Researchers have achieved accuracies of 99.4% and 94.3% in two algorithmic methods for monitoring, diagnosing or ruling out Parkinson’s disease going only by individuals’ spoken words.
The team built a set of 126 voice markers (i.e., “features”) touching everything from tone, pitch and loudness to enunciation, pace and pause ratio.
Additionally, they had their best model analyze 25 isolated Spanish-language words pronounced by each study subject (50 Parkinson’s patients and 50 healthy controls).
The study’s lead author is Federica Amato, a PhD candidate in computer engineering at the Polytechnic University of Turin in Italy. Senior author is electronic engineer and computer scientist Juan Rafael Orozco-Arroyave, PhD, of the University of Antioquia in Medellín, Colombia.
In a study posted in Health Information Science and Systems, the team explains:
From an engineering perspective, the human vocal signal can be seen as a quasi-periodic train of air pulses that are shaped by the resonances of the vocal tract. The frequency of the train pulses, i.e., the number of glottal contraction per second, represents the fundamental frequency or pitch, while the resonance frequencies of the oropharyngeal cavities account for the vocal formants. Fundamental frequency is influenced by the intrinsic features of the phonatory system and is distinctive of [a] single speaker to a large extent. It is also influenced by anatomical characteristics dependent on the speaker’s gender.”
The team incorporated such engineering insights to devise their high-performing machine learning model for classifying Parkinson’s.
The model’s promising performance proves out the concept of automated Parkinson’s evaluation using voice recordings, Amato and colleagues note.
What’s more, they suggest, the approach may be easily enough transportable to consumer platforms that a simple smartphone app would be a sensible development.
The authors add:
This work confirmed the possibility of a speech-based Parkinson’s disease classification, suggesting new promising methodologies for vocal feature analysis. Furthermore, the usage of features extracted from common words gives rise to a new perspective on passive speech-based monitoring of Parkinson’s patients. Specifically, given the high precision reached by our algorithm, it may be employed in the home monitoring of motor fluctuations in Parkinson’s subjects, as well as a decision support system in early diagnosis.”
The study is available in full for free.