Utilizing electroencephalography for brain monitoring contributes to eliminate subjective factors in predicting neurological outcomes after a cardiac arrest. Both the clinical features and continuous EEG recordings have the potential to improve prediction performance of neurological recovery. This study aimed to develop a multimodal model that takes clinical information and EEG data as input and automatically neurological recovery for patients after cardiac arrest. In our approach, each EEG recording was normalized with a Z-score normalization to handle the baseline shift problem. We extracted the last 1 minute of an EEG recording. Wavelet transformation was applied to reduce the noise in signals. Our base model had a ResNet module to capture EEG features and an Adaboost module to capture clinical features. The outputs of the two modules were integrated together with the attention mechanism. And finally peer-to-peer learning was performed by the framework. We adopted a distribution-wise cross evaluation method to select a robust model with good generalizability. Our team, PA_HB, received a challenge score (true-positive rate at a false-positive rate of 0.05) of 0.313 on the online validation set.