Consciousness, a fundamental aspect of human experience, is often understood as comprising two key elements: arousal (wakefulness) and awareness (subjective experience). While various clinical measures exist to assess responsiveness, directly quantifying the distinct contributions of arousal and awareness to consciousness remains challenging. This lack of precise neurophysiological metrics hinders our understanding of the neural mechanisms underlying different states of consciousness, including those observed during sleep, under anesthesia, and in patients with disorders of consciousness (DoC), such as unresponsive wakefulness syndrome (UWS) and minimally conscious state (MCS). Existing indices like the perturbational complexity index (PCI) provide valuable insights but have limitations, particularly in differentiating between states with varying levels of awareness, such as REM sleep and ketamine-induced anesthesia, both of which exhibit low arousal but differing levels of awareness. This research aims to address this gap by developing a novel, interpretable metric capable of simultaneously quantifying arousal and awareness in diverse states of consciousness. The use of deep learning techniques, along with techniques for interpreting the models' decisions, will allow for a more nuanced understanding of the neural correlates of consciousness.

Previous research has attempted to quantify consciousness using various neurophysiological approaches. The perturbational complexity index (PCI), derived from EEG responses to transcranial magnetic stimulation (TMS), has been successful in distinguishing between conscious and unconscious states. However, PCI cannot differentiate between REM sleep or ketamine-induced anesthesia and healthy wakefulness, and it requires multiple trials. Other studies have explored the use of resting-state EEG features, such as the spectral exponent, to differentiate between anesthetic states. Although these measures show promise, they often lack the ability to simultaneously quantify both arousal and awareness. This study builds upon these previous efforts, incorporating the strengths of deep learning models to provide a more comprehensive and interpretable measure of consciousness that addresses the shortcomings of existing methods.

The researchers developed an Explainable Consciousness Indicator (ECI) using a convolutional neural network (CNN) to analyze EEG data. Three datasets were used: sleep (NREM and REM), general anesthesia (ketamine, propofol, xenon), and disorders of consciousness (UWS and MCS). TMS-EEG data were acquired, with TMS applied over various brain regions in each condition. Resting-state EEG data were also collected under anesthesia and in DoC patients. The CNN was trained to classify EEG data into high/low arousal and high/low awareness states using a leave-one-participant-out (LOPO) cross-validation method, which is a form of transfer learning that trains the model on data from most participants to test on the remaining participant. Layer-wise Relevance Propagation (LRP) was then applied to interpret the CNN's decisions, identifying which brain regions contributed most to the classification. The accuracy of the CNN was compared with that of linear discriminant analysis (LDA) and support vector machines (SVM). The relationship between ECI and PCI was also explored.

The ECI successfully distinguished between different levels of arousal and awareness across various states of consciousness. In sleep, NREM sleep showed low arousal and awareness, while REM sleep showed low arousal but high awareness, and wakefulness showed high arousal and awareness. Under general anesthesia, ketamine induced low arousal and high awareness, while propofol and xenon induced low arousal and low awareness. In DoC patients, MCS patients exhibited high arousal and high awareness, whereas UWS patients showed high arousal and low awareness. The ECI showed a strong positive correlation with PCI, validating its effectiveness. The CNN model demonstrated superior performance compared to LDA and SVM in classifying arousal and awareness. Using LRP, the researchers found that parietal regions were most important in classifying the EEG data into the different states, suggesting a key role of the parietal cortex in the neural correlates of consciousness. The ECI performed well even with resting-state EEG data, demonstrating its robustness and applicability in clinical settings.

The results of this study demonstrate that the ECI is a robust and reliable neurophysiological indicator capable of simultaneously quantifying arousal and awareness across a range of conscious states. The ability to distinguish between states like REM sleep and ketamine-induced anesthesia highlights the ECI's advantage over previous measures. The strong correlation with PCI adds to its validity, confirming that the deep learning approach captures relevant information about the neural correlates of consciousness. The identification of the parietal cortex as a key brain region involved in consciousness aligns with previous research and provides further insight into the neural underpinnings of conscious experience. The ECI's effectiveness with resting-state EEG suggests its practical utility in clinical settings where TMS may not be feasible.

The ECI provides a novel and valuable tool for assessing levels of consciousness in diverse clinical settings. Its ability to distinguish between arousal and awareness, its strong correlation with PCI, its performance with resting-state EEG, and its interpretability through LRP make it a significant advancement in the field. Future research could focus on validating the ECI in larger, more diverse patient populations and exploring its real-time application for continuous monitoring of consciousness.

The study's sample size was relatively small, limiting the generalizability of the findings. Although the ECI was shown to perform well with few trials, real-time application was not tested. Although efforts were made to minimize artifacts, sensory stimulation from the TMS could have influenced the results. The ECI distinguishes between high and low states but doesn't provide a continuous measure or functional assessment of consciousness.