Table 2.

Summary of deep learning architectures used in EEG-based healthcare.

Architecture	Key features	Complex	Input latitude	Typical applications	Strengths	Limitations	Representative studies
CNN	Spatial feature extraction, convolution layers	Low–Med	High rep., Med length/channels, Low topology.	Seizure detection, sleep staging, BCI	Captures local spatial patterns, less manual feature engineering	Needs large datasets, weak temporal modeling	Acharya et al. (2018) and Supratak et al. (2017)
RNN/ LSTM	Sequential modeling, temporal dependencies	Med–High	Med rep., High length, Low–Med channels, Low topology.	Emotion recognition, seizure prediction	Models long-term dependencies	Sensitive to data imbalance, slower training	Tripathi et al. (2017)
Transformer	Self-attention, parallel processing	High	High rep./length, Med channels, Low topology	Cognitive workload estimation, emotion recognition	Captures global dependencies, interpretable attention	Computationally heavy, needs large data	Klein et al. (2025)
CNN–LSTM Hybrid	Combines spatial + temporal modeling	Med–High	High rep./length, Med channels, Low topology	Seizure detection, sleep staging	Joint spatio-temporal learning	High computational cost	Roy et al. (2019)
GNN	Graph-based electrode modeling	Med–High	Med rep./length, High channels/topology	Motor imagery, emotion recognition	Captures inter-channel connectivity	Graph construction complexity	Xu et al. (2024)