Biomedical Engineering - Theses
Permanent URI for this collection
Search Results
1 - 3 of 3
-
ItemInvestigation of Speech Imagery for Brain-Computer Interfaces( 2023-06)Brain-computer interfaces can restore various forms of communication in paralyzed people who have lost their ability to articulate intelligible speech. These devices, known as speech neuroprostheses, serve as direct interfaces between living brains and artificial electronic components. The field of research is rapidly moving toward clinical trials with potential users benefiting from chronic brain implants. However, research with non-target participants is essential to accelerate the development of such devices. This thesis investigated the use of speech imagery tasks in study participants implanted with intracranial electrodes to improve the performance of speech neuroprostheses. Four original studies were conducted to address practical considerations related to: (1) the identification of discriminative features for keyword detection from brain recordings, (2) the characterization of brain activation patterns during the production of isolated speech sounds, (3) the implementation of a closed-loop algorithm for real-time speech synthesis from brain recordings, and (4) the importance of brain coverage and task instructions to synthesize intelligible artificial sounds during silent speech. In all four studies, patients with medication-resistant epilepsy were temporarily implanted with intracranial electrodes, either stereotactic electroencephalography (SEEG) depth or electrocorticography (ECoG) surface electrode arrays, for the localization of the seizure onset zone prior to brain resection. They were asked to perform overt and silent speech tasks while their brain signals were recorded. The first study adopted a traditional decoding approach based on discrete trial classification and found that high-gamma activation in the superior temporal gyrus (STG) was the most discriminative feature for keyword detection during overt speech. No discriminative feature was found during imagined speech, which highlighted the need to go beyond trial-based designs. The second study introduced a voice-based cursor control task through the production of isolated phonemes. Onset and sustained neural activation patterns were observed in the STG but trajectory reconstructions from intracranial signals remained low. The mismatch between the control task and visual feedback also constituted an important challenge for restoring intuitive communication. The third study characterized the trade-off between decoding performance and execution times in the proposed algorithm for real-time speech synthesis from intracranial recordings. Purely based on acoustic features of speech, the model could be trained with a small patient-specific dataset and immediately tested in any language with no assumption on brain coverage. The fourth study tested the proposed closed-loop system in ten participants with various coverage of brain areas. Artificial sounds were synthesized from the STG during overt speech in three participants and from the precentral gyrus during mimed speech in two participants. Human perceptual judgments supported the fact that some of these sounds were occasionally intelligible. Unfortunately, no artificial sounds were synthesized during imagined speech. These four studies made specific contributions to the field of research by going beyond trial-based design, understanding the gap between voice and brain control, synthesizing artificial sounds in real time under clinical constraints, and rethinking the utility of mimed and imagined speech in able-bodied study participants. Altogether these four studies added further evidence toward the feasibility of closed-loop speech neuroprostheses that continuously synthesize intelligible artificial sounds. Such brain-to-audio systems have the potential to restore functional communication at a conversational rate of 150 words per minute, outperforming state-of-the-art brain-to-text systems that operate at 62 words per minute. This thesis concludes with a thorough discussion on the limitations of the proposed brain-to-audio approach and future directions to overcome the remaining barriers to clinical translation in target patients.
