Multimodal Deep Learning for Depth of Anesthesia Prediction

Abstract

Bispectral Index (BIS) is a widely adopted measure of anesthetic depth, which relies on Electroencephalography (EEG) to record brain’s electrical activity. Contemporary research in machine learning, aimed at replicating the BIS, focuses mainly on EEG recordings. Such models might underperform as this approach limits their perspective, excluding other important biological signs, such as blood pressure, heart beat or preoperative context, which have major influence on brain patterns. To address these limitations, this thesis proposes a multimodal deep learning strategy to accurately predict the bispectral index. Utilizing the public VitalDB database, the study combines extracted EEG frequency power spectra with static case context and intraoperative vital signs. The predictive capabilities of four distinct models: XGBoost, Long Short-Term Memory (LSTM), a vanilla Transformer, and an iTransformer, were evaluated on a filtered cohort of 383 surgical cases. Experimental results indicate that integrating static patient context did not improve the predictive capabilities of most architectures, mostly driven by the lack of a physiological truth to anesthetic depth. The LSTM architecture emerged as the most effective model, achieving a Root Mean Absolute Error of 4.63 and capturing 89.56% of the variance in the BIS value. Ultimately, all models were able to replicate the BIS at a satisfactory level. Furthermore, the results for model bias and anomaly handling indicate that the models have a good understanding of the target, and might be efficiently leveraged on a more truthful target variable.