Combination of EEG and HRV parameters for nociception assessment on anesthetized subjects

Abstract

Maintaining a good level of anesthesia and nociception along a surgery is essential for the both the intraoperative and the postoperative outcomes. General anaesthesia (GA) is the combination of hypnotic effect, in order to achieve unconsciousness of the patient, and analgesia to pervent nociception. Analgesia has been historically administered based on clinical signs during the procedure. However, in the recent years several monitors have appeared on the market in order to try to achieve a metric that is not arbitrary to aid the physicians and medical doctors. Since these monitors are relatively new and there is not yet a golden standard for nociception measurement, there is still plenty of research to carry out. Hence, the aim of this work is to develop mathematical regression models with the purpose of predicting the level of nociception during general anaesthesia. As which are the physiological parameters that carry more relevant information is still uncertain, different combinations of electroencephalography (EEG) and heart rate variability (HRV) parameters have been tried as inputs for these prediction models. The reference target for the models is a combination of the concentration of anaesthetic drug (remifentanil), the consciousness of the patient (qCON index), and the response to tetanic stimuli. The data used for this study was gathered during 475 gynecological ambulatory procedures in Hospital Clinic of Barcelona. The data was split in two different sets of patients, patients for training the models and patients for testing the models in a 1:2 ratio. As the data gathering took five years, in order to prevent the bias in time due to change of the data collector or by the learning process of the collection of data, we grouped the patients chronologically; every three patients the first was re-grouped unto training set and the two remaing unto test set. The selection of the features was carried out through three different methods, RReliefF algorithm, random forest predictors’ weights and a custom algorithm based on quadratic regression. The regressive models trained were quadratic, random forests and shallow neural networks. After predicting the output for every model, an exponential average was calculated patient by patient for smoothing the output. To assess the different models’ output, a fitness function was designed and developed that contained the correlation with the reference, correlation with concentration of remifentanil, prediction probability (pk) for prediction and detection of movement. Besides the score, the behaviour of the different models before and after tetanic stimuli, the output of the models vs. the concentration of remifentanil and the output of the model for awake patients was analyzed. The current index used in the company Quantium Medical S.L.U., qNOX was also assessed using these metrics. Only 5 of the 18 developed models were able to outperform qNOX before the exponential averaging but 10 out 18 afterwards. Before the exponential averaging only hybrid subsets (EEG + ECG) were able to outperform qNOX however, afterwards the EEG only also outperformed it, all these sets and models achieved a movement prediction probability (pk) greater than 0.7. ECG parameters alone were not good predictors and the shallow neural networks failed with every subset.