EEG provides valuable biological information for BCI tasks. In the process of studying EEG, many complex preprocessing methods have been generated. These methods require researchers to manually select and process data and features, which takes a long time and has high professional requirements and is highly subjective. This paper proposes the TwoM model with a spatiotemporal filtering mechanism, which adaptively filters effective data by learning spatial and temporal features, addressing the overfitting problem caused by small samples and low signal-to-noise ratios. TwoM has been successfully validated on three EEG-based depression datasets, including MODMA, achieving strong results without preprocessing, demonstrating the model’s robustness.