Custom Care Management Algorithms that Actually Reveal Risk

Article Summary

Care management is a tool for population health that focuses scarce healthcare resources on the sickest patients. Care management leaders need to know who those sickest patients are (or may be). The static risk models typically used for stratifying patients into risk categories only paint a partial picture of health and are ineffective for modern care management programs. Custom algorithms are now capable of predicting risk based on multiple risk models and multiple data sources. They help care management teams confidently stratify patient populations to paint a complete picture of care needs and efficiently deliver care to those who need it most.

This article explains how custom algorithms work on static risk models to normalize risk scores and improve patient stratification, care management, and, ultimately, population health management.

Graphic showing icons of medical professionals and patients linked together with a laptop and smartphone in the center

Care management is a population health strategy tool that helps patients achieve their healthcare goals and overcome socioeconomic barriers to care. These achievements are guiding principles of value-based care. Though organizations deploy care management programs differently, it is the bedrock of every successful population health approach.

A primary care management component is patient stratification: the process of separating patient populations into high-risk, low-risk, and rising-risk groups, which is how health systems identify patients most likely to benefit from a care management program. But patient stratification, and the entire population health strategy, is only as effective as the data and risk models it employs. Accurate and effective risk scoring depends on multiple models, multiple data sources, and custom care management algorithms capable of blending them all into a comprehensive patient stratification visualization.

Care Management Risk Modeling Under Value-Based Care

Care management practitioners and goals have changed as healthcare has moved from fee-for-service (FFS) to value-based care. Under FFS, a payer uses care management techniques to identify members of the population, target them for enrollment, and plan appropriate care services. The payer uses claims and clinical data, which are run through a black box risk model, to assign static risk scores.

Under value-based care, healthcare organizations (HCOs) have more complex risk modeling needs because their goals center on the health of multiple patient populations. These organizations’ goals involve multiple payers and multiple risk models, which tend to complicate care management. Because most organizations have limited care management resources, identifying those patients who will respond best to care management tactics is important. HCOs need to incorporate a dynamic risk scoring methodology over the top of these static risk models to meet the patient stratification demands of value-based care.

The Problem with Static Risk Model Scores

Different risk models produce unique risk scores. For example, Conifer and Milliman Advanced Risk Adjuster (MARA) are two popular models used by payers to produce static scores. A patient with a Conifer risk score greater than 30 is considered high-risk, as is a patient with a MARA risk score greater than 3.5. But comparing these scores is like comparing apples to oranges. Though they have similar meaning, they are on completely different scales. Also, these models were originally developed without access to EMR data.

Typical clinical risk models also produce unique scores, such as predictive risk, admissions risk, Charlson/Deyo, and HHS-HCC. These models were built with limited access to clinical datasets. In managing population health, care managers need to effectively compare patient risk levels, then target and coordinate care appropriately. But static, incongruent scores make it nearly impossible to compare risk levels to stratify patients and create precise lists of patients for assigning to care management programs. Normalizing risk scores is necessary for precise stratification.

How to Normalize Static Risk Model Scores with Custom Care Management Algorithms

HCOs need an analytics solution that can ingest the vast amounts of data that come from their patient populations. The solution must also work with disparate data sources, including claims, EMRs, and clinical applications. A solution that can nimbly handle large datasets from a broad representation of data sources establishes the process for normalizing risk scores.

To normalize scores (i.e., standardize them for comparison in a risk stratification process), care management teams must be able to build custom care management algorithms from all datasets, so they can accurately stratify patients into risk categories.

Comprehensive stratification algorithms consider chronic conditions, risk, utilization, medication, and social determinant variables, each of which must be weight-adjustable. For example, the utilization variable could be weighted by any combination of ED visits, hospital admissions, skilled nursing facility stays, specialist visits, or ICU stays. The risk variable could be weighted by predicted risk, rising risk, readmission risk, HHS-HCC risk, and Charlson-Deyo risk. A custom algorithm must play on the strengths of multiple risk models and data sources, while removing the fragmentation imposed by individual, mismatched scores and disparate data sources.

Effective Patient Stratification Accounts for Patient Complexity

Typical risk models don’t take into account the complexity of patients. Instead, they rely on static datasets. A clinical- or claims-based patient stratification algorithm only looks at a portion of the patient profile, which limits the care management team to only a partial view of the risk variables impacting each patient.

Custom care management algorithms reveal patient populations with complex care needs that otherwise would be missed using traditional risk scoring models. Clinical, claims, and socioeconomic data are all required for effective patient stratification. Though robust geographic-based socioeconomic datasets still need to be collected, the value of being able to ingest this data cannot be overstated. Clinical and claims data alone cannot tell the care management team if a patient can afford to pay for medication or understand discharge instructions. Patient stratification care management software should include modules for consuming all these data sources, especially socioeconomic data, to generate a list of patients that can then be reviewed and assigned to a care management program.

In a properly designed patient stratification tool, custom algorithms should also use machine learning to predict which patients will respond to care management tactics; e.g., who will be readmitted and who will be most impacted by a care program. Care management teams should be able to save algorithms and use them for generating updated lists of care management patient candidates daily. Algorithms should be developed in an open-source environment and shared between HCOs to support diverse, widespread population health initiatives.

Population Health Demands Next-Level Patient Stratification

Population health initiatives require knowing as much about patient populations as possible. More robust data enables this understanding, shows the care gaps, stratifies the sickest patients, and reveals how they are utilizing resources.

Using multiple risk models, multiple data sources, and dynamic patient stratification algorithms, care management teams can confidently target populations for care management resources: a critical process for meeting the financial and clinical challenges of value-based care.

PowerPoint Slides

Would you like to use or share these concepts?  Download this presentation highlighting the key main points.

Click Here to Download the Slides

Diversity in the Workplace: A Principle-Driven Approach to Broadening the Talent Pool

This website stores data such as cookies to enable essential site functionality, as well as marketing, personalization, and analytics. By remaining on this website you indicate your consent. For more information please visit our Privacy Policy.