Development and testing of a neuro-fuzzy classification system for IOS data in asthmatic children

Respiratory disorders can be difficult to diagnose because their symptoms are sometimes similar to one another. A Co-active Neuro-Fuzzy Inference System (CANFIS), however, is capable of classifying and characterizing features of major respiratory diseases such as asthma after it is adequately trained with a data set containing data patterns sampled from both healthy and unhealthy subjects. The attributes of these data patterns can be physical parameters of the lung\u27s mechanics or respiratory airway impedances obtained using pulmonary function tests. Impulse Oscillometry (IOS) is a patient-friendly pulmonary function test which provides impedance values of the respiratory airways as a function of frequency. This research employs IOS impedance measurements and parameter estimates of respiratory impedance models as attributes of data patterns for training two CANFISs using the backpropagation algorithm. Both CANFIS use data sets containing 112 patterns belonging to four different classes representing asthma disease severities: asthma, small airway disease (SAD), mild SAD and normal. The first CANFIS utilizes IOS impedance measurements and parameter estimates of the augmented RIC model, while the second CANFIS utilizes parameter estimates of the extended RIC model instead. The two CANFISs respectively produce a 97.32% and 95.54% memorization accuracy and a 100% and 95.83% generalization accuracy. The CANFIS\u27s performance is further investigated using a yeast data set from the University of California at Irvine\u27s machine learning repository, which consists of 1484 data patterns of 8 attributes belonging to 10 different classes, and a memorization accuracy of 45.14% is achieved. The sensitivity and specificity of the asthma severities for the best CANFIS memorization is 100% and 96.34% respectively for the asthma group, 96.23% and 98.31% for the SAD group, 100% and 96.84% for the mild SAD group and 91.67% and 98% for the normal group. The promising classification performances obtained by the two CANFISs, which can aid physicians in disease detection, diagnosis and management, can be improved by incorporating automatic fuzzy rules extraction, increasing the number of data patterns and reducing the CANFIS\u27s training time by controlling the learning parameters of the backpropagation algorithm