Researchers have created machine learning tools that utilize acoustic voice signals to screen for anxiety disorders (AD) and major depressive disorder (MDD) that occur simultaneously. The research team conducted interviews with participants both with and without these co-existing conditions using a secure telehealth platform. During the study, participants took part in a semantic verbal fluency test, where they had to list as many animals as they could within one minute. From the recordings, the researchers extracted acoustic and phonemic characteristics and employed machine learning methods to differentiate between those who had comorbid AD/MDD and those who did not. The findings reveal that a brief one-minute semantic verbal fluency test is a reliable method for screening these disorders.
The mental health crisis in the United States is well-known. By 2021, 8.3% of adults were diagnosed with major depressive disorder (MDD), and 19.1% were experiencing anxiety disorders (AD). The situation worsened during the COVID-19 pandemic. Despite the high rates of AD and MDD, the percentage of those receiving diagnoses and treatment remains low—36.9% for anxiety disorders and 61.0% for depressive disorders—due to various social, perceptual, and structural challenges. Automated screening tools could play a significant role in addressing this issue.
In an article published in JASA Express Letters by AIP Publishing on behalf of the Acoustical Society of America, a team of researchers developed machine learning algorithms capable of screening for comorbid AD/MDD through acoustic voice signals derived from a one-minute verbal fluency test. This research team consists of experts from the University of Illinois Urbana-Champaign, the University of Illinois College of Medicine Peoria, and Southern Illinois University School of Medicine.
Mary Pietrowicz, one of the authors, explained, “This research stemmed from the understanding that people suffering from anxiety disorders and major depressive disorder frequently experience delays in both diagnosis and treatment. Finding voice signals that reflect various psychiatric, neurological, and even upper-gastrointestinal health issues encouraged us to delve deeper into AD/MDD.”
Each of the disorders—AD, MDD, and the combination of both—has distinct acoustic characteristics. Identifying comorbid AD/MDD is particularly difficult since the acoustic indicators for anxiety and depression often conflict.
Pietrowicz noted, “A lot of existing research fails to recognize these differences and doesn’t focus on the specific traits of comorbid AD/MDD.”
In their study, Pietrowicz and her colleagues interacted with female participants who either exhibited or did not exhibit comorbid AD/MDD. They conducted the recordings via a secure telehealth platform and had the participants perform a semantic verbal fluency test in which they needed to name as many animals as possible against a clock.
From the recordings, the researchers analyzed acoustic and phonemic features and utilized machine learning techniques to effectively differentiate between participants with and without comorbid AD/MDD. The outcome demonstrated that a simple one-minute semantic verbal fluency test can serve as a dependable screening tool for these conditions.
Pietrowicz added, “The group suffering from AD/MDD generally opted for more basic words, showed less variability in the length of their phonemic responses, and exhibited lower levels and inconsistency in phonemic similarity.”
Looking ahead, Pietrowicz intends to investigate the biological mechanisms underlying these findings and aims to improve the screening model. However, she emphasized that creating an effective diagnostic tool will necessitate a substantial amount of data from various demographics and conditions.
“Our current priority is to broaden the scope, diversity, and methods of our data collection, all while implementing cutting-edge analytical techniques to improve the accuracy of our model and enhance our insights into these acoustic signals,” Pietrowicz concluded.
