A children’s encephalitis condition detection model combining spatiotemporal features

Abstract Background Pediatric viral encephalitis is characterized by sudden onset and rapid progression, and accurate early assessment of disease severity is essential for preventing severe complications and long-term neurological sequelae. However, conventional EEG-based assessment relies heavily on manual interpretation, which is subjective, time-consuming, and may limit its utility for early clinical screening. To address these challenges, we propose a TFG–CBNN (time–frequency fusion and GAF-encoded CNN–BiLSTM network) model for the automated classification of encephalitis severity in children. The proposed framework integrates time–frequency feature fusion, Gramian Angular Field (GAF) encoding, and CNN–BiLSTM classification into a unified pipeline. Specifically, multichannel EEG signals are fused with multiband frequency-domain features and encoded into a sequence of GAF images, thereby preserving temporal information, amplitude relationships, and cross-channel dependencies. A CNN extracts spatial features from the GAF images, while a BiLSTM models temporal dependencies across the image sequence. Results Experimental results show that TFG–CBNN achieves an accuracy of 97.04% on EEG recordings from 20 pediatric patients and outperforms the baseline models, including MobileNet, ResNet50, CNN–LSTM, EfficientNet, and ConvNeXt, in accuracy, precision, and F1 score. Conclusions These findings demonstrate the potential of the proposed framework for computer-aided severity assessment and clinical decision support in pediatric viral encephalitis.