Early detection and monitoring of Alzheimers Disease and Related Dementias using non-semantic linguistic and acoustic features of speech derived from hearing aids

Abstract Alzheimer’s disease and related dementias (ADRD) are a serious national health concern that affected 5.8 million people in 2020 and are expected to increase by 40% over the next decade. There is evidence that the functional, psychological, pathological, and physiological changes underlying ADRD may emerge many years prior to the clinical manifestation of cognitive symptoms, which is increasing the interest in early detection and monitoring to inform disease prediction and management at both the individual and population level. Given the need for improved measures to understand and treat ADRD, several divisions of the National Institute on Aging have called for improved methodologies for prognosis, diagnosis and/or treatment monitoring of aging related cognitive decline that are more sensitive to early cognitive changes, less costly and noninvasive. Recent advances in technology for hearing aids (HA), speech analysis, and machine learning (ML) present tremendous opportunities to provide cost-effective, user-friendly cognitive measures that can be readily used, or adapted, for persons living in remote, urban, and peri-urban communities. There are over 14 million people in the US that use a HA between the ages of 50 and 85 years-old, which corresponds directly to the study population for the onset and progression of ADRD. These 14+ million HA users are also at 3X greater risk for developing ADRD, which strengthens the justification for a focus on this high-risk population. Phase I successfully identified a set of non-semantic features of voice and speech that can predict mild cognitive impairment (MCI) with an AUC of 0.87 (90% CI 0.82-0.93). The information gained through the Phase I work affirmed the feasibility of using a HA for the analysis of non-semantic linguistic and acoustic features of speech indicative of early changes in cognitive health. The ability to extract voice features in the HA is a key aspect of maintaining privacy for the user outside of clinical or structured conversations (i.e., during the person’s normal activities of daily living). Overall, we believe that passive cognitive monitoring is needed for the long-term study of how ADRD progresses and for improving screening, so that treatment can be provided earlier.