter tuning for a full comparative study.