Fig 1.
Distribution of 30s EEG epochs in sedation and sleep datasets used in this study.
Similar number of MOAA/S epochs were present in all six channels in the sedation dataset.
Fig 2.
Illustration of the proposed artificial intelligence framework to predict sleep stages in sedation dataset.
The raw EEG signal was first filtered and segmented into 30s epochs. After identifying the optimal model to classify different sleep stages on the MGH sleep dataset, it was later used to predict sleep stages in the UMCG sedation dataset. One-to-one comparison was then made between the MOAA/S scores and predicted sleep stages. This analysis was performed separately across all six channels.
Fig 3.
Illustration of different training and testing combinations for sleep stage predictions and correlation with sedation states developed in this study.
The machine learning model was trained on sleep data to predict sleep stages on the sedation data which were then correlated with sedation states.
Table 1.
Model type, hyperparameters, grid search range and the final best performing parameters obtained during training the machine learning models to predict sleep stages.
Table 2.
The performance of the machine learning algorithms to differentiate between W and other sleep stages during three and five class binary classification tasks on the test set sleep stages during three and five class binary classification tasks on the test set.
Fig 4.
Heat map of the weights assigned to the 44 features by the random forest algorithm across all channels in the sleep data during training.
The weights of the features with more discriminatory information are indicated with high intensity color (normalized to 1). Features 12, 21,31, 41 and 44 (burst suppression ratio, normalized alpha power with respect to the total power, normalized band powers with respect to the theta band power, Renyi entropy, and fractal dimension, respectively) were the top five discriminatory features.
Fig 5.
Heatmap of the performance of the random forest algorithm to differentiate between awake (MOAA/S score = 5) and different levels of sedation using sleep staging models for all three drugs.
The colorbar represents the AUC values obtained by the model. Here WN = trained on wake (W) and nonrapid eye movement (N) sleep stages; WR = trained on W and rapid eye movement R; WN1 = trained on W and N1; WN2 = trained on W and N2; WN3 = trained on W and N3; MOAA/S = Modified Observer’s Assessment of Alertness/Sedation (MOAA/S) scale; M54 = MOAA/S 5 versus MOAA/S 4; M53 = MOAA/S 5 versus MOAA/S = 3; M52 = MOAA/S 5 versus MOAA/S = 2; M51 = MOAA/S 5 versus MOAA/S 1; M50 = MOAA/S = 5 versus MOAA/S = 0.
Table 3.
The performance of the random forest algorithm to differentiate between awake (MOAA/S score = 5) and different levels of sedation using sleep staging models for all three drugs.