Correlation dimension of electroglottographic data from healthy and pathologic subjects.

This paper considers the effects of nonstationarity, noise, and finite data sets on the estimation of the correlation dimension of time-series data characterizing the vibratory behavior of the vocal folds. The electroglottographic signal of sustained /a/ phonations from 10 healthy subjects and 20 subjects with vocal fold pathologies were used to reconstruct the state space in successively higher embeddings using the method of lags. The dimension values were calculated from the scaling region (the level area of the slope plots) which did not increase for higher embeddings. Reasonably defined scaling regions were found in all of the data from the healthy subjects and from five of the pathologic subjects, with values saturating between the first and second embeddings. The EGG data from those five pathologic subjects were nearly periodic. From the remaining 15 subjects, the scaling regions were highly constricted with nonconstant slopes, so that dimension values could not be confidently estimated. The results suggest that the correlation dimension is a highly subjective measure which is not usefully applied to abnormal EGG data. It is recommended that, if used, correlation dimension statistics need to be presented cautiously, and graphical presentation of the data should be included.