Top Five Elements of an Effective Readmission Risk Score
Readmission risk scores are pervasive in healthcare. With increasing options in risk models (e.g., LACE and LACE+ indexes, and scores from EMR and analytics vendors), health systems are overwhelmed with choices. For many organizations, partnering with an analytics vendor to develop a readmission risk score that meets their unique needs is the highest value, most effective option.
Risk stratification uses risk scores to separate patients into high-risk, low-risk, and rising-risk groups; it’s essential in value-based healthcare. Health systems use risk-level insight to adjust their interventions and improve outcomes by identifying and proactively managing high-risk, high-cost patients. For example, when an organization uses a readmission risk score for patients with chronic obstructive pulmonary disease (COPD), it can plan interventions (e.g., follow-up visits and medication management) to help these patients avoid returning to the hospital.
In the past few years, reducing unnecessary readmissions has become increasingly important for healthcare organizations. Under the Affordable Care Act’s 2012 Hospital Readmission Reduction Program (HRRP), CMS may penalize healthcare organizations for unacceptably high readmission rates among Medicare and Medicaid patients in general and with specific conditions, such as COPD; this makes reliable risk prediction a health system imperative.
Developing an Effective Readmission Risk Score: A Five-Part Checklist
The following checklist describes five elements that enable the effective development of a readmission risk model.
1. Identifies at-risk patients early.
By identifying patients at risk of readmission as soon as possible after the patient is admitted to the hospital, and before discharge, an effective risk model enables clinicians to intervene quickly. A hospitalized patient is a captive audience, allowing clinicians and frontline caregivers to provide patient education and set up transitions of care to the outpatient setting. These interventions become more difficult and less effective once the patient has gone home. By tailoring the identification method and timing to the needs of the use case, developers can identify the majority of at-risk patients within hours of admission and days prior to discharge.
2. Separates patients relevant to the disease-specific identification method and intervention strategy from all other in-hospital patients.
Using the COPD example, if an organization wants to intervene to reduce COPD readmissions, it needs to stratify COPD inpatients separately from other inpatients. Some established risk identification methods don’t make this distinction; instead, they calculate risk for all patients together—meaning the risk score and intervention strategy are not tailored to the unique needs of patients with COPD.
In a disease-specific identification method, developers determine how to identify the desired population while patients are still in the hospital. Identifying patients while they are still in the hospital is more challenging than most people know because, while diagnosis codes are typically used to identify patients retrospectively, those codes are not available until after discharge. Developers must use other data to infer diagnosis, such as historical diagnoses, progress notes, chief complaint, or specific medications that were used to treat the patient.
Many health systems prefer a method with high sensitivity more than one with high specificity—meaning it includes most of the organization’s patients with COPD and perhaps a few that will end up other diagnoses (versus a specific method that finds only patients with COPD, but doesn’t catch all COPD admissions). Each organization will determine an acceptable level of error for its identification method.
3. Uses organization-specific data to train a disease-specific model.
Health systems inherently treat different patients with distinct interventions; in addition, patient populations differ by region and by facility. An effective risk model allows the health system to customize the tool by including features unique to its organization.
Organizations often have tens of thousands of historical visit records related to a given disease state; this organizational and disease-specific data is a tremendous asset and can provide hundreds of potentially predictive features to test and include in readmission risk model.
4. Model performance exceeds existing models.
Risk model performance comparison is a critical factor in adoption, as end users will likely compare a novel tool to one they’re currently using. It is important for developers to determine whether the new approach is as good as or better than existing tools and support their assertion with data. A custom model can exceed the accuracy of LACE and other models by 10 to 20 percent, which can translate to hundreds of additional correct predictions and more efficient care delivery as a result.
5. Developers collaborate with and consult domain experts.
Risk model development starts and ends with domain experts. Developers engage domain experts (who are also eventual end users) in every step, from defining a population and deciding what level of error to accept, to determining interventions and documenting sources. This way, domain experts add the value of their clinical understanding and are more likely to adopt the model, as they understand its relevance as a result of being been a part of its development.
In the case of COPD, domain experts can include hospitalists, pulmonologists, and operational leaders. A clinical analyst, who’s familiar with the content and data, understands clinical concepts and has clinician relationships that can also help bridge the gap between data and the clinical setting.
Domain expert collaboration is a unique benefit of working with an analytics vendor that offers expertise, transparency, and follows the five-part checklist to develop a score (versus implementing a simple risk score). As end users, the domain experts are more likely to adopt a model they’ve helped develop or implement, and strong user adoption means fewer readmissions.
From Identification to Intervention
Identifying at-risk patients is a readmission risk model’s first task. The next step focuses on intervention strategies to prevent readmission, using information gleaned from the prediction model.
Again, end user buy-in is an important part of getting clinicians to rely on risk models for intervention strategies. Developers who continue to engage domain experts after production and gather feedback will likely produce the most useful, adoptable risk prediction tools.
Health Systems Need the Right Risk Model
Under value-based healthcare and the HRRP, reducing readmissions will continue to be an important goal for health systems. As these systems work to reduce readmissions using risk models, they need to keep the five key elements of an effective risk model in mind, including partnering with an analytics vendor to identify at-risk patients early and according to the target population, using organization-specific data, and prioritizing end user buy-in by having developers demonstrate the value of the new tools compared to existing models.
Would you like to learn more about this topic? Here are some articles we suggest:
- Prescriptive Analytics Beats Simple Prediction for Improving Healthcare
- Patient Flight Path Analytics: From Airline Operations to Healthcare Outcomes
- Hospital Readmissions Reduction Program: Keys to Success
- Reduce Readmissions with Predictive Analytics and Process Redesign
- Driving Down Costly COPD-related Readmissions with NOREADMITS Bundle