Development of Brain Computer Interface Systems and Machine Learning-Based Analysis of Brain Consciousness States
Year: 2026
Type: dissertation
Abstract: This thesis presents the development of a Brain-Computer Interface (BCI) system designed to facilitate daily-life interactions for individuals with physical disabilities, alongside a machine-learning-based analytical framework for identifying consciousness states in patients with disorders of consciousness.First, addressing the limitations of conventional BCI systems-such as delays in external stimulus responses and the lack of seamless, real-time continuous control-we designed a novel signal processing method for the online identification of specific neural oscillatory patterns with high accuracy and low latency.(1) To assist in communication and emergency calling, we integrated Canonical Correlation Analysis (CCA) with a sliding window approach, enabling real-time analysis of steady-state visual evoked potentials (SSVEPs).This led to the development of a brain-controlled mobile phone dialing system.Online experiments involving five participants (N=420 trials) demonstrated a high average classification accuracy of 98.1%.(2) To enhance the quality of daily life, we introduced a signal enhancement technique based on ensemble averaging, which formed the core of a brain-controlled robotic arm operation system.Validation through online experiments with five participants (N=500 trials) confirmed a robust average classification accuracy of 81.2%.(3) Building upon these core technologies, we further integrated multiple BCI paradigms within an Internet of Things (IoT) framework to achieve unified environmental control.The resulting integrated BCI system provides comprehensive support for mobility, communication, and smart home interaction for individuals with physical disabilities.Next, we developed machine learning methodologies for the analysis of electrophysiological signatures associated with states of consciousness, including sleep, coma, and brain death.(1) For automatic sleep stage classification, we proposed a feature extraction method combining Turning Tangent Empirical Mode Decomposition (2T-EMD) with Multitaper Power Spectral Density (MT-PSD).A II Random Forest (RF) classifier was subsequently implemented in an automatic deep sleep (N3 stage) recognition system.Evaluation on data from 41 subjects confirmed a high classification accuracy of 95.6% for the N3 stage.(2) To enhance generalizability, we developed a personalized system optimized for individual users, which further improved the classification of sleep-related brain dynamics.Furthermore, by integrating with a portable BCI device, we realized a system capable of real-time deep sleep detection and quantitative sleep quality assessment.(3) For the automatic discrimination of pathological consciousness states (e.g., coma and brain death), we employed the Variational Mode Decomposition (VMD) algorithm to decompose EEG signals into intrinsic mode functions, which were then classified using a machine learning model.Tests on EEG datasets from eight patients achieved a remarkable classification accuracy of 99.59%.In summary, this research establishes the practical efficacy of an integrated BCI system for supporting daily activities in individuals with physical disabilities, and a validated, EEG-based machine learning framework for the precise identification of consciousness states in clinical populations.
Language: English
Subfield: Cognitive Neuroscience
Field: Neuroscience
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7205037382