Evaluation of Transformer-Based Models in Optimizing Invasive and Non-Invasive Brain-Computer Interfaces: Recurrent Neural Networks to Enhance Communication Speed for Locked-In Syndrome Patients

Brain-Computer Interfaces (BCIs) have been proposed as assistive technologies for Locked-In Syndrome (LIS) patients that can facilitate communication based on decoding of neural signals. Traditional BCI systems based on recurrent neural network (RNN) models exhibit certain constraints in terms of decoding accuracy, communication speed, and response latency. The current study aims to assess the effectiveness of transformer-based frameworks in optimizing the efficiency of both invasive and non-invasive BCI systems as compared to classical RNN models. A computational-clinical study design was used which involved participation of 48 LIS or severely paralysed participants. Subjects were grouped in accordance with their involvement in invasive or non-invasive BCI groups, and assessments were conducted during a period of eight weeks of intervention. Neural activity data processing was done with the help of two different approaches, including transformer-based model application and RNN application, assessing communication speed, decoding accuracy, latency, and error rates of both systems. Results suggest that transformer-based neural decoding frameworks proved to be superior to RNNs in terms of all evaluated criteria. Invasive transformer-based BCI demonstrated the best results concerning communication speed, decoding accuracy, lowest latency, and lowest error rates. Non-invasive transformer BCIs also yielded better results than RNN-based BCIs.