Machine learning- and statistical-based voice analysis of Parkinson’s disease patients: A survey

Artificial Intelligence-Based Voice Assessment of Patients with Parkinson’s Disease Off and On Treatment: Machine vs. Deep-Learning Comparison

“…We would like to stress that most of our previous work within the same context point to the same conclusion, even when using transfer learning and comparing architectures [ 20 , 24 ]. Moreover, the plethora of studies involving voice analysis for PD, albeit showing a trend towards the usage of CNNs in the last few years, still achieve equally relevant results with ML methods [ 52 ]. However, let the reader be reminded that accuracies and trends in specific tasks with limited datasets can only point out a “direction” for future studies to take, define a more thorough baseline methodology and present all the possible viable alternatives, and it is not recommended to draw strict conclusions on the accuracy of specific models.…”

“…Taking a closer look at Table 1 and in general on the overview of the current state-of-the-art of voice analysis for PD (as detailed in [ 52 ]), we believe that the strengths of this work can be summarized as: the extensive set of well-recorded, validated data; a comprehensive approach comparing ML and CNN methodologies; and the usage of a broad spectrum of acoustic features.…”

“…To build a comprehensive baseline for voice analysis for PD detection, we strove to present a thorough literature review [ 52 ], especially directed towards feature-based methodologies, which we will refer to for more detailed statistics. For the sake of completeness, a brief overview of relevant works is also presented in Table 1 : both ML and DL approaches are covered, with the main limitations often involving small or poorly recorded datasets as well as a general lack of interpretability.…”

“…To face these challenges, we strove to build an extensive, well-prepared dataset involving a grand total of 426 subjects, without relying on crowdsourced, non-validated data. Other common datasets, as detailed in [ 52 ], are those presented by Little, Naranjo and Tsanas that involve a small amount of PD subjects, along with Sakar’s one involving 180 PD subjects recorded in unspecified conditions with unspecified devices, as well as the larger mPower dataset of over 1000 subjects, but the recordings for which are crowdsourced and unverified.…”

Voice Disorder Multi-Class Classification for the Distinction of Parkinson’s Disease and Adductor Spasmodic Dysphonia

“…This is because, in general, voice analysis for PD detection often relies on some specific prosodic features like the fundamental frequency (F0), jitter, shimmer, Ssignal-to-noise ratio, or mel-frequency cepstral coefficients (MFCCs). However, such a subset of features alone may lead to underestimation and, moreover, there is no consistency or proven performance boost in using a certain subset with respect to another, as evidenced in a recent review [20]. As such, here, we extracted a large number of features employing a toolbox based on the INTERSPEECH 2016 feature set [21], which contains a vast amount of low-level descriptors (average, quartiles, delta coefficients, etc.)…”