Artificial Intelligence Can Improve Risk Adjustment Data Validation for Managed Care Plans Serving Duals

Introduction

In prior blogs, we discussed how risk adjustment data validation (RADV) is used by the Centers for Medicare & Medicaid Services (CMS) to ensure the accuracy of payments made to Medicare managed care plans.  The RADV process also helps ensure that appropriate health care resources are available for plans’ members.  This is particularly important for plan members who qualify for both Medicare and Medicaid, known as dual-eligible beneficiaries (referred to as “duals” throughout this blog), as they tend to be sicker and require more health care than Medicare-only beneficiaries.

The current approach to RADV relies on the manual review of medical records to determine the accuracy of information submitted by plans on the health status of their members.  Manual validation is a labor-intensive process.  It involves plans having to ask physicians and other clinicians for medical records, as well as teams of humans having to review the documentation.  The process is complicated by the fact that medical records are in different formats across the various electronic medical record systems that may exist within a plan’s provider network.

Due to the laboriousness of the RADV process, CMS must limit the number of records that can be reviewed.  Plans are also limited in their ability to comprehensively review medical records prior to submitting diagnoses to CMS.  That increases the likelihood plans submit diagnosis codes to CMS that are not properly supported in medical records, which can result in the plan having to return payments to CMS.

The RADV process is even more challenging for Medicare managed care plans whose members are predominantly comprised of duals—i.e., individuals who qualify for both Medicare and Medicaid.  There are over 12 million duals across the country, with more than a third of them enrolled in managed care plans.  Below, we discuss challenges plans may face in the RADV process for duals and how the use of artificial intelligence (AI) can be used to create a more efficient validation process.  First, we highlight key attributes of duals and the types of plans that specialize in coverage for them.

Managed Care and Duals

Duals receive health insurance coverage through two distinct programs—Medicare and Medicaid—that offer different benefits.  Through Medicare, duals generally receive coverage for most preventive, primary, and acute health care services, as well as prescription drugs.  Through Medicaid, duals can receive coverage for services such as long-term care services and supports (LTSS) and certain behavioral health services.  Having to manage these two programs can result in duals being vulnerable to fragmented and poorly coordinated health care.

Adding to the complexity of coordinating benefits is that not all providers participate in both Medicare and Medicaid networks.  Moreover, there is little incentive for Medicare and Medicaid to engage in activities that might benefit the other program. For example, states have little incentive to reduce hospital admissions because Medicare would realize most of the savings. Similarly, Medicare has little incentive to prevent nursing home visits, where Medicaid pays for most of the care.  These discontinuities limit the ability of data sharing across providers, as well as care coordination.  They can also result in inadequate monitoring and documentation of chronic conditions for annual RADV reviews.

Managed care has been viewed as a pathway to more integrated care.  CMS has promoted the managed care model for duals through multiple programs, such that 37% of duals were enrolled in Medicare managed care by 2018, up from 18% in 2009.  The major plan types for duals include:

• Dual-eligible special needs plans (D–SNPs), which are the most widely used plans focusing on duals. D-SNPs are essentially traditional Medicare Advantage (MA) plans that limit enrollment to duals. The level of Medicaid integration required by D-SNPs is limited, as their requirements of coverage for specific Medicaid services, including LTSS and behavioral health, vary substantially across states.  The result is that plans may only cover a limited subset of Medicaid services in some states.  It should be noted that the Bipartisan Budget Act of 2018 includes provisions to improve the coordination of Medicare and Medicaid benefits by D–SNPs beginning in 2021.

• Fully integrated dual-eligible (FIDE) SNPs make up about 10% of D–SNPs. FIDE-SNPs are generally more integrated with Medicaid than regular D–SNPs. For example, they must have a capitated Medicaid contract to provide a range of services that includes LTSS, provide both Medicare and Medicaid benefits through a single managed care plan, and take steps to integrate member materials and other administrative processes.

• Program of All-Inclusive Care for the Elderly (PACE) plans are provider-sponsored plans that serve Medicare enrollees who are 55 or older and need the level of care provided in a nursing home. The program aims to keep people living in the community instead of nursing homes. PACE plans rely on an interdisciplinary health team that uses a model of care based on adult day-care centers, as well as therapeutic and medical services. PACE plans are fully integrated because they provide all Medicare- and Medicaid-covered services. Dual status is not required for PACE enrollment, but virtually all PACE enrollees are full-benefit duals. Note also that PACE plans are not subject to the RADV process.

• Medicare Medicaid Plans (MMPs) currently operate in 9 states as part of a demonstration project through the Centers for Medicare and Medicaid Innovation. MMPs are highly integrated relative to D–SNPs, including FIDE–SNPs, because they provide all or almost all Medicaid-covered services. Also, more of their administrative processes have been combined. Depending on the results of the demonstration, the use of MMPs could be expanded in the future.

