Deep learning models using intracranial and scalp EEG for predicting sedation level during emergence from anaesthesia

Abstract

Background

Maintaining an appropriate depth of anaesthesia is important for avoiding adverse effects from undermedication or overmedication during surgery. Electroencephalography (EEG) has become increasingly used to achieve this balance. Investigating the predictive power of intracranial EEG (iEEG) and scalp EEG for different levels of sedation could increase the utility of EEG monitoring.

Methods

Simultaneous iEEG, scalp EEG, and Observer's Assessment of Alertness/Sedation (OAA/S) scores were recorded during emergence from anaesthesia in seven patients undergoing placement of intracranial electrodes for medically refractory epilepsy. A deep learning model was constructed to predict an OAA/S score of 0–2 vs 3–5 using iEEG, scalp EEG, and their combination. An additional five patients with only scalp EEG data were used for independent validation. Models were evaluated using the area under the receiver-operating characteristic curve (AUC).

Results

Combining scalp EEG and iEEG yielded significantly better prediction (AUC=0.795, P<0.001) compared with iEEG only (AUC=0.750, P=0.02) or scalp EEG only (AUC=0.764, P<0.001). The validation scalp EEG only data resulted in an AUC of 0.844. Combining the two modalities appeared to capture spatiotemporal advantages from both modalities.

Conclusions

The combination of iEEG and scalp EEG better predicted sedation level than either modality alone. The scalp EEG only model achieved a similar AUC to the combined model and maintained its performance in additional patients, suggesting that scalp EEG models are likely sufficient for real-time monitoring. Deep learning approaches using multiple leads to capture a wider area of brain activity may help augment existing EEG monitors for prediction of sedation.