Hybrid Correlation-Neural Network Synergy for Gait Signal Classification

Abstract

This chapter presents a thorough discussion on the development of a robust algorithm for pathological classification of human gait signals. The technique involves the extraction of time and frequency domain features of the correlograms obtained by cross-correlating the gait signals with a reference, and subsequently employing a pre-trained Elman’s recurrent neural network (ERNN) for automatic identification of healthy subjects and those with neurological disorder, and also the type of disorder. To assess the performance of the algorithm, stance, swing, and double support intervals (expressed as percentages of stride) of 63 subjects, either healthy, or suffering from Parkinson’s disease (PD), Huntington’s disease (HD), or Amyotrophic Lateral Sclerosis (ALS), have been processed by the proposed algorithm for a period of approximately 300 s. The performances of ERNNs are also compared with those already reported for back propagation neural network (BPNN), learning vector quantization (LVQ), and least-square support vector machine (LS-SVM) based classification algorithms. With time-domain features, the proposed modular ERNNs outshined the other classifiers by attaining 90.3–98.5 % classification accuracy for binary classification jobs, and an accuracy as high as 87.1 % for the four-class classification problem. With frequency-domain features, classification into healthy and pathological subjects has been studied, and in this case also, the best performance of 81.6 % mean accuracy was achieved employing ERNN.