Complexity Analysis Using Nonuniform Embedding Techniques for Voice Pathological Discrimination

Automatic voice pathology detection systems are receiving an increasingly growing interest due to the their advantages compared to traditional systems (non invasiveness, objectivity and reduction in time and cost analysis). In this respect, the complexity analysis has been also introduced within the automatic detection framework for providing more reliability and higher system performance. However, such an analysis needs for an initial reconstruction stage that is devoted to map the studied voice recordings onto phase space. Mapping is usually carried out using a uniform embedding. Nonetheless, the uniform embedding might be no the optimal for biosignal processing having different timescales, i.e. a short lag is optimal for high frequency components whereas a long lag is related to low frequency components and modulations, so, to get a unique optimal value lag leads to inadequate processing for both timescales. To cope with this drawback, this work discusses a non-uniform embedding approach that is validated by using two complexity features. The attained results using k-nn and Support Vector Machines show superior performance in comparison to those reached by uniform embedding, and thus providing evidence of the validity of the discussed training technique.