The detection of stages of anaesthesia is performed by using vital signs like heart rate, blood-pressure or oxygen saturation for example. The anaesthetist is reacting to the values of these signs to de- or increase the concentration of the narcotic drugs to control the stage of anaesthesia of the patient. But the regular used vital signs are not indicating the depth of consciousness directly. In addition the patterns of the electroencephalogram (EEG) indicate the stages of anaesthesia while performing a surgical operation [1–4]. By assessing these EEG patterns the anaesthetist is getting a meaningful value to estimate the state of the patient. The structure of the electroencephalographic patterns differs throughout the different stages of anaesthesia. So to react to the patterns of EEG directly could be a heavy task for the anaesthetist. A simple EEG-based value is needed to add a robust indicator for classification of stages of anaesthesia. Therefore an automatic approach of a classification-algorithm could be implemented into an embedded system. This system reacts only to the variation of the frontal one-channel pattern of EEG and is indicating three stages of anaesthesia: “Wake”, “Anaesthesia” and “Asphyxia”.
To establish an autonomic device which is able to classify the different stages of anaesthesia the design of a classifier system is necessary in a first step. The training of the classification model should be established before implementing this classification model to an embedded system or microcontroller-based system. So the training itself has to be performed on a standalone computer or computer grid. In a second step the hardware itself has to be constructed. This hardware will capture the frontal one-channel EEG and processes a feature extraction. These features are than presented to the implemented classification approach. At least the classification system will calculate the different stages of anaesthesia and present it to the anaesthetist in a simple way.
2.1 Design of the classification model
The overall database is defined by datasets of the frontal one-channel EEG collected from different patients and chirurgical operations. The training data set is defined by 18 records collected throughout different surgical operations. The validation data set includes 20 additional independent records. In a first step features were extracted and used to train different classification approaches. To enable a linguistic interpretation of the classification system fuzzy rules were selected to evaluate the extracted features sets. By the means of multiobjective evolutionary algorithms simple fuzzy rules were generated . These rules are able to distinguish between the three anaesthetic stages “Wake”, “Anaesthesia” and “Asphyxia”. The number of rules is limited to five rules for each class. To enhance the robustness of the classification system there were populations of rule sets generated. These populations include up to 18 independent rule sets. The population itself was established by the applied training algorithm . The constructed classification system has a very simple structure connected to an appropriate value of robustness. A linguistic interpretation of the rule sets is possible. So there were generated efficient classifiers to support the anaesthetist in defining the actual anaesthetic stage of the patient. In addition by using fuzzy rules a linguistic interpretation could be performed to estimate the importance of the extracted EEG features.
2.2 Construction of the embedded system
An embedded system should be used to analyse the frontal one-channel EEG in an autonomic way. The construction of this device has to be transposed in a simple and efficient way. The generated population of fuzzy rules should be implemented to work as an essential part of the firmware. Therefore some hardware system approaches were presented in [6–9].
Populations of fuzzy rule sets were generated by applying the algorithms presented in . In a first step the selection of the linguistic variables and the corresponding rule structure for each fuzzy rule set was optimized. Therefore different Multiobjective Evolutionary Algorithms (MOEA) were used . The Non-dominated Sorting Genetic Algorithm II (NSGAII) presented in , the Strength-Pareto Evolutionary Algorithm 2 (SPEA2) and the Indicator-based Evolutionary Algorithm (IBEA) [11, 12] generated populations of fuzzy rule sets. In a second step the number of fuzzy rule sets was reduced by using a genetic algorithm applied for each population. As presented in  for each MOEA-specific population a subset of fuzzy rule sets was extracted. The genetic algorithm determined for the NSGAII-optimized population 3 fuzzy rule sets as an optimal classification model. Out of the SPEA2-optimized population also 3 fuzzy sets were specified. The IBEA-optimized population provides 5 fuzzy rule sets after applying the genetic selection algorithm.
In Addition to the results presented in  these 11 different fuzzy sets can be used to classify the stages of anaesthesia in common. To raise the efficiency the genetic selection algorithm can be applied to this new population of fuzzy rule sets. The generation count was limited to 500. As a main result of this selection procedure 7 fuzzy rule sets were found to classify the stages of anaesthesia of 20 independent features sets of validation data. The classifi-cation performance of these fuzzy rule sets is presented in table 1.
The results presented in table 1 illustrate the median values of the decision vectors generated for each cell of 20 performance matrices. The matrices itself correspond to the classification output of each fuzzy set of the produced population. Each line of the table matches the manual rated classification of the experts. Each column represents the median degree of concordance in percentage between the manual rated epochs of EEG and the results received by the population of fuzzy rule sets. The anaesthetic stages “Wake”, “Anaesthesia” and “Asphyxia” indicate an accordance of 78.42%, 86.77% and 85.65%. The given anaesthetic stage “Asphyxia” was not classified as “Wake” by the automatic approach as presented in table 1.
