Late-life depression is a large public health problem that will only escalate with population aging in decades ahead. Depressive symptoms are well known to worsen symptoms of comorbid medical problems, pain and anxiety, causing a spiral of functional decline, disability, and loss of quality of life for older people. Rates of depression are higher among older adults who require health related assistance, including 38% of home health care, 23% of residential care, and 49% of nursing home care recipients. Despite its frequency, recognition of depression as a treatable condition is low. Innovative new approaches are needed to easily detect depression using objective markers that don't rely on subjective appraisals by older people, families, or providers. Vocal biomarkers of depression have been successfully identified in adults who are 65 years of age or younger, but no studies have examined if these methods may be applied with older people. This gap is significant because aging vocal qualities are confounded by changes associated with primary aging, sex differences, health conditions and medication use in older people. The purpose of this study is to identify vocal markers of late-life depression that will allow later development of individualized depression detection monitoring devices that overcome attitudinal barriers by providing objective biological information. Two specific study aims are to evaluate vocal patterns in older adults that are (1) distinct from and comparable to those identified for younger adults, and then (2) use machine learning to study and construct algorithms to predict depression. The observational, repeated measure design will enroll 250 participants aged 65 years and older from community settings. Audio recorded interviews will assess depression level using the PHQ-9, reading using the phonetically balanced Rainbow and Grandfather Passages, and free speech using a set of mood-neutral open-ended questions. Two interviews will be conducted for each participant using PHQ-9 scores to determine timing of follow-up interviews of either 2 weeks (PHQ-9 less than 10) or 8 weeks (PHQ-9 of 10 or greater). Interview segments will be separated using timestamps and vocal features will be extracted using openSMILE software. Each participant will serve as his/her own control in linear panel analysis. Support Vector Machine, Lasso regression, and K- means clustering will be used to predict presence and levels of depression based on PHQ-9 scores. Model performance will be assessed using the percent correcting predicted, false positive rate, and false negative rate. Outcomes from this study will guide additional research and the development of in- home monitoring and other devices that assure easy, objective identification of depressive symptoms that advance early, effective self-care approaches and treatments.
Late-life depression interacts with other medical problems in a way that makes symptoms of all conditions more severe, leading to worsening health, loss of function and quality of life, and increasing disability. This devastating sequence may be avoided by identifying vocal biomarkers of depression that will hasten identification and treatment that averts unnecessary suffering. The goal of our study is to identify vocal patterns that are ?biomarkers? of late-life depression so that easy-to-use monitoring systems and devices can be later developed to promote well-being and health.