Unplanned hospital and skilled nursing facility (SNF) admissions among the Medicare population negatively impact the U.S. health care system. Unplanned admissions are costly, unnecessary, and can result in serious health consequences for patients. These visits reveal that patients may have received poor care during a prior hospital stay or during care coordination in post-acute care situations. It exposes patients to medical risk and can result in an adverse event.
Adverse events include outcomes such as a patient contracting a hospital-acquired infection. Predicting both unplanned hospital and SNF admissions and adverse events before they occur is the goal of the CMS AI Health Outcomes Challenge.
Unplanned Admission: When an individual arrives at a hospital or SNF due to an urgent and/or unexpected condition.
Adverse Event: A negative consequence of care resulting in unintended injury or illness.
Physicians need actionable data and more accurate predictive capabilities so that they can provide appropriate resources to the highest risk patients, at the right time.
Using CMS Claims data for Medicare Part A (hospital) and Medicare Part B (professional services), this challenge ask participants to leverage deep learning AI models to identify unplanned hospital and SNF (skilled nursing facility) admissions and adverse events before they occur. CMS is looking for solutions that build transparency and trust with clinicians and patients, relaying output information back to human users in explainable and comprehensible formats.
The solution could be used to drive interventions such as care management or home visits according to individual beneficiary risk, while also providing valuable insight to the Innovation Center on the success of individually-tailored interventions for Medicare beneficiaries. This information, in turn, could be used to develop new Innovation Center models that could result in more effective, more efficient care for all Medicare beneficiaries.
There are different predictive outputs that an AI solution can produce to effectively work with clinicians to predict unplanned hospital admissions and adverse events. Please note these are a few examples of outputs, but entries need not be limited to the following approaches.
- Beneficiary Risk Classification: The AI model will classify risk stratification groups (low, medium, high), and bucket each beneficiary into a risk group.
- Beneficiary Risk Clustering: The AI model will cluster beneficiaries into risk stratification groups.
- Event Prediction Probability Score: The AI model will predict the probability of an event occurring such as an unplanned hospital admission or adverse event and deliver an associated confidence interval justifying the strength or weakness of the prediction.
The following information provides an overview of the datasets used in the challenge and how these datasets must be acquired and used throughout the course of the challenge.
- A summary of the Medicare Part A and B Limited Data Det (LDS) for competition use; and
- A summary of the Data Use Agreement and LDS request process
In Stage 1, the participants may request 5 years of data, from approximately 2008 through 2012.
In Stage 2, the finalists may request an additional, continuous 5 years of data for the same set of beneficiaries, from approximately 2013 through 2017.
Summary of Medicare Part A and B Datasets
The Medicare Part A and B limited data set (LDS) is expected to contain continuous data for approximately 2.5-3 million beneficiaries.
The LDS contain beneficiary-level health information but exclude specified direct identifiers. The LDS is comprised of standard analytical files based on institutional and non-institutional claim files.
These files contain beneficiary level claims data submitted by the institutional facility provider and include a unique encrypted beneficiary identifier which is consistent within the files, across the files and over time. Physician identifiable numbers are encrypted. The data files contain provider number, claim from and through date, diagnosis and procedure information, revenue center codes, and cost and payment information
|Institutional File||Care Type Description|
|Home Health Standard||Home health care includes a wide range of health care services that can be given in a patient’s home for an illness or injury.|
|Hospice Standard Analytic File||Hospice is a specialized type of care for those facing a life-limiting illness, their families, and their caregivers.|
|Inpatient Standard Analytic File||Inpatient refers to a patient who stays in a hospital while under treatment.|
|Outpatient Standard Analytic File||Outpatient refers to a patient who receives medical treatment without being admitted to a hospital.|
|Skilled Nursing Facility Standard Analytic File||Nursing care provided by registered professional nurses, bed and board, physical therapy, occupational therapy, speech therapy, social services, medications, supplies, equipment, and other services necessary to the health of the patient.|
Please review each institutional file’s data dictionary to get an in-depth understanding of the file’s variables, variable descriptions, possible values, field names, and data types.
Data Dictionary Link:
These files contain beneficiary level claims data submitted by the non-institutional provider (individual physician or provider, durable medical equipment supplier). They include a unique encrypted beneficiary identifier which is consistent within the files, across the files and across time. Physician identifiable numbers are encrypted. The files contain place of service code, provider specialty code, diagnosis and procedure information, claim from and through dates, cost and payment information, units codes, etc.
|Institutional File||File Type Description|
|Carrier Standard Analytic File||Carrier’s can be physician or supplier claims with information by date of service.|
|Durable Medical Equipment Standard Analytic File||Durable Medical Equipment consists of, but is not limited to, wheelchairs (manual and electric), hospital beds, traction equipment, canes, crutches, walkers, kidney machines, ventilators, oxygen, monitors, pressure mattresses, lifts, nebulizers, bili blankets and bili lights.|
Data Dictionary Link
Summary of the Data Use Agreement (DUA) and Limited Data Set (LDS) Request Process
Participants will be required to follow the CMS process for requesting the LDS and sign a DUA with CMS in order to receive a claims data set that encompasses Medicare FFS Parts A/B data for a random 5% sample of Medicare beneficiaries.
The DUA will include specific rules and requirements to receive, store, and protect the LDS.
Please note, Launch Stage participants should not apply to purchase any additional CMS LDS for their own use in the competition as CMS will offer the LDS for the competition free of charge to the selected participants in Stages 1 and 2. If Launch Stage participants apply to purchase additional LDS for their own use prior to their selection as a participant, CMS will not reimburse the costs of the additional LDS. CMS will provide further guidance to Stage 1 and Stage 2 participants on data use issues at a later date.
Stage 1 and 2 participants may not augment or link the LDS at the beneficiary level for competition use, but may use and link other de-identified public data sets to the LDS (e.g., Census data). Participants are not required to use other de-identified public data, however, they will be evaluated on their ability to identify other data sets and types of information that will be useful to further refine their solutions following the competition.
Participants will only be allowed to use the LDS for the purposes of the challenge and will be expected to destroy the files and submit a Certificate of Disposition (COD) form to close their DUA once the competition is complete. While this LDS will not be offered to Launch Stage participants, they may refer to the following links to learn more about the LDS structure, to inform their entry.
Key Resource Links