Decision support tool that uses novel feature extraction and machine learning approaches for early identification of Alzheimer's patients who are candidates for palliative and/or hospice care.

Nearly 6 million Americans had Alzheimer’s disease (AD) in 2019, a figure that will double by 2050. This progressive disease can be difficult to manage at home for patients, and for their families and caregivers. Timely and appropriate referral to palliative care can help. The medical literature suggests that as Alzheimer’s progresses, palliative services can help improve quality of life and reduce the total cost of care (currently on track to exceed $578 billion/year by 2050). Benefits include avoiding unnecessary hospitalizations, providing support for in-home caregivers, and alleviating distressing symptoms. Despite these benefits, too few Alzheimer’s sufferers are referred to palliative care services and, when they are referred, it is often too late. Studies suggest that existing legacy tools do a poor job of timely identifying palliative care candidates among AD patients. Even in cases where palliative care suitability is clear, medical professionals often lack the time required or are uncomfortable discussing advance care planning. The aim of this research is to improve timely and appropriate palliative care referrals for AD patients. We plan to develop and validate a novel clinical application of cutting-edge machine learning techniques to identify AD patients for earlier palliative care intervention. We will predict 12-month mortality as a proxy for palliative care appropriateness, building upon previous research but also addressing its limitations. Our specific objectives are to (a) utilize six years of CMS national Medicare claims data to generate the most detailed analysis to date of AD patient utilization history, (b) develop a rich feature set of relevance to AD disease progression, encompassing medical utilization, clinical, functional, socio-behavioral, and demographic dimensions, (c) train and evaluate an array of supervised ML classifiers to predict 12-month mortality, and (d) develop a risk stratification score that may be used clinically to rank-order AD patients in terms of their appropriateness for referral to palliative care. Our risk stratification score will combine both the clinical appropriateness for palliative care (i.e., the need) and the likelihood of a successful referral (i.e., the feasibility). The aforementioned “feasibility” element is especially novel, and potentially represents a “missing link” that has hindered prior research. If our Phase I effort is successful, the outcome will be a validated novel data-driven approach to risk-stratify AD patients for earlier palliative care intervention. In a future Phase II proposal we would seek to demonstrate clinical efficacy by productizing and deploying the risk stratifier into a real-time clinical decision support system and prospectively evaluating this methodology in a clinical environment. Our proposal is responsive to the NIH/NIA’s mission “to conduct research leading to the development of innovative products and/or services that may advance progress in …caring for and treating AD/ADRD patients.” The planned work is specifically aligned with NIA Priority Topic DBSR-C, which calls for innovations to support “evidence-based methods, technologies, and interventions to reduce the burden of caregiving for persons with AD.”