Algorithm to ID Voice Changes May Aid in Early Detection of Parkinson’s
AI tool uses phone camera to detect change in voice, facial features
A new machine-learning algorithm — a type of artificial intelligence or AI — that integrates facial expression and voice recordings using just a smartphone camera could help clinicians in the early detection of Parkinson’s disease, a study has found.
Combining data on changes in voice and facial features with age and sex allowed the algorithm to distinguish people at the early stages of Parkinson’s from aging healthy individuals.
“These results provided the evidence to show that voice and facial expressions analyzed with a deep-learning classifier could effectively discriminate between patients with early-stage [Parkinson’s disease] and control individuals” without the neurodegenerative disorder, the researchers wrote, adding, “This analysis mimicked the real-world situation in community screenings.”
If further developed, this technology could have the ability to aid in the early detection of Parkinson’s without the “often expensive” wearable sensors needed for assessing patients’ motor function, the researchers said.
“Voice and facial expression analysis, which only requires a webcam or a smartphone with a camera, is a convenient, relatively affordable tool for detecting [Parkinson’s] … [that] may benefit patients … that live in underdeveloped areas without access to a movement disorder specialist,” the team wrote.
Voice, facial changes mark early Parkinson’s
The study ,“An integrated biometric voice and facial features for early detection of Parkinson’s disease,” was published in the journal npj parkinson’s disease.
Motor symptoms, including tremors and slow movements, are the hallmark of Parkinson’s disease. They occur due to the loss of neurons, or nerve cells, that produce the neurotransmitter dopamine, a chemical messenger essential for muscle control. Levodopa, a precursor to dopamine, and its derivatives have long been one of the gold standards for Parkinson’s treatment.
However, patients often develop non-motor symptoms, such as reduced facial expressions and voice changes, before the classic motor symptoms. Such voice changes include lower volume, slower tempo, frequent pauses, and shortened speech. Facial masking, officially hypomimia, can be another early symptom.
While these changes can be seen as early as five years before a Parkinson’s diagnosis, they are “often ignored and considered normal aging phenomena, which may delay the diagnosis and optimal treatment of [Parkinson’s],” the researchers wrote.
Early detection of these changes could help detect Parkinson’s sooner and identify patients in need of medication to prevent disease progression.
In contrast to the evaluation of motor symptoms, which involves complex sensors, assessing voice and face movement can be easily done with a camera, either a webcam or a smartphone, according to researchers. Moreover, they note, this could be done remotely.
To test this, a team led by researchers in Taiwan used smartphones to record both voice and facial expressions of participants while reading an article. The team then used machine-learning algorithms to analyze the recordings.
In total, the analysis included 186 patients and 185 healthy participants, who served as controls. Among the patients, 119 had early and 67 had late Parkinson’s. Ages were matched for controls and early stage Parkinson’s patients.
Testing algorithms among patients
First, the team used the results from 112 patients during “on” time and 111 controls to train the machine-learning algorithms. “On” time occurs when levodopa or its derivatives are working well and symptoms are controlled; “off” time occurs when treatment effects wear off and symptoms return before another dose of levodopa can be taken.
To assess the algorithms’ effectiveness, they tested them against a validation group, comprised of patients during their “off” times or treatment-naïve patients (total 74) and an equal number of controls. The team tested a total of nine different algorithms.
Because patients in “on” time are harder to differentiate from healthy individuals, the researchers reasoned that algorithms trained with it would show “optimal diagnostic performance in identifying drug-naïve patients with [Parkinson’s] from healthy, aged individuals, in a real-world situation.”
Voice analysis results showed that Parkinson’s patients took longer to read the article, and paused more while reading. Also, the patients’ voice pitch and volume were reduced compared with the healthy controls. Facial expression analysis revealed that Parkinson’s patients had a significantly reduced eye-blinking rate compared with controls.
In the training group, combining the information of both voice and facial expression, along with age and sex into the algorithm was able to discriminate early-stage Parkinson’s patients from controls with an area under the receiver operating characteristic (AUC) diagnostic value of 0.85.
The AUC is a statistical test of how well a given measurement can distinguish between two groups. AUC scores can range from 0.5 to 1, with higher values reflecting a better ability to tell the two groups apart.
Next, they tested whether an algorithm incorporating all these parameters showed diagnostic potential when used on treatment naïve or on “off” time patients. Again, the AUC diagnostic value reached 0.90, confirming the algorithm’s effectiveness.
Overall, these findings demonstrate that specific facial expression and voice changes could be used to accurately diagnosed Parkinson’s with a simple video recording test, according to researchers.
“Our results showed that the integrated biometric features of voice and facial expressions combined with deep- learning classifiers could assist the identification of early-stage PD patients from aged controls,” the team concluded.
About the Author
