AI-based biomarker tool may serve as a promising aid for depression screening

Key takeaways:

• The AI-based tool correctly identified depression in around 71% of patients.
• It also correctly ruled out depression in close to 74% of patients.

An AI model demonstrated efficacy at identifying moderate to severe depression through speech patterns, according to a cross-sectional analysis published in Annals of Family Medicine.

The model also showed an ability to identify or rule out depression in certain diverse groups.

The U.S. Preventive Services Task Force currently recommends that clinicians screen all adults for depression and adults aged younger than 65 years for anxiety disorders.

However, some estimates suggest that depression screening rates are under 4% in outpatient settings, Alexa Mazur, BA, from Kintsugi Mindful Wellness, Inc., and colleagues noted.

“Even when identified to undergo screening, patients with depression are included [less than] 50% of the time,” they wrote. “Thus, there is a substantial opportunity and need to improve primary care screening for depression.”

In the study, the researchers examined the efficacy of an AI-based voice biomarker tool (Kintsugi Voice, Kintsugi Mindful Wellness, Inc.) using several predictive measures and the voice samples of 14,898 participants from the United States and Canada.

Participants had answered the question, “How was your day?” with at least 25 seconds of free-form speech. Kintsugi Voice then assessed voice biomarkers tied to depression, such as hesitations, frequent and longer pauses and speech cadence.

The researchers then compared voice biomarker results with self-reported results from the Patient Health Questionnaire-9 (PHQ-9).

The tool had three outputs, which included:

• signs of depression not detected;
• signs of depression detected; and
• further evaluation recommended for uncertain cases.

The analysis included 10,442 voice samples to train the model, with 4,456 samples used in a validation set.

Mazur and colleagues found that the model showed a sensitivity of 71.3% (95% CI, 69-73.5), meaning it correctly identified depression in 71% of people who had it.

The tool also had a specificity of 73.5% (95% CI, 71.5-75.5), correctly ruling out depression in close to 74% of people who did not have it. Researchers calculated positive predictive and negative predictive values for Kintsugi Voice and 69.3 (95% CI, 67.1-71.5) and 75.3 (95% CI, 73.3-77.2), respectively.

Among subpopulations, the model showed highest sensitivity among Hispanic (80.3%; 95% CI, 72.6-86.6) and Black (72.4%; 95% CI, 64-79.8) populations, whereas specificity appeared greatest among Asian or Pacific Islander (77.5%; 95% CI, 72.8-81.8) and Black (75.9%; 95% CI, 69.3-81.7) populations, “which all had wider CIs relative to the full sample and white subpopulation,” the researchers noted.

Sensitivity and specificity also significantly varied among men and women. For example, researchers determined sensitivity of 74% (95% CI, 71.4-76.5) and 59.3% (95% CI, 54-64.4) for women and men, respectively, while reporting specificity of 68.9% (95% CI, 66.2-71.4) for women and 83.9% (95% CI, 80.8-86.7) for men.

Mazur and colleagues also pointed out that those aged younger than 60 years had a specificity (71.8%; 95% CI, 69.6-73.9) and sensitivity (71.9%; 95% CI, 69.5-74.2) with narrower CIs vs. those aged 60 years or older.

The researchers explained that many mental health diagnostic and screening tools have performances of 60% to 90% for both sensitivity and specificity, meaning that Kintsugi Voice’s performance vs. PHQ-9’s “suggests it might be effective for an [machine learning] device to assist in screening and identifying individuals with depression.”

They concluded that an AI tool for depression screening “holds promise because it could increase the proportion of patients screened, without undue clinician clerical burden.”