Abstract: Objective To identify and detect brain functional characteristics related to consciousness fluctuations under the critical state with propofol anesthesia by confident learning and then use them to classify consciousness states. Methods Atotal of 29 patients who underwent thoracic and abdominal surgery in the First Affiliated Hospital of Xi'an Jiaotong University from October 2019 to October 2020 were selected. They were aged 18‒45 years, BMI 18‒26 kg/m2, American Society of Anesthesiologists (ASA) class Ⅰ to Ⅱ, without neurological and psychiatric diseases or history. Propofol was administered using a target‑controlled infusion system. The initial drug concentration of propofol was set to 1.0 mg/L, which then increased with a step‑size of 0.2 mg/L at every 6 min until unconsciousness was achieved. Meanwhile, the patients were requested to press buttons to capture their consciousness states and collect their electroencephalography (EEG) data. Then, the collected EEG data were preprocessed and divided into 5 s epochs according to the sound stimulation. EEG epochs were assigned to consciousness or unconsciousness according to behavioral response following each stimulus. The state when patient's consciousness states were fluctuated was defined as a critical state. Totally 110 brain functional characteristics, including power spectrum of EEG signal, signal complexity, interregional functional connectivity and brain network properties, were extracted for each EEG epoch. The forward selection method was employed for feature screening. Then, confident learning was used to clear training sets. Finally, three machine learning algorithm models including linear discriminant analysis (LDA), logistic regression (LR), and support vector machine (SVM) were used to calculate the accuracy rate of consciousness state classification before and after confident learning. Results Most labels cleared by confident learning were located in the critical state of consciousness fluctuations. The classification accuracy rate of LDA, LR and SVM classification models for consciousness state before confident learning was (85.3±3.7)%, (85.3±3.9)%, and (85.3±3.8)%, respectively, and the classification accuracy rate after confident learning was (93.5±2.0)%, (92.9±1.8)%, and (93.3±1.0)%, respectively, with an average increase of 7.93% compared to those before confident learning. The posterior α‑power, frontal‑parietal PLI‑δ, and clustering coefficient‑δ of the epochs that were initially labeled as unconsciousness (without responding to auditory stimuli) but re‑labeled as consciousness by confident learning were significantly different from stable unconsciousness state (P<0.001). At the same time, there was no significant difference compared with steady consciousness state (P>0.05). Although there were statistical significant differences between indeterminate consciousness state and stable consciousness states and table unconsciousness state in the fast and slow wave‑α‑frontal, permutation entropy‑θ‑frontal, permutation entropy‑θ‑central, and α‑proportional‑frontal (P<0.001), these data were significantly closer to stable conscious state than stable unconsciousness state. Conclusion The use of confident learning effectively improves the classification of different consciousness states in the critical state, which provides methodology support for more accurate intraoperative awareness monitoring.
|