When you build a data analytics product for ACOs, it is vital that members are attributed to providers. Unfortunately, given the plurality of data vendors and the lack of need in a lot of other cases for this attribution, a lot of the raw data available does not have direct attribution for members. In our healthcare products Makalu and Everest, we needed to make sure we had attribution model correctly set in.

Attribution Methodologies

According to the Milliman Healthcare Reform Briefing Paper, Susan E. Pantely, there are different methods of patient attribution:

Patient-Based vs. Episode-Based Attribution

Patient-based attribution method assigns the costs of each patient to a provider or providers. The attribution covers the overall episodes of care of the patient. This is the most common method of attribution.

Episode-based attribution assigns patient to the provider for a single episode. The episode is the period from onset of symptoms to the completion of treatment. In some conditions, where a patient has suffered from chronic disease like diabetes, the episode is defined in prior such as of 12 months.

Single vs. Multiple Attribution

In Single Attribution method, the patient is assigned to the provider with the highest percentage of services or total cost. The patient is only attributed to the provider if the percentage is above the threshold (which is from 25% to 35%), otherwise the patient is not attributed to any of the provider.
Multiple attribution method is used when single attribution method is insufficient for patient’s care. In multiple attribution method, a patient or episode of patient is assigned to multiple providers.

Prospective vs. Retrospective Attribution

Prospective attribution method assigns members to the provider on historical claim data. The assumption in this method is that the member will use the same providers in the future as they have done in the past. The prospective attribution method provides quality and cost reports on timely basis which will be very helpful for further analysis and planning by the providers. The physicians know in prior that which patient is assigned to him.
Retrospective attribution method assigns patients to the providers on the basis of their actual utilization. But the problem here is that the physician has no knowledge of which patients are assigned to him.

Patient Attribution Models

There are different patient attribution models used by ACOs. The two most commonly used models are:

• Dartmouth Patient Attribution Model
• PGP Patient Attribution Model

Dartmouth Patient Attribution Model

Dartmouth model is the model which maintains longitudinal relationship between patient and provider and prevents high-cost patient being excluded from the ACO.

In this model patients are assigned to the provider based on the analysis of patient’s historical data. The analysis is done by taking into consideration the patient’s visit trend and the specialty of provider according to the claim data from last two years. The providers may be primary care providers, medical specialists or surgical specialists.

Primary care providers are given the highest priority.  If the patient has visited primary care provider even for a single episode, s/he will be assigned to primary care provider even when s/he has visited medical specialist or surgical specialist. If s/he has visited multiple primary care providers, s/he will be assigned to the one whom s/he has visited the most. In case both have same number of visits, s/he will be assigned to the provider with the greatest number of days between the first and the last visit. In case of this being identical too, the provider with latest visit is assigned.

The second highest priority is medical specialist followed by surgical specialist. The assignment procedure is same as that of primary care provider.

PGP Patient Attribution Model

The PGP model is a retrospective patient attribution model. In retrospective method, the providers are unaware about which beneficiaries are being assigned to them until the end of the performance  in order to ensure that all the patients are treated the same.

According to PGP model, the patients are assigned to PGP on the basis of the largest share of outpatient evaluation and management (E&M) services received. As PGP might not be able to provide all type of services to patient, certain E&M services are excluded from attribution methodology (such as emergency department visits).

One of the pitfalls of this method is that the providers are not able to plan future care delivery services according to the needs of the patient as the patients to be assigned are unknown.

According to a finding, the PGPs included in this model provide approximately 80 to 90 percent of the outpatient E&M services for their assigned beneficiaries, and nearly two-thirds of these beneficiaries are retained from year-to-year. It gives some clues to providers for their future plans.

The membership period normally remains for one year after the attribution of patient to provider. The patient can change the provider inside or outside the ACO provider list. In such case, there needs to be appropriate rule to adjust such changes.       

Deerwalk Methodology

As we built the models based on examples of other ACO models, we did an in-depth study of what kind of data is available to us before building a logic for attribution that helps define providers for us.

We take a look at the last two years worth of claims data for a particular member. If a patient has had no valid outpatient E&M visit within the two-year window, the patient would not be assigned. The Dartmouth model is very good in that it uses specialty to determine actual attribution. The retrospective approach introduced by the PGP model helps in the fact that all patients are treated equally. To us, we think including patients who have had an outpatient E&M visit recently means we can have all patients at equal footing, irrespective of how much they have spent.

We take the claim type as professional so that we exclude facilities from showing up on our attribution. If the patient has visited only one provider in the claim period, he is attributed to that provider. If the patient has visited multiple providers, we take the one that the patient has visited the most number of times (based on encounter).

If these cases still get a tie between providers (in case of two providers having same number of visits), we then look for the provider whom the patient has been seeing the longest. The biggest difference between first and last visits then helps determine which provider the patient is attributed to.
If none of the above processes work, we go with the provider the patient has seen most recently.

Challenges in Patient Attribution

There are several challenges in patient attribution and it has been a complex problem. It is because instead of the access on data of patient’s medical services and cost, it is accessed by insurer or other entity who assign patients to the physician.

Another type of challenge is the exclusivity of patients. If a patient is assigned to multiple ACOs, then it is difficult to keep track of who is accountable for the patient. The problem is that if the patient seems to be of high cost, both ACOs try to disassociate the patient from them. Patients often see several physicians for potentially overlapping care. For example, Medicare FFS beneficiaries annually see a median of two PCPs and five specialists working in four different practices (Pham et al., 2007).

Current visit- and procedure-based reimbursement systems pay physicians to treat particular medical problems, not to manage a patient’s overall care. Physician specialists focus on narrow medical issues that are referred to them. Building an attribution model out of this model is fraught with dangers.

Attributing physician responsibility for care is further complicated, in that FFS patients have the freedom to choose their providers; they face few limitations or preapproval requirements on seeing multiple doctors and seeking multiple opinions or treatment options. Since the data shows equal evidence of each visit, our metrics are likely to get compromised.
That said, it is likeliest that we get very close to correct attribution in more cases than not. It is a very close approximation and until we get data that is actually completely correct, we work with this model.

References

• David Axene, John Bertko, Dan Dunn, Bela Gorman, Steve Lieberman, Joachim Roski, Mark Zezza,  (January 2011), ACO Toolkit, Engelberg Center For Health Care Reform | THE DARTMOUTH INSTITUTE
• Jerry Cromwell, Michael G. Trisolini, Gregory C. Pope, Janet B. Mitchell, and Leslie M. Greenwald,    (March 2011), Pay for Performance in Health Care: Methods and Approaches
• Susan E. Pantely, (2011),  Milliman Healthcare Reform Briefing Paper
• G. Hackbarth, ( April 2009), Reforming America’s health care delivery system. Statement to the Senate Finance Committee
• Ingenix , Symmetry episode treatment groups: Issues and best practices in physician episode attribution
• Pham, H. H., Schrag, D., O’Malley, A. S., Wu, B., & Bach, P. B., Care patterns in Medicare and their implications for pay for performance

Download PDF