Rehabilitation Monitoring of Post-Laryngectomy Patients through the Extraction of Vocal Parameters

This paper deals with the analysis of substitution voices of post-laryngectomy patients with the aim of identifying a methodology to track the effectiveness of rehabilitation therapies. The available data-set includes 22 patients that had undergone Open Partial Horizontal Laryngectomy (OPHL) of type II and 10 patients that had undergone OPHL of type III. A pre-processing algorithm that relies on the spectral kurtosis has been designed to remove non-harmonic frames from the available recordings of patients, thus minimizing the number of outliers among the extracted parameters. Such an algorithm has been tuned starting form the results of a control group of 10 healthy subjects. After this preliminary step, from the harmonic frames a series of parameters have been extracted that belong to spectral domain (tilt, kurtosis, entropy and Soft Phonation Index) and cepstral domain (Mel-Frequency Cepstral Coefficients, MFCCs, and Cepstral Peak Prominence Smoothed, CPPS). Then, the patients have been subdivided into two classes according to the index I (Intelligibility) of the auditory perceptual scale INFVo and a Kolmogorov-Smirnov two-samples test has been run, which has highlighted that low-band MFCCs, spectral entropy and spectral kurtosis show the best discrimination capability of substitution voices. This outcome has been confirmed by an alternative method that is based on the performance of classification algorithms. A classification accuracy of about 81% has been obtained using a logistic regression model that was trained with median of MFCC3, range of MFCC4 and 95° percentiles of MFCC6 and MFCC9. The same accuracy has been provided by a coarse decision tree algorithm trained with skewness of MFCC1, median of MFCC3 and 95° percentile of spectral entropy.