Modelling of bispectral index using artificial neural networks during inhalational anaesthesia

The bispectral index (BIS), a value derived from the electroencephalograph (EEG), has been proposed as a measure of anaesthetic effect. Therefore, this paper shows the development of a system to model BIS using artificial neural networks (ANNs) based on measured systolic arterial pressure (SAP), heart rate (HR), and end-tidal anaesthetic agent concentration (Etaa). It was trained and tested in 8 patients during inhalational anaesthesia using desflurane anaesthetic gas. The network was designed a four-layer network with one input layer, two hiddel layers, and one output layer. The determination of the optimum nodes of two hidden layers was used an automated regularization algorithm. Also, the network was used an early stopping method to find the minimum cost function of different 8 algorithms which were built-in functions of the neural networks toolbox of MATLAB. According to the off-line analyses fo 8 patients, the Leveberg-Marquardt algorithm can be obtained a minimum cost function (i.e. 0.0135). Then, using this algorithm to iterate further 2000 epochs, the best cost function can be finally reduced to a 0.00064 value. Furthermore, using the final weighting function to model the BIS of these 8 trained patients, the errors between the clinical and model values of BIS were acceptable except one patient which the raw data of BIS, SAP, HR and Etaa had missed several segments due to nearly end of operation. If we didn't take into account of this patient, the average and standard deviation of cost function of those 7 patients were 0.020±0.026.