Monitoring the depth of anesthesia using Autoregressive model and Sample entropy

Abstract

Anesthesia is an important part in modern surgery, and the way how to effectively monitor the depth of anesthesia (DOA) is core issue in the anesthesia work. Since anesthetics mainly affected the brain of patients, it is very effective to monitor DOA by electroencephalogram (EEG). This paper proposes a method for monitoring DOA using EEG. First, the sample entropy (SampEn) of EEG were calculated as a feature vector. Simultaneously, the Burg recursive algorithm was used to solve the autoregressive model (AR model) and AR coefficients were extracted as feature vectors. Later, according to the characteristics of uneven distribution of sample points, the weighted k-nearest neighbor (WKNN) classifier was selected. The Anesthesia was divided into awake, mild, moderate and deep by WKNN classifier. According to the results, the correlation coefficient between the SampEn of the EEG and Bispectral Index (BIS) is above 0.8. There is a correlation between the first five orders of AR coefficient and the BIS index, and the correlation of the second order reaches 0.8. Through the validation of 30 patients, this method can assessment of DOA effectively and accurately.

Boosting framework via clinical monitoring data to predict the depth of anesthesia

“…Compared with ECG and other clinical indicators commonly used in previous studies [1], we selected four clinical indicators EMG, ETCO2, remifentanil dosage, and flow rate that were more closely associated with BIS to predict the depth of anesthesia. Besides, in previous studies on the depth of anesthesia, several popular machine learning algorithms such as DT, KNN, and SVM, were used to build prediction models [4,6,8]. In this paper, a boosting-based prediction model to predict the depth of anesthesia was built based on four clinical monitoring data.…”

“…The K-nearest neighbor (KNN) [7] was used in clinical practice to classify states of anesthesia as awake, mild, moderate, and deep [8]. However, these methods are sensitive to parameters and easy to overfit, and have low precision.…”