Uncertainty modeling for classification and analysis of medical signals

Medical time-series signal classification and analysis is an interesting problem for both medical and computer science applications. Detection and evaluation of gait locomotion and mild neurological dysfunction in humans and animals by subjective visual assessment is a difficult medical task. Design and implementation of computational methods that analyze medical signals may be used for screening and characterization of disease phenomenon. We have developed an objective method for analyzing kinematic data and tested the methodology using several equine gait models. Features were extracted from the movement of key body parts of horses trotting on a high-speed treadmill using uncertainty-based information models. Starting with a gait lameness data set, temporal features were extracted using continuous wavelet analysis guided by combined uncertainty models for wavelet selection. A wavelet is selected from a set of wavelets if its transformation is best, in a maximum uncertainty type that combines together fuzzy and probabilistic models. A time-sequence composition process was used to create feature vectors that capture transitions of the transformed signals within a moving time window. A neural network was then trained via back propagation to classify the transformed signals. The method was successful at detecting mild lameness and for differentiating right forelimb from left forelimb lameness. Tests performed using this method showed that a horse's poll is needed for detecting lameness, and that adding one leg point to it can identify the side of lameness. A second set of experiments started with neurological spinal ataxia gait data. Signal features were directly extracted using combined uncertainty modeling of signal self-information. Fuzzy clustering successfully separated normal from pathological data, in feature space. A sequential selection algorithm was able to select a minimal set of features that may be used to characterize the disease. These results show that uncertainty modeling may be used in disease analysis and recognition. They also encourage further investigation of uncertainty models that integrate together other types of complex information in time-series signals.