Parkinson’s disease (PD) is a neurological condition that causes tremors, stiffness, and difficulty walking, balancing, and coordinating. Dopamine levels diminish due to nerve cell destruction in the brain, resulting in Parkinson’s symptoms.
PD patients frequently complain about variable impairment of voice emission. These patients may experience speech problems even at the prodromal stage of the condition. Symptoms of Parkinson’s disease normally appear gradually and worsen over time, eventually leading to severe voice impairment in more advanced stages of PD.
The current clinical voice assessment methods in PD are solely on qualitative evaluation. This includes spectral analysis that reveals various irregularities in certain voice qualities in individuals with PD, including the reduced fundamental frequency and harmonics-to-noise ratio and increased jitter and shimmer. However, the human voice is complicated that comprises high-dimensional data based on an exponential number of features.
As a result, in addition to an independent examination of specific voice features, more advanced techniques capable of analyzing and dynamically combining high-dimensional datasets of voice features are required that would accurately classify the objectives of voice samples in PD.
Machine learning methods have enabled the automatic classification of voice impairment in various neurologic illnesses with high accuracy. However, only a few exploratory studies have been reported on the use of machine learning analysis in PD to date. It’s important to see if machine learning can distinguish between patients in different stages of the disease to see if it can recognize the effect of disease severity.
A new study by researchers in Italy and Jordan studied the voice of Parkinson’s disease patients in a large and clinically well-characterized cohort. This study is the first to classify voice in Parkinson’s disease patients based on the stage and severity of the disease and the effect of chronic L-Dopa medication. All diagnostic tests were evaluated for sensitivity, specificity, positive and negative predictive values, and accuracy.
The IRCCS Neuromed Institute and the Department of Systems Medicine at Tor Vergata University in Rome, Italy, recruited participants for the study. The participants included 115 individuals with Parkinson’s disease and 108 age-matched healthy subjects (HS). All of the patients were native Italian speakers who did not smoke. There were no reports of bilateral or unilateral hearing loss, respiratory diseases, or any non-neurologic disorders affecting the voice cords among the subjects.
Participants were given a specific speech task to perform with their regular voice strength, pitch, and quality for voice recordings. The task included a persistent emission of a close-mid front unrounded vowel.
The researchers used OpenSMILE (an open-source toolkit for audio feature extraction and classification)to pre-process each speech sample for feature extraction. They collected 6,139 voice attributes from each speech sample. They used the Correlation Features Selection algorithm (CFS) to find (uncorrelated) voice qualities substantially linked with the class. As a result, the original dataset was stripped of duplicated and/or useless information. The Information Gain Attribute Evaluation (IGAE) methodology, based on Pearson’s correlation method, was then used to rank all of the selected features in order of relevance by assessing the information gained for the class.
The researcher further employed the discretization pre-process to improve the accuracy of the results by identifying the best dividing point from the two classes and assigning a binary value to the features.
Given the study’s limited dataset, the team used a Support Vector Machine (SVM) classifier to achieve a binary classification. To limit the number of selected features required for the machine learning study, they employed only the first 30 most relevant characteristics as ranked by the IGAE. The sequential minimal optimization strategy was used to train the SVM. Using an optimization approach that tries to reduce the model classification error, different combinations of hyperparameter values were examined.
The machine learning results demonstrate that voice is abnormal in Parkinson’s disease, as evidenced by high diagnostic accuracy in voice discrimination between PD patients and healthy people.
The researchers also performed ROC analysis to determine the optimal diagnostic cut-off values for differentiating between HS and PD, early-stage and mid-advanced-stage patients, and mid-advanced-stage patients on and off medication.
The team observed that great statistical accuracy was gained by machine learning in distinguishing early-stage patients from HS patients. With this, they remark that early-stage PD patients have subclinical voice impairment. They believe that the high accuracy in distinguishing early-stage patients from HS reflects machine learning’s ability to discern subclinical voice impairment in PD, given that 32% of early-stage patients did not have a clinically overt voice impairment.
To determine the effect of L-Dopa on voice, the researchers compared patients’ OFF and ON therapy. This study found that L-Dopa improves voice quality in patients with mid-stage Parkinson’s disease. Furthermore, their clinical evaluation showed that L-Dopa improved voice less than other motor symptoms, indicating that L-Dopa has a weaker clinical effect on axial signs in PD. They observed high diagnostic accuracy in comparing patients on and off therapy, indicating that L-Dopa has a significant effect on voice in PD.
The researchers hope that their research will encourage the use of machine learning speech analysis for telemedicine techniques in Parkinson’s disease in the future.