Discover more from Health Tech Stack
HTN Risk Adjustment Discussion Readout
Insights from the live online referral discussion on Tuesday, Feb 8th, 2022
Two years into the pandemic, Zoom fatigue is real. That's why I am even more impressed if so many people show up for an online "mini-conference" and bring their unique perspectives to a fruitful discussion. After the specialist referral discussion last month, we had another online session with the health tech nerds community about risk adjustment in value-based care". The participants had a wide range of backgrounds. Value-based care providers, health plans, actuaries, and a former regulator joined the discussion and shared their thoughts on the topic. In today's post, I will summarize some of the main insights from our conversation.
As these "mini-conferences" are so insightful and exciting, I plan to organize them more regularly. If you're interested in participating and getting the invites, please sign up here. Before jumping into the article below, if you're new to risk adjustment, you can read some of the basics in this article covering the basics of risk adjustment.
Introduction to Risk Adjustment
Daena Bell and Erica Rode from Milliman kicked off our discussion with an introduction to risk modeling - Milliman is a global actuarial and consulting firm with quite some expertise in risk adjustment. It's worth looking a bit at the terminology - at its core, a risk model is a predictive model that assigns a risk score to each individual. The risk model can be used to predict various factors, such as utilization, the likelihood of acute events, or the total cost of care. Risk Adjustment is the process of using these risk scores for a specific use case. Here are some of the most common applications:
Payment arrangements: Risk scores are used to compensate organizations based on their members' risks. This finds application in Medicare Advantage Organizations, ACA Marketplace plans, ACOs performance payments, and other commercial value-based payment models.
Care management: Risk scores can be used to identify members that are likely to experience adverse events - providers and health plans can use these prediction models to design specific interventions for these groups. This is also known as risk stratification.
Quality measurement: To accurately measure a provider's performance, risk scores can be used to remove some of the bias in the patient data. Certain providers see sicker patients, so this needs to be considered when evaluating their outcomes.
Given the advancements in data science over the last decade, the world of risk models is also subject to change. The workhorse for actuaries and risk modelers remains to be the linear regression. The advantage of this model is that it can be easily interpreted and works well on smaller datasets. However, risk adjusters are exploring more complex models such as decision trees and neural networks. These models might be better suited to detect non-linear relationships between input variables and predict outcomes more accurately. But these complex models can also be computationally intense to train and need large data as input. Furthermore, the parameters of these models are harder to interpret.
An important thing to consider for risk model calculation is biased data. If the population used to train the risk model is substantially different from those used for the prediction, the predictions would be biased and less accurate. We discussed the bias in the next section.
Critical voices about the current risk adjustment system
The risk adjustment system put forward by CMS for Medicare Advantage and Direct Contracting has been under quite some criticism lately. Read this article in health affairs if you want to learn more about the controversy. We also had a few critical voices in the group:
CMS-HCC model disease prevalence: There is a potential considerable statistical bias in the risk model that CMS applies, which is only trained on data from fee-for-service Medicare but does not include encounter data from Medicare Advantage. Because fee-for-service Medicare providers don't have incentives to be very careful in their risk coding, the model might underestimate the prevalence of a particular condition. This bias is sometimes called coding intensity bias. A solution for this could be to include MA encounter data into the CMS risk model.
Models can underestimates the risk for underrepresented groups: Risk models can only be as good as the underlying data - groups without access to health care or with little health data overall might not be reflected properly in the government risk model. This means that the compensation for conditions in certain marginalized groups might be too low. This can in turn strengthen the bias in the data, if Medicare Advantage Organizations are less financially incentivized to care for these groups and capture their data. An idea to fix this is to take non-medical data into account. We talked more about this during the social determinants of health part.
Adverse incentives: Risk scores for different diseases may encourage providers to do diagnostic tests that are not necessary. Doctors might be inclined to do a diagnostic test if a particular diagnosis is higher compensated than others, even if the clinical guidelines do not deem it necessary. [I guess we will never get incentives right in health care...]
Some interesting insights related to this criticism came from a former CMS regulator. CMS is aware of some of the issues related to risk adjustment, and they are trying to improve the methodology for risk adjustment. When setting these reimbursement policies, they need to balance the reporting burden with the reporting accuracy. Also, when deciding their risk model, it is not just about predicting costs but also about pushing certain policies.
