Researchers from Italy use machine learning to differentiate between different stages and severity of Parkinson’s disease by voice

Parkinson’s disease (PD) is a neurological condition that causes tremors, stiffness and difficulty walking, balance and coordination. Dopamine levels are reduced due to damage to nerve cells in the brain, resulting in Parkinson’s symptoms.

PD patients often complain of a variable disorder of voice distribution. These patients may have speech problems even in the prodromal phase of the condition. Symptoms of Parkinson’s disease usually appear gradually and worsen over time, eventually leading to severe voice impairment in the advanced stages of PD.

Current clinical methods of voice assessment in PD are based only on qualitative assessment. This spectral analysis reveals irregularities in certain sound qualities in people with PD, including a decrease in the basic frequency and the ratio of harmonics to sound and an increase in jitter and glare. However, the human voice is complex, containing high-level information based on an exponential number of characteristics.

As a result, in addition to independent testing of specific sound characteristics, advanced methods of analytical capability and aggregation of a high-volume data set of sound characteristics are required, which accurately classify the targets of sound samples in PD.

Mechanical study methods have made it possible to provide automatic classification of voice impairment in various neurological diseases with high accuracy. However, to date only a few research studies using the study of mechanical engineering analysis in PD have been reported. It is important to understand that machine learning can differentiate between patients at different stages of the disease to see if it can recognize the severity of the disease.

A new study by researchers in Italy and Jordan studied the voices of Parkinson’s disease patients in a large and clinically well-described group. This study is the first to classify voice in Parkinson’s disease patients based on the stage and severity of the disease and the effects of chronic L-Dopa medications. All diagnostic tests were evaluated for sensitivity, selectivity, positive and negative predictive and accurate values.

The Neuromed Institute of the IRCCS and the Department of Systemic Medicine at Thor Vergata University in Rome, Italy, recruited participants for the study. Participants included 115 people with Parkinson’s disease and 108 healthy subjects (HS). All patients were Italian-speaking and did not smoke. There were no reports of bilateral or unilateral hearing loss, respiratory disease, or any neurological disorders affecting the vocal cords.

Participants were given a specific task of speaking with the usual volume, volume and quality of sound recording. This task involves the constant release of a unique front-end soundtrack.

The researchers used OpenSMILE (an open source set for extracting and classifying audio features) to pre-process each speech sample. They collected 6139 sound attributes from each speech sample. They used the Correlation Features Selection Algorithm (CFS) algorithm to find sound (nonlinear) qualities that were significantly associated with the class. As a result, the original data set was deleted from duplicate and / or useless data. Subsequently, the Information Feature Assessment Methodology (IGAE), based on the Pearson correlation method, was used to evaluate all selected characteristics according to conformity by evaluating the information obtained for the class.

The researcher further used a pre-sampling process to improve the accuracy of the results by determining the best point of division from the two classes and assigning a binary value to the properties.

Given the limited set of data, the team used the Support Vector classification (SVM) to achieve binary classification. To limit the number of features selected for machine learning, they used only the 30 most suitable features listed by IGAE. For the SVM study, a minimal optimization sequence strategy was used. Using an optimization approach that attempted to reduce the model classification error, different combinations of hyperparameter values ​​were investigated.

The results of machine studies show that voice is uncommon in Parkinson’s disease, as evidenced by the high accuracy of diagnosis in voice discrimination between PD patients and healthy people.

The researchers also performed an ROC analysis to determine the optimal diagnostic values ​​for the difference between HS and PD, primary and secondary stage patients, and middle and advanced stage patients.

The team observed that great statistical accuracy was achieved through machine learning in distinguishing primary-stage patients from HS patients. However, they note that patients with early-stage PD have a subclinical voice defect. They believe that high accuracy in distinguishing primary stage patients from HS reflects the ability of the machine to detect subclinical voice defects in PD, given that 32% of primary stage patients did not have clinically obvious voice defects.

To determine the effect of L-Dopa on voice, the researchers compared the OFF and ON treatment of patients. This study showed that L-Dopa improves voice quality in patients with Parkinson’s disease in the middle stage. Furthermore, their clinical evaluation showed that L-Dopa improved voice less than other motor symptoms, indicating that L-Dopa had a weaker clinical effect on axial symptoms in PD. They observed high diagnostic accuracy compared to patients in treatment and therapy, indicating that L-Dopa had a significant effect on voice in PD.

The researchers hope that their research will encourage the use of machine speech analysis for telemedicine methods in Parkinson’s disease in the future.



Leave a Comment