EEG Based Recovery Prediction by convLSTM and XGBoost

Muhammad Zoraiz Ramay1 and Muhammad Akram2
1Department of Computer and Software Engineering, National University of Sciences and Technology, Islamabad 44000, Pakistan, 2Professor, CEME, NUST Pakistan


Aims: Cardiac arrests is leading cause of death globally confirmed by WHO. Severe brain injury is common with it and patients enter into coma. Among survivals of cardiac arrest, brain injury is leading contributor of death. Physicians based on EEG data offer a prognosis. Automated analysis of continuous EEG data has the potential to improve prognostic accuracy, reduce time and labor cost. In PhysioNet challenge 2023, we propose algorithm that use basic clinical information and continuous EEG recordings to predict the level of neurological recovery for cardiac arrest patients who are hospitalized in the state of coma.

Feature Extraction: Data comprises of 18 channel EEG recordings and basic clinical data. From EEG recordings, spectrogram is extracted as time-frequency feature. For robustness; time and frequency noise is added to spectrograms. RMS(root mean square) of a specific channel across hourly timestamps is summed up and it is done for all 18 channels; resulting in array of 18 for a single patient.

Algorithm: Our algorithm comprises of two parts convLSTM and XGBoost. convLSTM is good for capturing time changes and CNN is good for obtaining spatial features of images. Thus, convLSTM is good choice for getting hourly changes in EEG and within channels. Futhermore, we get hourly prediction on EEG which is helpful for doing interference at different hours. Two layers of convLSTM with max-pooling, batch-normalization in between them and after last convLSTM is appled. Flatten, dense layers are applied at end. XGBoost is used on clinical data and RMS array along with convLSTM dense layer to get output.

Results: Using 5 fold cross-validation, it received average of 0.56 at 72 hours.

Challenge Unofficial Score: 0.33 Team Name: MZR_UA