Social determinants of health in the risk model
Moving on, we discussed how social determinants of health could play a role in risk adjustment or at least in the risk stratification process. Quite a few people are currently working on Social Determinants of Health (SDoH) in health care these days. SDoHs consider the conditions in the environments where people live that affect their health. Factors include economic stability, education, health care access such as access to transportation and availability of doctors, social factors, and environmental pollution. In theory, providers and health plans could use these data points to predict better whether a patient is at risk and design targeted interventions. The group brought up a few challenges here:
Data Collection: There are no established methodologies to collect SDoH data. A common source is surveys via phone, mail, or at the doctor's office. However, these surveys are rarely standardized and are difficult to connect to other survey results. Another potential source of data is publicly available data on a community level, for example, ZIP-level income information, air quality data, school district performance data, or occupational information. For example, an asthma patient in Montana might be less at risk than in downtown Los Angeles. But all this data is not easy to handle at scale, as it comes in different file formats and has varying scope and methods for calculating their metrics.
Turning data into action: Even if data is available and tied back to a patient, it is still unclear how providers can use it for their clinical decision-making. Doctors are generally not trained to take SDoH into account for treatment decisions. Just giving them the raw data will not achieve much, even if it can predict certain conditions. Therefore, SDoH data needs to be analyzed, and specific interventions must be designed, such as access to care programs (like paying for a taxi to and from the doctor's appointment) or food security programs. A great example brought up during the discussion was to work with community institutions such as churches to drive awareness about medication adherence.
Include SDoH into the Risk Adjustment Reimbursement Model: There are thoughts to include the SDoH into the CMS reimbursement models. However, it isn't easy to accomplish. In particular, there are no standards (yet) for capturing SDoH data, and thus they are difficult to codify and include in a nationwide risk model. Also, the administrative burden of requiring physicians to collect SDoH data for their practices might not outweigh the benefits of more accurate risk adjustment. However, it might be a possible backdoor to push an industry-wide standard for SDoH data capture, expanding their use.
Operational challenges with risk gaps
As risk scores play an essential role in value-based care compensation, we continued our discussion by covering the operational challenges of capturing and reporting risk factors. Jay Srivastava from Cityblock shared his views on what a successful risk documentation program would look like.
A critical effort of many stakeholders in the risk adjustment process is to surface risk gaps and suggestions at the point of care. A risk gap means that the patient is likely to fall into a particular risk category but currently is missing the compliant documentation required by CMS. A great example for a risk gap, is if the patient had a chronic condition last year, but it has not been coded this year. Managed care organizations would ideally like to inform the providers about open risk gaps at the point of care so they can provide the appropriate diagnosis documentation. This in turn gets the managed care organization or value-based organization the full risk-adjusted reimbursement for their member, which can be shared with the providers or reinvested if savings are generated. However, this puts an added burden on the provider, and to make this program successful, the following points should be considered:
Motivation and Incentives: As risk adjustment may put some coding burden on the practice, they need the right incentives and motivation to implement the proper reporting processes. For example, if a practice only has 10% of their patients in value-based payment arrangements, there is little incentive to pay much attention to risk reporting. However, suppose the share is closer to 90% or there exist incentive programs from payers that pay bonuses on reporting accuracy, the practice will be much more inclined to pay attention to their risk adjustment program. A critical factor for success is buy-in from the practice management and leadership. They need to understand the mechanics of their risk adjustment program and impact on revenue, so they will sponsor implementing the right processes and commit resources.
Select the right tools & integrate them into the clinicians' workflows: There are a variety of vendors providing tools to help doctors identify and close risk gaps. The idea is to bring risk model outputs to the provider at the point of care so that the provider can act on the data. We discussed a few features that make these tools successful. First, the data needs to be well embedded into the clinical workflow, i.e., the EMR, and it should not come as a stand-alone Excel file sent monthly via email. The provider should be able to act on the data in as few clicks as possible, and it's imperative that they can trust the data. Some discussion participants shared that instead of giving doctors black-box risk scores and prompts, doctors would be much more likely to act on data if they get specific supporting evidence, such as references to past claims or medications.
Education & Training: Even if there is motivation at the practice management level, they must communicate the relevance of risk coding to the attending physicians. Health plans and value-based care providers need to invest in good risk adjustment training. People shared that dedicated training & practice sessions on properly using risk adjustment support tools can greatly improve their adoption.
Clinical Value: In the end, the risk adjustment model should not just be used for reimbursement purposes but also to guide patient outcomes. A good risk model can identify patients who need dedicated outreach, are at risk of acute episodes or are skipping their care. While the risk gap and the care gap, i.e., the gap between needed care and actual care, are not always overlapping, they are often related, and risk gaps can help identify gaps of care.