The below table summarizes key features and differences of each of the plans.  It should also be noted that duals can also enroll in other types of plans, such as regular Medicare Advantage plans and special needs plans for individuals who live in long-term care institutions or have certain chronic conditions.

Plan Characteristics	Dual-eligible Special Needs Plan (D-SNP): Regular (D-SNP)	Dual-eligible Special Needs Plan (D-SNP): Fully Integrated Dual-Eligible SNP (FIDE-SNP)	Medicare-Medicaid Plan (MMP)	Program of All- Inclusive Care for the Elderly (PACE)
Authorization	Permanent Program	Permanent Program	Demonstration	Permanent Program
States where plan is available	40	10	9	31
Number of plans	400	45	46	126
Enrollment (as of Jan 2019)	1,977,848	184,279	388,098	44,440
Contracting structure	Separate Medicare and Medicaid contracts	Separate Medicare and Medicaid contracts	Single 3-way contract with CMS & state	Single 3-way contract with CMS & state
Level of integration	Varies widely but generally low	High	High	High
Share of enrollees who are partial-benefit duals	28%	<1%	<1%	<1%
Passive enrollment	Allowed, for default enrollment only	Allowed, for default enrollment only	Allowed	Not allowed
Eligible for RADV	Yes	Yes	Yes	No

Source: Medicare Payment Advisory Commission (MedPAC).  Based on 2018 and 2019 Reports to the Congress.

Notes: Figures do not include Puerto Rico. The number of D–SNPs and FIDE SNPs are based on unique combinations of contract and plan number; the number of MMPs and PACE plans are based on unique contracts. The figures for the share of enrollees that are partial-benefit duals are based on enrollment data for December 2016.

RADV Challenges for Dual Managed Care Plans

The current RADV process is labor intensive for more Medicare managed care plans.  The process can be even more intensive for plans that service duals.  Below, we discuss how some of the member and plan attributes of plans that service duals can add to the burdens of the RADV process. 

• Duals are associated with more chronic conditions that require medical record documentation. Duals tend to have worse health status than the general Medicare or Medicaid populations. CMS reports that approximately 70% of duals have three or more chronic conditions compared to half of Medicare-only enrollees; approximately 30% of duals have six or more, compared to 16% of non-duals. Over 40% of duals have at least one mental health diagnosis (compared to 16% of Medicare-only beneficiaries).  Duals are also more likely to be disabled; nearly 40% of duals are eligible for Medicare due to disability, compared to less than 10% of the Medicare-only population.  Because duals are sicker, they are associated with more diagnoses that need to be documented on an annual basis.    

• Plans must obtain medical records for duals from wider array of providers. The poorer health status of duals also leads to higher utilization of health care relative to the Medicare-only population. Not only are duals more likely to use more in total, but they are also more likely to use more of each of the different types of services (e.g., hospital care, post-acute care, medications, etc…). Moreover, about half of duals receive LTSS which is generally not covered by Medicare. The diverse and intense health care needs of duals require interactions with a wide array of health care providers, increasing the chances that medical records will be located across health IT platforms with different formats.  That is, it is likely these platforms will involve different data structures and formatting, not only reflecting the customized needs of the various provider types and health systems (e.g., nursing homes versus hospitals), but also the diversity of personal preferences of how individual clinicians input information on their patients.  

• Plans with partial-benefit duals may not be able to fully integrate care coordination processes. There are two types of duals—“full benefit” and “partial benefit.” Full-benefit duals, comprising roughly 70% of all duals, qualify for the full range of Medicaid services covered in their state, including LTSS.  In contrast, partial-benefit duals receive assistance only with Medicare premiums and, in some cases, assistance with cost sharing.  The presence of partial-benefit duals can create an obstacle to greater integration, even for full-benefit duals. This is because enrolling both groups in the same plan makes it difficult to develop a single care coordination process that oversees all Medicare and Medicaid services (since states have little incentive to help finance the costs of care coordination for partial-benefit duals) or use a single set of integrated member materials (e.g., summary of benefits and provider networks).

• A minority of dual plans are fully integrated which limits access to medical records. The Medicare Payment and Advisory Commission (MedPAC) considers MMPs, FIDA-SNPs and PACE plans to have a high degree of integration. However, the large majority of duals are enrolled in D-SNPs, which generally exhibit limited integration of benefits, including provider networks, across Medicare and Medicaid.  States are not required to contract with D–SNPs to provide any Medicaid services, let alone services such as LTSS.  MedPAC estimates that less than half of full-benefit duals in D–SNPs are in plans where the parent company also operates a Medicaid LTSS plan in the same area.  Plans that do not have integrated Medicare and Medicaid managed care products will have additional burdens in retrieving all relevant medical record documentation of their enrollees.  They may have to coordinate with out-of-network providers and other health plans. 

