Design and Validation of a New Diagnostic Tool for the Differentiation of Pathological Voices in Parkinsonian Patients

Things to Consider When Automatically Detecting Parkinson’s Disease Using the Phonation of Sustained Vowels: Analysis of Methodological Issues

“…This is the largest corpus of its kind to date which is publicly available, with 5876 unique participants contributing a total of 65,022 recordings. Previous studies have used the mPower corpus not only to create and test PD detection models [12][13][14]25,27,30,36,37], but also to design real-time PD diagnosis tools and applications [26], classify voice impairment level [38], measure longitudinal reliability and stability of these metrics [20], quantify and improve diagnosis techniques on signals recorded in noisy environments [39,40], and even screen for symptoms of depression reported by PwPD [41].…”

“…Another common problem identified in the literature is that some studies employ data sets in which the average age difference of speakers with and without PD exceeds 5 years [10][11][12][13][14][15]. Generally, these studies report high recall in their trials of automatic PD detection.…”

“…This is usually more adequate when employing acoustic analysis or features proposed to be used in the stable part of the phonation, such as complexity or frequency and amplitude perturbation features [18]. Therefore, some studies take corpora distributed with the full recording and shorten them to as short as one second [14,19,20]. Nevertheless, the onset and offset contain articulatory information and might add extra differentiation between classes, especially when employing certain features such as Mel-Frequency Cepstral Coefficients (MFCC) to characterize them.…”

“…It is uncommon to find studies which employ different corpora in model training and testing (cross-corpora validation) [22]. In some cases, the authors do not use cross-validation, and divide the corpus into training and testing subsets randomly [14,[22][23][24][25][26]. This increases the uncertainty of the results when using small corpora (usually no more than 3 h of recordings), as the testing partition is not large enough to be considered representative.…”

“…Moreover, some studies where there are more than one recording per participant ensure that recordings from a speaker used to train a model are not considered to test or validate the same model (subject-wise split) [19,25,27], whereas other studies randomly select the recordings employed in the training and testing processes or intentionally use recordings from the same speakers in training and testing subsets (record-wise split) [11,14]. The latter studies do not consider that when both training and testing processes contain speaker identity information encoded in sustained vowels, their classifiers may be detecting characteristics of specific speakers in addition to PD patterns.…”