Accountable Care Organizations (ACOs) involve groups of healthcare providers
who voluntarily come together to deliver coordinated, high-quality care to aligned
beneficiaries. Many ACOs, such as the Medicare Shared Savings Program and the
ACO REACH program, can participate in alternative payment models that differ
from the prevalent Fee-for-Service model. In these alternative payment models,
providers and payers share financial risk to align the ACOs’ financial incentives
with the dual aims of reducing the total cost of care and improving the quality
of care. In other words, ACOs could profit by keeping their patients healthy and
preventing unnecessary hospitalization. However, to make this financial structure
work as intended, there needs to be a Risk Adjustment (RA) model to change
reimbursement proportional to a beneficiary’s risk; otherwise, ACOs may enroll
only healthy patients, i.e., adverse selection. While most ACOs adopt RA models
for this reason, the original RA methodology has mostly stayed the same over
the last several decades. As a result, some ACO participants have found ways to
“game” the system: to receive disproportional payments for the risk they bear.
To mitigate the waste, the federal government has added various post-adjustment
mechanisms, such as mixing the risk-adjusted benchmark with historical spending,
adjusting by a coding intensity factor, capping risk score growth rate, and
incorporating health equity incentives. Unfortunately, those mechanisms build
on top of each other in nonlinear and discontinuous ways, causing their actual
effects - and efficacy - to be difficult to disentangle and evaluate. In this paper,
we will summarize our lessons from operating one of the most successful ACOs in
the nation to help rebuild the RA model based on a data-driven approach. Next,
we outline the characteristics of an ideal RA model. Then, we propose a new one
that addresses such requirements, eliminating the need for a multi-step process
involving nonlinear and discontinuous staging. Finally, we provide experimental
results by applying this model to our ACO data and comparing them with the
current RA implementation.