EEG-based methods for diagnosing awareness in disorders of consciousness

A brain–computer interface (BCI) is an advanced neurotechnological system that enables direct communication between the brain and the external environment by bypassing conventional neuromuscular pathways. This capability offers valuable insight into the assessment of awareness in acute clinical states. Clinically, diagnosing disorders of consciousness (DOC) remains a significant challenge, largely because current practice relies heavily on behavioral indicators of consciousness —markers that are often ambiguous and prone to misinterpretation. To address these limitations, electrophysiological and neuroimaging techniques have been explored, with electroencephalography (EEG) standing out for its non-invasiveness, portability, high temporal resolution, and robustness. As a result, EEG-based methods and BCI-inspired protocols have emerged as promising tools for improving the diagnosis and prognosis of DOC, particularly in detecting cognitive motor dissociation (CMD), a condition frequently overlooked by standard clinical scales. Despite this promise, the clinical translation of these approaches remains constrained, primarily due to a shortage of sufficiently powered validation studies. In this thesis, I evaluate the effectiveness of several popular EEG-based methods and systematically compare the performance of deep and shallow classification models on a large, novel dataset acquired using a motor imagery (MI)-based command-following paradigm in accessing awareness with DOC patients. Specifically, I extracted measures including classification accuracy, brain rhythms, effective connectivity and the perturbational complexity index PCI, from MI, idling and functional electrical stimulation FES epochs. These were contrasted against Coma Recovery Scale–Revised (CRS-R) scores, the current clinical gold standard. Furthermore, state-of-the-art deep learning models (EEGNet, DeepConvNet, and EEGConformer) were evaluated alongside a shallow classifier, employing leave-one-trial-out cross validation scheme on the full and windowed trial segments. In addition, analysis was evaluated at the sessional level to account for variability in diagnostic states. The findings confirm that EEG contain valuable information regarding the state of awareness of DOC patients. In particular, the classification accuracy and the μ-/β-band separability of MI power spectral density(PSD) features, as well as centro-parietal δ- band connectivity during MI and resting, correlate statistically significantly with CRS-R. Moreover, metric-specific thresholds separating awareness from non-awareness could be determined. I further provide useful insights on the ability of these metrics to detect CMD and rectify the false-negative vulnerability of CRS-R. At the same time, this work highlights the risk of statistical misuse of such metrics, which can lead to over-optimistic assessments of latent awareness. Furthermore, the thesis also reveals that deep learning architectures may be prone to overestimation of results when applied to DOC populations. This research supports the potential of open-loop BCI DOC diagnosis and highlights the need for further development, validation and standardization to establish clinically deployable systems.