An artificial intelligence technology that supposedly detects stress in people鈥檚 voices does not deliver consistent results even for individuals taking the test twice in a row, a study has found.
During the second year of the covid-19 pandemic, US healthcare company test 鈥� based on AI technology from 鈥� as a free online tool. Cigna鈥檚 website initially claimed the 鈥渃linical-grade AI technology underlying the test can listen for signs of stress鈥� in people鈥檚 voices based on a 60-second audio sample.
鈥淚f an individual were to take this test right now, and then a minute later, they would likely receive two different scores,鈥� says at Arizona State University. 鈥淚f the tool is not repeatable per participant, how are users to know if the results provided are accurate and reliable?鈥�
Advertisement
Yawer and her colleagues evaluated the StressWaves test by asking 60 people to take the test two times with minimal or no break between. They also had participants fill out a self-reported questionnaire called the , a well-established test for measuring stress, for comparison.
Their experiment showed that the StressWaves test failed to provide repeatable results for each individual and did not align with the self-reported stress questionnaire鈥檚 scoring. 鈥淭he tool did not measure what it claimed to measure, which in this case was psychological stress,鈥� says Yawer.
Yawer鈥檚 coauthors disclosed a competing financial interest as co-founders of the speech analytics company Aural Analytics. The research is published in a peer-reviewed journal.
, co-founder of Ellipsis Health, described the study of his company鈥檚 technology as having a small sample size and coming from a competitor with a conflict of interest. He also cited 鈥渘umerous peer-reviewed papers鈥� showing that his company鈥檚 AI can in people鈥檚 voices.
But Aratow acknowledged an ambiguity in the description of the AI as a 鈥渃linical-grade鈥� technology for stress detection. He says StressWaves is an adaptation of the company鈥檚 AI that is designed to detect stress, 鈥渂ut because stress is not a clinical construct 鈥� unlike anxiety and depression 鈥� this specific use case is not clinical-grade鈥�, he says.
Since 快猫短视频 reached out, Cigna鈥檚 website has been updated to remove the 鈥渃linical-grade鈥� phrasing.
鈥淲e stand by our decision to provide tools such as this, which seek to support mental well-being, during a period in history when the need was acute,鈥� says a Cigna spokesperson.
Voice-analysis AIs could be useful in healthcare if they are properly evaluated in a clinical setting, says at Imperial College London.
鈥淚t鈥檚 really good that researchers are going out and evaluating these tools,鈥� says Coppock. 鈥淲hat they found is concerning, but I think [it] is quite [common] in the space of audio classification.鈥�
Scientific Reports
