An Analysis of Automated Parkinson’s Diagnosis Using Voice: Methodology and Future Directions

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].…”

“…In most of our experiments, especially those employing the Neurovoz and ItalianPVS corpora, the dimensionality of the feature vectors was larger than the number of recordings. In order to remove redundant or irrelevant information, we employed two dimensionality reduction techniques, similarly to other previous studies [12,27,37]. The first technique we utilized to reduce dimensionality was PCA from the Python scikit-learn module [57].…”

“…We also incorporated a hyperparameter grid search as described in the next subsections to provide the highest possible accuracy with these models. Although there are several examples of studies employing Deep Neural Networks (DNN) to assess PD [14,19,[25][26][27]36,37,58], most of the works found in the literature employ approaches with simpler models, as the corpora are usually not large enough to train DNN.…”

“…During the testing phase, each tree provides its own score, and then the average score is taken as the collective forest decision. Several studies report high classification accuracy employing RFCs for PD assessment in similar scenarios [12,19,25,27,30,36,37,59]. Table 10 lists the hyperparameter ranges we used in the grid search:…”

“…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.…”