Toward Interpretable Sleep Stage Classification Using Cross-Modal Transformers.

Accurate sleep stage classification is significant for sleep health assessment. In recent years, several machine-learning based sleep staging algorithms have been developed, and in particular, deep-learning based algorithms have achieved performance on par with human annotation. Despite improved performance, a limitation of most deep-learning based algorithms is their black-box behavior, which have limited their use in clinical settings. Here, we propose a cross-modal transformer, which is a transformer-based method for sleep stage classification. The proposed cross-modal transformer consists of a cross-modal transformer encoder architecture along with a multi-scale one-dimensional convolutional neural network for automatic representation learning. The performance of our method is on-par with the state-of-the-art methods and eliminates the black-box behavior of deep-learning models by utilizing the interpretability aspect of the attention modules. Furthermore, our method provides considerable reductions in the number of parameters and training time compared to the state-of-the-art methods. Our code is available at https://github.com/Jathurshan0330/Cross-Modal-Transformer. A demo of our work can be found at https://bit.ly/Cross_modal_transformer_demo.

Decoding of Hand Gestures from Electrocorticography with LSTM Based Deep Neural Network.

Hand gesture decoding is a key component of controlling prosthesis in the area of Brain Computer Interface (BCI). This study is concerned with classification of hand gestures, based on Electrocorticography (ECoG) recordings. Recent studies have utilized the temporal information in ECoG signals for robust hand gesture decoding. In our preliminary analysis on ECoG recordings of hand gestures, we observed different power variations in six frequency bands ranging from 4 to 200 Hz. Therefore, the current trend of including temporal information in the classifier was extended to provide equal importance to power variations in each of these frequency bands. Statistical and Principal Component Analysis (PCA) based feature reduction was implemented for each frequency band separately, and classification was performed with a Long Short-Term Memory (LSTM) based neural network to utilize both temporal and spatial information of each frequency band. The proposed architecture along with each feature reduction method was tested on ECoG recordings of five finger flexions performed by seven subjects from the publicly available 'fingerflex' dataset. An average classification accuracy of 82.4% was achieved with the statistical based channel selection method which is an improvement compared to state-of-the-art methods.