Skip to Main Content
An ANFIS based neuro-fuzzy classifier with a pruning algorithm was implemented and applied to the classification of sleep-waking states-stages in infants, using the sleep pattern detection system of P.A. Estevez (2002) to generate the inputs. Including artifacted pages, an average of 88.2% of expert agreement was achieved for testing data. As a result of the training process and pruning, rules and parameters that defined a fuzzy classification system were also determined. Analyzing the rules obtained for sleep-stage NREM-1, it was found that the main rule matched the expert rule to classify NREM-1. Additional rules were discovered that complement the classification and may provide additional information about the characteristics of this sleep stage. This is a promissory result, and further research is needed in this topic. Future work includes implementation of a clustering algorithm to determine the initial parameters of the system, training the system with a different objective function, such as the max-type error function described in J.S.R. Jang and C.T. Sun (1993), and evaluating the performance of different T-norms at layer 2 in Figure 1. The development of a general methodology for rule discovery and interpretation is also of interest.