Can AI detect depression based on the way a patient talks?

In 2016, the U.S. Preventive Services Task Force (USPSTF) recommended screening for depression in the general adult population when adequate systems are in place to ensure accurate diagnosis, effective treatment and appropriate follow-up. Despite this recommendation, it’s estimated that roughly 4.1% of primary care patients are screened for depression. A new study, published in Annals of Family Medicine, evaluated a machine learning (ML) tool designed to analyze patients’ free-form speech to detect signs of a depressive episode, specifically, the Kintsugi Voice Biomarker Technology.

From February 2021 through July 2022, 14,898 U.S. adults completed the Patient Health Questionnaire-9 (PHQ-9) and recorded a voice response to the prompt, “How was your day?” for at least 25 seconds. From that free-form speech, the ML technology correctly identified depression in 71% of patients who had it and correctly ruled out depression in 74% of patients who did not have it. The tool flagged roughly 20% of cases as uncertain, recommending further evaluation by a clinician.

“Findings from this study suggest that harnessing ML technology to evaluate speech for the detection of signs of a depressive episode is effective compared with the PHQ-9 at a cutoff score of 10,” the authors of the study wrote. “This study supports that the use of ML technology as a clinical decision-support tool might be a step toward universal depression screening, a primary care objective recommended by the [USPSTF].”

Although the technology is not a replacement for primary care evaluations, it does have the potential to serve a crucial role as a decision-support system in primary care screenings. The authors of the study clarify that the results of the study are meant to demonstrate the efficacy of the ML device when used by qualified clinicians, particularly primary care physicians/family medicine doctors, as an adjuvant tool to assist in depression screenings.

Since the tool can operate in the background of a visit, analyzing physician-patient conversations, it has the potential to increase depression screening rates without contributing unnecessary burden to physicians or patients. Despite early results that are promising, authors acknowledged the need for additional studies aimed at weighing the acceptability of augmenting primary care workflows with ML technology and determining other potential conditions that might influence depression voice biomarker analysis.