Enhancement of Readiness Potentials’ pre-processing chain in a Brain-Computer Interface for nonresponsive patients

Over the last years, an increasing number of patients survive severe brain injury thanks to recent improvements in intensive care. Nevertheless, the recovery of brain capacities and motor skills is not always complete: it may happen that the patient maintains vital functions but does not show overt awareness of himself and the environment around them. From the diagnostic point of view, behavioral assessments can provide general and summary information on the patient’s state of consciousness, based on evidence found in attitudes and responses, voluntary or not. Despite this, such type of assessment can lead to a relevant number of misdiagnoses or inaccurate diagnoses, depending on the severity of brain injury, which may reveal incomplete or inconsistent behaviors. The challenge is to design a low-cost non-invasive assessment method that, starting from electroencephalographic (EEG) and electromyographic (EMG) signals, can detect and distinguish different levels of consciousness. This method relies on a detailed pre-processing chain, which allows to clean up raw signals and bring to light the real information contained in the biopotential. This information is then fed to the machine learning algorithms that allow classification, useful for differential diagnosis. The following thesis focuses on some steps of the pre-processing chain, in order to improve the quality of the obtained Readiness Potential (RP), allowing a more accurate classification. In particular, the covered topics are: (1) set-up of temporal filtering parameters, in order to extract from raw EEG signal the frequency bands which contain useful information; (2) implementation of a novel method for jitter compensation based on Residue Iteration Decomposition (RIDE), an algorithm of iterative subtraction which separates different clusters of ERP’s components according to their time-locking to stimulus onset, response times, or estimated latencies and reconstructs ERPs by re-aligning the component clusters to their most probable trial latencies; (3) comparison of RIDE method with the previous Woody’s method and testing alternative methods which combine both techniques for jitter compensation, relying on the strengths and weaknesses of each of them; (4) analysis of event-related potentials from patients with hemiplegia in order to assess the robustness of the pre-processing chain.