A hybrid autoencoder framework of dimensionality reduction for brain-computer interface decoding

As the scale of neural recording increases, Brain-computer interfaces (BCIs) are restrained by high-dimensional neural features, so dimensionality reduction is required as a preprocess of neural features. In this context, we propose a novel framework based on deep learning to reduce the dimensionality of neural features that are typically extracted from electrocorticography (ECoG) or local field potential (LFP). A high-performance autoencoder was implemented by chaining convolutional layers to deal with spatial and frequency dimensions with bottleneck long short-term memory (LSTM) layers to deal with the temporal dimension of the features. Furthermore, this autoencoder is combined with a fully connected layer to regularize the training. By applying the proposed method to two different datasets, we found that this dimensionality reduction method largely outperforms kernel principal component analysis (KPCA), partial least square (PLS), preferential subspace identification (PSID), and latent factor analysis via dynamical systems (LFADS). Besides, the new features obtained by our method can be applied to various BCI decoders, without significant differences in decoding performance. A novel method is proposed as a reliable tool for efficient dimensionality reduction of neural signals. Its high performance and robustness are promising to enhance the decoding accuracy and long-term stability of online BCI systems based on large-scale neural recordings.