Olfactory Phenotypes as Non-Invasive Biomarkers for Alzheimer's Disease: A Machine Learning Approach - PROJECT SUMMARY/ABSTRACT There are currently 50 million people suffering globally with Alzheimer’s disease (AD). 95% of the population over age 65 is concerned about their dementia risk and 80% are interested in dementia screening. There is a critical need for accessible and cost-effective biomarkers that can be used to identify those on the ADRD continuum – including the asymptomatic stages – not only in research and specialty-care centers, but in community-based and primary care settings as well. This information could dramatically improve referrals for early clinical trial enrollment, the triage process for specialty evaluation, and comprehensive care planning. The methods used must be appropriate for point-of-care, community, or at-home deployment while maintaining accuracy and predictive value. Olfactory (sense of smell) dysfunction (OD), in combination with machine learning (ML) algorithms, is a promising non-invasive biomarker for ADRD. We have previously demonstrated the reliability of the Affordable Rapid Olfactory Measurement Array (AROMA) to objectively measure OD and categorize olfactory phenotypes (patterns of correct and incorrect responses to various odorants and multiple concentrations). AROMA uses essential oils, which are complex blends of odor molecules and may be more reflective of “real world” olfaction than the single chemicals used in most other tests. This is because when scents are encountered in real life, the brain processes and recognizes the odorant combinations making up each complete scent differently from the individual component chemicals. Our research with AROMA in ADRD has shown that AROMA can distinguish cognitively unimpaired (CU), mildly cognitively impaired (MCI), and AD patients from one another. Additionally, olfactory phenotypes were detected using machine learning and differentiated between disease states. Our algorithms had 100% sensitivity, 83% specificity for correctly classifying CU versus MCI/AD. Algorithms tasked with classifying MCI versus AD had 100% sensitivity, 75% specificity. We propose longitudinal testing of CU, MCI, AD subjects (n=324 men and women > 55 years) over 3 years to assess changes in OD, functional status, and neurocognition. A group of neurologic controls will be included to ensure olfactory phenotypes are specific for ADRD. Using traditional statistics and machine learning techniques to examine the relationship of AROMA performance with ATN-biomarkers and clinical markers of disease (Aim 1); define predictive models using AROMA data to predict changes in function and frailty (Aim 2); and develop a streamlined ADRD-version of AROMA using only the scents and concentrations of highest influence (Aim 3). Our long-term goal is for point-of-care olfactory biomarker data, analyzed in real-time by ML algorithms, to be widely accessible to meaningfully inform clinical, research, and caregiver decisions.