The costs of Medicare services are disproportionately distributed within the beneficiary population. For example, 60% of Medicare’s 2015 reimbursement to providers participating in Duke Connected Care (DCC) was for care of just 10% of the total population. About a third of these beneficiaries were persistently high-cost in 2014, and many others live with long-term serious illnesses. In addition, 10% of Medicare’s 2015 reimbursement to DCC providers in our Medicare Shared Savings Program (MSSP) was for care furnished to beneficiaries in their last 90 days of life.

Palliative care (PC) medicine provides specialized and comprehensive care for individuals with serious illness. Though hospice care is a type of palliative care, PC extends far beyond the subset of patients nearing end of life. Instead, PC seeks broadly to palliate symptoms, address psychosocial needs and assure that providers deliver care that is consistent with the goals and wishes of seriously ill patients and their families, regardless of life expectancy.

This RFA proposes to develop a predictive model that would enhance DCC’s ability to identify patients appropriate for this type of care. Greater efficiency in patient identification and targeting expands the capacity to engage with patients by reducing time spent assessing patients who will not benefit from this intervention.

DCC proposes the development of machine learning models to identify, stratify and visualize patients with unmet symptomatic needs, in the setting of progressive, life-limiting or serious illness and high rates of modifiable care utilization. Critically, this project will merge Medicare claims data for the DCC MSSP population with clinical data available in the new Epic Caboodle data warehouse.

Ultimately, this project will enable DCC to enhance the quality and value of care by achieving greater concordance between patients’ goals and the care they receive, improve the quality of life reported by patients and families and reduce unwanted utilization of costly health care services. Without good analytic tools to identify a useful cohort of patients who need this intervention, the required manual effort would be unmanageable with current staffing levels.