Real-time classification of neural signal from motor cortex through multiple recording sessions

The development of Brain Computer Interfaces (BCIs) provides a way for the brain to interface directly with external devices. In recent years, BCIs could translate signals generated by brain activity into control signals without the involvement of peripheral nerves and muscles to restore lost sensorimotors function, for example caused by spinal cord injury or neurodegeneration. For this purpose, the barrier to the adoption of the invasive BCIs is getting lower thanks to progress in neuroscience field, to smaller and more performing electrode arrays and to advances in AI field. The combination of these elements is reflected in the B-Cratos project, a closed-loop neural interface that exploits the intra-cranial EEG and advanced Deep Learning algorithms to reach fine control of a prosthetic arm that, at the same time, stimulates sensory feedback response in the patient brain. The first goal of this work is to present a Deep Learning model, starting from a dataset collected by researchers of DPZ (Deutsches Primatenzentrum), that can classify signals recorded via implanted microelectrode arrays. The dataset is based on the brain activity of two purpose-bred macaque monkeys that were trained in a delayed grasping task with a wide range of different objects. The classification consists in two steps: a first binary classifier that detects movement activation, a second multiclass classifier that identifies the graspable objects, grouped in 48 classes. Bidirectional Recurrent Neural Network (BRNN) models shows higher performance, in particular in the multiclass task with an accuracy of up to 72%, representing a significant advancement over previously published results. The second goal is to transfer the BRNN model on other datasets of the same animal but recorded in different days. This is needed to make the classification model robust to the evolution of the neural signal due to mechanical movement of the implant, inflammatory processes, brain plasticity etc. Before this, dimensionality reduction algorithms have been investigated in order to have the same number of electrode channels in each dataset. Principal Component analysis (PCA) and Linear Discriminant analysis (LDA) have been tested on the different datasets. After a little pre-training of the BRNN model on a reducted dataset, an accuracy around 40-50% is reached. The overall results are promising, showing that the different objects and trial phases shape the neural activity in such a unique way that makes it possible to classify them, on the other hand, it is yet not possible to demonstrate significant results in transfer learning through time.