-
ItemDecoding Sensorimotor Rhythms for Brain-Computer Interfaces( 2021)Brain-computer interfaces (BCIs) have great potential to improve the quality of life for people with severe motor disabilities. By measuring and interpreting neural activity, BCIs can predict and express intention through an external computerised device. This creates an alternative mechanism for communication and control for people with paralysis. Volitional changes in oscillatory activity near the sensorimotor cortex, known as sensorimotor rhythms (SMRs), can be measured with invasive techniques, such as electrocorticography (ECoG), or with non-invasive techniques, such as electroencephalography (EEG). This work explored novel approaches for decoding SMRs from EEG with the aim of utilising neurophysiological principles. This thesis also investigated the viability of vascular electrocorticography (vECoG) as a SMR-based BCI modality. The common spatial patterns (CSP) algorithm is used to linearly combine information from multiple EEG electrodes in order to accentuate SMR activity. Typical patterns of SMR activity derived by this method were characterised using publicly available EEG datasets. It was found that both neurophysiologically probable and improbable patterns of activity were commonly extracted and that selecting for, and adapting to, neurophysiologically probable activity could improve decoding performance. These findings highlight the importance of considering spatial filter adaptation in EEG BCI decoder design. The range of decoding algorithms available to BCI practitioners is extensive and diverse. A universal method for explaining the predictions of EEG decoders in terms of neurophysiologically relevant factors was investigated. The validity of the explanations was demonstrated using simulated EEG data. The method was also employed to compare the behaviour of four categories of decoders using real, pre-recorded EEG data. The results indicated that all decoders were able to harness neurophysiologically plausible electrodes and cortical sources to make accurate predictions. However, the influence of artifactual activity was also found to contribute to high decoder accuracy. These findings emphasise the need to understand the predictive behaviour of decoders and the proposed method may be useful a tool to help BCI researchers achieve this understanding. Vascular electrocorticography (vECoG) measures intracranial neural activity by chronically implanting a stent-electrode array within the brain vasculature. By omitting the need to penetrate the skull, this minimally invasive technique has the potential to safely record high fidelity neural information and be used as a long-term, clinically useful BCI. Data from the first-in-human clinical trial of the Stentrode was used to characterise vECoG signal quality. Participant-specific re-referencing was shown to improve signal quality and the maximum bandwidth of the signal was found to be consistent with previous animal studies. A multiclass, online SMR decoder was also implemented and tested with a single participant. Discriminatory activity from multiple motor imagery classes could be observed, however, signal non-stationarity affected online decoding performance. Together, the signal quality and multiclass decoding results suggest that vECoG has significant potential as a recording modality for a clinically useful BCI. Overall, the findings presented in this thesis contribute to two key facets of SMR BCI research. In terms of decoding SMRs from EEG, they demonstrate the need to harness neurophysiological phenomena, avoid contamination from artifactual activity and improve the interpretability of complex decoding algorithms. Furthermore, the vECoG results add to the growing body of evidence implicating this modality as a beneficial way for interfacing with the brain in a minimally invasive manner.
-
ItemDecoding upper-limb kinematics from electrocorticography( 2017)Brain-computer interfaces (BCIs) are technologies for assisting individuals with motor impairments. Activity from the brain is recorded and then processed by a computer to control assistive devices. The prominent method for recording neural activity uses microelectrodes that penetrate the cortex to record from localized populations of neurons. This causes a severe inflammatory response, making this method unsuitable after approximately 1-2 years. Electrocorticography (ECoG), a method of recording potentials from the cortical surface, is a prudent alternative that shows promise as the basis of a clinically viable BCI. This thesis investigates aspects of ECoG relevant to the translation of BCI devices: signal longevity, motor information encoding, and decoding intended movement. Data was assessed from a first-in-human ECoG device trial to quantify changes in ECoG over multiple years. The mean power, calculated daily, was steady for all patients. It was demonstrated that the device could consistently record ECoG signal statistically distinct from noise up to approximately 100 Hz for the duration of the study. Therefore, long-term implanted ECoG can be expected to record movement-related high-gamma signals from humans for many years without deterioration of signal. ECoG was recorded from patients undertaking a two-dimensional center-out task. This data was used to generate encoder-decoder directional tuning models to describe and predict arm movement direction from ECoG. All four patients demonstrated channels that were significantly tuned to the direction of motion. Significant tuning was found across the cortex and was not focused on primary motor areas. Decoding significantly above chance with a population-vector approach was achieved in three of the four patients. Decoding accuracy was significantly improved by weighting the population vector by each channel's tuning signal-to-noise ratio. Hence, directional tuning exists in high-frequency ECoG during movement preparation, and movement angle can be decoded using population vector methods. Having confirmed the existence of direction-related information in the recorded data, artificial neural network models were created to decode intended movement direction. A convolutional neural network (CNN) model had significantly higher decoding accuracy than a fully connected model for all four patients for decoding movement direction. Training models on data from all patients and testing on a single patient improved decoding performance for all but the best performing patient with the CNN model. Decoding using data from multiple time-points with a CNN model and averaging the results boosted accuracy when using the mode of the outputs. Overall, it was demonstrated that artificial neural network models can decode intended movement direction from ECoG recordings of a two-dimensional center-out task. This thesis presents results that demonstrate ECoG has the desired signal properties for a clinically-relevant BCI. ECoG is shown to be robust over multiple years, encode direction-related information and can be decoded with high accuracy.