• Passive enrollment of duals creates care coordination challenges. Passive enrollment has been used in most of the dual managed care programs, particularly in the MMP demonstration. With passive enrollment, beneficiaries are automatically enrolled in a dual plan unless they indicate they do not want to participate in the program.  Plans have reported a couple challenges related to passive enrollment: (1) not being able to locate between 20% to 35% of their enrollees because of outdated contact information; and (2) difficulties conducting assessments when large numbers of beneficiaries were passively enrolled at the same time.  Not being able to quickly assess enrollees will lead to delayed care and potentially underreported diagnosis codes.  To the extent these issues persist, plans may have difficulties efficiently allocating care coordination resources across their members.  For example, many assessments are necessarily conducted by phone.  However, phone interactions may not meet the RADV criteria of a “face-to-face encounter,” potentially leading to insufficient medical documentation of diagnoses, particularly for higher-risk enrollees with more chronic conditions. 

• Limited Resources to Conduct RADV. Many plans serving duals are smaller in size compared to general Medicare managed care plans. Even for dual plans, that are highly integrated, RADV may be challenging due to not having the scale to support a dedicated workforce to facilitate manual medical record reviews.

It should also be noted that managed care plans for duals also often need to submit data to states, as well as CMS, for program integrity purposes.

Potential Future of RADV Audits with Artificial Intelligence (AI) Innovation

Health IT is at an inflection point where it is feasible to eliminate the burden of manual RADV audits. The state of artificial intelligence (AI) and electronic health record (EHR) interoperability provides the capability to implement an automated RADV audit with a staggering amount of productivity improvement.  There are several features of AI that are particularly conducive to assisting Medicare plans serving duals, as detailed in the table below. 

Challenge for Dual Plans	How AI can Help
Duals are associated with more chronic conditions that require medical record documentation. Because duals are often sicker, they are associated with more diagnoses that need to be documented on an annual basis.	AI is scalable. AI can make it possible for plans to review all of their records in minutes on an on-going basis. This is critical for plans that serve duals who have a high prevalence of chronic conditions that may be susceptible to underreporting or inadequate documentation. Moreover, AI encompasses machine learning—the process by which computing systems learn by experience. Hence, more data will lead to better performance.
Plans must obtain medical records for duals from wider array of providers. The diverse and intense health care needs of duals require interactions with a wide array of health care providers, increasing the chances that medical records will be located across health IT platforms with different formats.	AI is adaptable to varying data formats. Preferences for how data is coded into electronic medical records can vary substantial across providers, particularly for different specialties. AI can read in data that is in different formats, including text from medical notes, as well as voice messages and images from radiology and other tests. What really sets AI apart, is that its algorithms are designed to analyze and find patterns in unstructured and unformatted data.
Plans with partial-benefit duals may not be able to fully integrate care coordination processes. The presence of partial-benefit duals can create an obstacle to greater integration, even for full-benefit duals.	AI is capable of working across different EHR systems. Given that many duals are not enrolled in plans that can offer the full spectrum of health care services they need within a single integrated provider network, it is likely that the medical records for duals will rest in multiple EHR systems. Due to its adaptability and scalability, AI technology can facilitate data sharing and care coordination even when providers are operating within different health systems with different health IT platforms and interfaces.
A minority of dual plans are fully integrated which limits access to medical records. Plans that do not have integrated Medicare and Medicaid managed care products will have additional burdens in retrieving all relevant medical record documentation of their enrollees.	AI protects patient information. Not only can AI technology be deployed across health systems, but an AI approach would not require the sharing of any personal health information (PHI) across data systems or other points of contact. In contrast, manual reviews require the exchange of medical records containing PHI via multiple points of contact, which creates data security vulnerabilities.
Passive enrollment of duals creates care coordination challenges. Not being able to quickly assess enrollees will lead to delayed care and potentially underreported diagnosis codes.	AI can be used to efficiently allocate care management tools. AI will not only help address RADV coding inaccuracies prior to when payments are made, but also can be used to help plans allocate more intensive resources to the patients that need them the most. To the degree that dual plans have access to demographic and prior Medicare or Medicaid data on new members, AI can also be used to predict resource-use needs even before initial assessments are taken.
Limited Resources to Conduct RADV. Even for dual plans that are highly integrated, RADV may be challenging due to not having the scale to support a dedicated workforce to facilitate manual medical record reviews.	AI streamlines plan administrative burdens. AI reduces the need for manual record reviews. Instead that staff can remain focused on care management activities. Consequentially, plans will save time and money on manual record reviews. AI technology can also handle increasing enrollment and adjustments in compliance rules without requiring a huge team of internal staff to keep pace with the changes.