Spatio-temporal representation learning with selective state space models for EEG-based depression detection

Abstract

Electroencephalogram signals, as key cognitive biomarrs, capture subtle brain activities and are crucial for diagnosing mental disorders like depression. However, owing to the multi-channel and high sampling rate of acquisition devices, EEG data exhibit high dimensionality and long sequences. Most existing studies focus on analyzing individual domains, such as temporal, structural, or state features, making it challenging to effectively capture and represent the correlations among these features across multiple channels while avoiding the loss of key information in long sequential data. In addition, variations in patients’ conditions result in differences in detection times, further increasing the complexity of data processing. To tackle these challenges, we propose the TSS-SSM framework, which combines spatio-temporal representation learning of temporal, structural, and state correlations with selective state-space model to effectively handle the complex features of EEG signals. First, by segmenting EEG signals into adaptive time slices and using multiple GCNs, we effectively extracted structural relationships between brain regions. The integration of LSTM networks and attention mechanisms enabled us to model the historical states of EEG segments and retain critical information from past states in continuous time sequences. Then, by integrating SSM and a selection mechanism, our model highlights important brain activity events and prevents them from being overlooked in long sequences. Experimental results on the public MODMA dataset and the real-world dataset from Xiangya Hospital demonstrate that TSS-SSM achieved significant performance improvements, with ACC values of 0.9481 and 0.8836, respectively, and its effectiveness was further validated through extensive ablation studies.