Investigating the Trade-off Between Model-based and Deep Learning Decoders for Brain-machine Interfaces

Severely paralyzed patients encounter major difficulties with daily tasks, restricting their independence and quality of life. Over the last few decades, brain–machine interfaces (BMIs) have emerged as game-changing paradigms, enabling patients to voluntarily control external devices (e.g., robotic arms) through neural signals. Among the key components of BMIs, the decoder, which interprets user intentions from neural activity, has significantly improved with deep learning. These deep-learning-based decoders have demonstrated state-of-the-art performance in BMI control. However, due to their complex model architecture, they typically require greater computational resources for training. This can pose a significant challenge for practical deployment, particularly in real-world BMI applications where computational constraints and environmental variability may differ from controlled laboratory settings. Models such as the Kalman filter (KF) are still popular in clinical and assistive applications due to their comparable performance to deep learning methods while requiring much fewer computational resources. In this study, we compared three decoding approaches using electrocorticography (ECoG) data from patients: the multiple linear regression model; the KF, which incorporates temporal dynamics; and a long short-term memory (LSTM) based recurrent deep-learning model. This comparison highlights the trade-off between computational cost and performance in real-world ECoG BMI applications.