Photo credit: www.sciencedaily.com
Digital voice recordings are emerging as useful tools for assessing cognitive health, providing a non-invasive approach to monitor potential cognitive decline. Recent research indicates that analyzing various speech features—such as rate, clarity, intonation shifts, and pauses—can reveal early indicators of cognitive impairments when these factors deviate from typical patterns.
Nevertheless, the use of voice recordings raises significant privacy concerns. These recordings can contain identifiable information, including gender, accents, emotional tones, and nuanced speech traits that may pinpoint individual identities. The introduction of automated processing systems further complicates these issues, increasing the likelihood of re-identification and the inappropriate use of personal data.
A recent study conducted by the Boston University Chobanian & Avedisian School of Medicine introduces an innovative computational framework that utilizes pitch-shifting, a technique that adjusts the pitch of audio recordings, to safeguard individual identities while retaining crucial acoustic features necessary for cognitive assessments.
“By utilizing methods like pitch-shifting for voice obfuscation, we have shown how to reduce privacy risks while maintaining the diagnostic utility of voice features,” stated Vijaya B. Kolachalama, PhD, FAHA, an associate professor of medicine and the study’s corresponding author.
The research team leveraged data from the Framingham Heart Study (FHS) and DementiaBank Delaware (DBD) to apply pitch-shifting at various intensities, incorporating additional modifications such as altering the speed of speech and adding background noise. These adjustments aimed to disguise vocal attributes in responses from neuropsychological evaluations. The effectiveness of speaker obfuscation was evaluated using equal error rate metrics, and the diagnostic potential was analyzed through the accuracy of machine learning models tasked with distinguishing among cognitive states, including normal cognition (NC), mild cognitive impairment (MCI), and dementia (DE).
Utilizing the obfuscated audio files, the investigatory framework successfully identified differences between NC, MCI, and DE in 62% of cases from the FHS dataset and 63% from the DBD dataset.
The researchers emphasize that their findings advance both the ethical implementation and practical use of voice data in medical evaluations, highlighting the critical need to balance patient privacy with the accuracy of cognitive health investigations. Kolachalama remarked, “These discoveries set the stage for establishing standardized, privacy-oriented protocols for the future use of voice-based assessments in both clinical and research environments.” She is also an associate professor of computer science and a founding member of the Faculty of Computing & Data Sciences at Boston University.
Results from this research have been published in the journal Alzheimer’s & Dementia: The Journal of the Alzheimer’s Association.
This initiative received funding from the National Institute on Aging’s Artificial Intelligence and Technology Collaboratories (P30-AG073104 and P30-AG073105), the American Heart Association (20SFRN35460031), Gates Ventures, and multiple grants from the National Institutes of Health (R01-HL159620, R01-AG062109, and R01-AG083735).
Source
www.sciencedaily.com