A Low-Complexity Combined Encoder-LSTM-Attention Networks for EEG-based Depression Detection

Despite the high performance of existing state-of-the-art deep learning models for depression detection using electroencephalography (EEG), they incur a heavy computational burden. In this paper, we propose an efficient model consisting of a cascade of an encoder, long short-term memory (LSTM), and attention mechanism networks. The encoder compresses data into a lower-dimensional latent space. The LSTM models the temporal variations in brain rhythms. The attention mechanism rectifies the problem of compressed data in sequence-to-sequence models and efficiently leverages parallelism. Compared with recent state-of-the-art, our proposed depression detection model shows better performance and efficiency with a validation accuracy of 99.57% on subject-dependent experiment and a testing accuracy of 84.93% on subject-independent experiment with a total number of 4,355 parameters. The proposed model has resulted in 99.65% reduction in complexity compared with the state-of-the-art EEG-based depression detection models. The results of this study indicate the effectiveness of the proposed model design and the usefulness of the combined encoder, LSTM, and attention modules. These networks serve as mitigating factors for the computational load, which is vital for future research on multi-tasking mental health monitoring using AI-enabled EEG wearables.