Machine learning- and statistical-based voice analysis of Parkinson's disease patients: a survey

The preliminary diagnosis and evaluation of the presence and/or severity of Parkinson's disease is crucial in controlling the progress of the disease. Real-time, non-invasive methodologies based on machine learning-enhanced voice analysis are gathering more interest as the potential of this field unveils. Specifically, acoustic features are employed in many machine learning techniques, and could also function as indicators of the overall state of the subjects' voice: this review aims at identifying the most widely employed and promising feature-based machine learning methodologies, evidencing baselines and state-of-the-art solutions. A total of 102 works plus 5 review articles were selected from the IEEE Xplore, PubMed, Elsevier, and Web of Science electronic databases. A statistical assessment is performed identifying the most frequently used features as well as those deemed as most effective; an overview of algorithms, public datasets, toolboxes, and general metadata is also performed. According to our results, Jitter, Shimmer, Harmonic-to-Noise Ratio, Fundamental Frequency, and Mel Frequency Cepstral Coefficients are the mostly adopted features. In addition, it is worth noting a fair prevalence of glottal-like models and additional filtering options, such as Detrended Fluctuation Analysis.