Advancing Parkinson’s disease diagnoses through speech-based machine learning

Parkinson’s disease (PD) is the second most common neurodegenerative disease, affecting some 10 million people worldwide. In later stages, PD causes gross motor dysfunctions including bradykinesia, akinesia, tremor, and postural instability. By this time, as much as 50% of dopaminergic neurons are irreversibly damaged — making early detection crucial for effective treatment.

Because voice and speech disorders often manifest in patients with PD as early as five years before gross motor dysfunctions, the development of speech-based machine learning (ML) systems has emerged as a path to expedite disease detection. Still, there is room for improvement in distinguishing prodromal PD speech from that of healthy controls (HCs).

Mittapalle Kiran Reddy and Paavo Alku introduced a two-layer wavelet scattering network (WSN) that provides a framework to characterize speech-based information related to pitch, articulation and amplitude modulation.

“Most studies on automatic speech detection from patients with PD focus on extracting features that effectively characterize the information into two main aspects of speech production: articulation and phonation,” said Reddy. “Using a WSN avoids the need to use separate methodologies to represent each aspect of speech production and therefore reduces the inconsistencies in feature representation.”

Using a popular speech corpus database, the team examined speech signals from a variety of speaking tasks performed by 50 PD patients and 50 HCs.

“Our experimental results indicate that the ML systems developed using features extracted with a WSN provide better PD speech detection compared to other existing systems,” said Reddy. “While extensive investigations are still needed to establish the reliability of the proposed approach, the results are promising for PD screening and diagnoses.”

Source: “Automatic detection of parkinsonian speech using wavelet scattering features,” by Mittapalle Kiran Reddy and Paavo Alku, JASA Express Letters (2025). The article can be accessed at https://doi.org/10.1121/10.0036660 .