By applying the general training algorithm as presented in  populations of optimized fuzzy sets were generated. These fuzzy rules are able to classify stages of anaesthesia by evaluating the frontal one-channel electroencephalogram. The fuzzy rules were optimized using the multiobjective evolutionary algorithms NSGAII, SPEA2 and IBEA. For each MOEA an independent population of fuzzy sets was generated. The size of each population was reduced by a genetic algorithm. The remaining fuzzy sets of each population were used to create a new population of effi-cient fuzzy rules. To reduce the computation effort in an additional way the genetic search algorithm was applied again to this new population. The number of fuzzy sets was reduced from 11 to 7. In relation to the results as presented in  the performance was raised lightly by using this new reduced fuzzy rule set to classify the validation data set. In a next step this new population of fuzzy rule sets has to be implemented into an embedded system to enable an online classification of anaesthetic stages.
Courtin, R. F.; Bickford, R. G.; Faulconer Jr., A.: The Classification And Significance of Electro-Encephalographic Patterns Produced by Nitrous Oxide-Ether Anesthesia During Surgical Operations, Proceedings of the Staff Meeting Mayo Clin 25, No. 8, pp. 197-206, 1950 Google Scholar
Wenzel, A. : Robuste Klassifikation von EEG-Daten durch Neuronale Netze - Untersuchungen am Beispiel der einkanaligen automatischen Schlafstadien- und Narkosetiefenbestimmung, Shaker Verlag Aachen, 2005Google Scholar
Baumgart-Schmitt, R.; Walther C.; Wenzel, A.: Multi criteria evaluation of sleep and anesthesia by neural networks, fuzzy rules, evolutionary algorithms and support vector machines, IFMBE Proceedings, Vol. 25, No. 4, pp. 2254-2255, 2009Google Scholar
Walther, C.: Multikriteriell evolutionär optimierte Anpassung von unscharfen Modellen zur Klassifikation und Vorhersage auf der Basis hirnelektrischer Narkose-Potentiale, Shaker Verlag Aachen, 2012 Google Scholar
Walther, C.; Baumgart-Schmitt, R.; Menz, C.; Trommer, D.; Krautwald, M.; Sturm, K.-P.; Jäger, U.: Multi-criteria Model-driven Feature Selection Using Fuzzy Ruled Classification Of Stages Of Anaesthesia, Proceedings of the 55th International Scientific Colloquium, Ilmenau, Germany, pp. 519-522, 2010 Google Scholar
Fraenzel, N.; Weichert, F.; Wenzel, A.; Ament, C.: A prototyping system for smart wheelchairs, Biomedizinische Technik, Vol. 59, pp. 902-905, 2014 Google Scholar
Faenger, B.; Becker, F.; Brandl, M.; Fränzel, N.; Holder, S.; Köhring, S.; Lutherdt, S.; Michaelis, A.; Weichert, F.; Schumann, N. P.; Scholle, H. C.; Zimmermann, K.; Augsburg, K.; Ament, C.; Linß, G.; Witte, H.; Pezoldt, K.; Wenzel, A.: Mobil-itätssysteme für Best Ager (50plus): Neue Präventionsansätze. In: Dienstbühl, I.; Scholle, H. C.; Stadeler, M.: Prävention von arbeitsbedingten Gesundheitsgefahren und Erkrankungen: 20. Erfurter Tage, pp. 363-367, 2014Google Scholar
Menz, C.: Entwicklung einer Softwarebasis zur Echtzeitanalyse von EEG-Daten auf einer Mikrocontrollerplattform, Project Report, University of Applied Sciences Schmalkalden, Germany, 2010Google Scholar
Hess T.; Hopf, M.: Entwicklung eines mikrocontrollerbasierten Systems zur Signalextraktion, Verarbeitung und echzeitfähigen Klassifikation von biometrischen Signalen, Project Report, University of Applied Sciences Schmalkalden, Germany, 2011 Google Scholar
Deb, K.; Pratap, A.; Agarwal, S.; Meyarivan, T.: A Fast Elitist Multi-Objective Genetic Algorithm: NSGA-II, IEEE Transactions on Evolutionary Computation 6, pp. 182-197, 2000 Google Scholar
Zitzler, E.; Laumanns, M.; Thiele, L.: SPEA2: Improving the Strength Pareto Evolutionary Algorithm, Research Report, Swiss Federal Institute of Technology (ETH) Zurich, 2001 Google Scholar
Zitzler, E.; Künzli, S.: Indicator-based Selection in Multiobjective Search, Proceedings of the 8th International Conference on Parallel Problem Solving from Nature (PPSN VIII), Springer, pp. 832-842, 2004 Google Scholar
About the article
Published Online: 2015-09-12
Published in Print: 2015-09-01
Conflict of interest: Authors state no conflict of interest. Material and Methods: Informed consent: Informed consent has been obtained from all individuals included in this study. Ethical approval: The research related to human use has been complied with all the relevant national regulations, institutional policies and in accordance the tenets of the Helsinki Declaration, and has been approved by the authors’ institutional review board or equivalent committee.