Implementing Risk Stratification in Primary Care: Challenges and Strategies

Abstract

Introduction: Primary care risk stratification (RS) has been shown to help practices better understand their patient populations' needs and may improve health outcomes and reduce expenditures by targeting and tailoring care to high-need patients. This study aims to understand key considerations practices faced and practice experiences as they began to implement RS models. Methods: We conducted semistructured interviews about experiences in RS with 34 stakeholders from 15 primary care practices in Oregon and Colorado and qualitatively analyzed the data. Results: Three decisions were important in shaping practices' experiences with RS: choosing established versus self-created algorithms or heuristics, clinical intuition, or a combination; selecting mechanisms for assigning risk scores; determining how to integrate RS approaches into care delivery. Practices using clinical intuition found stratification time-consuming and difficult to incorporate into existing workflows, but trusted risk scores more than those using algorithms. Trust in risk scores was influenced by data extraction capabilities; practices often lacked sufficient data to calculate their perceived optimal risk score. Displaying the scores to the care team was a major issue. Finally, obtaining buy-in from care team members was challenging, requiring repeated cycles of improvement and workflow integration. Discussion: Practices used iterative approaches to RS implementation. As a result, procedural and algorithmic changes were introduced and were influenced by practices' health IT, staffing, and resource capacities. Practices were most successful when able to make iterative changes to their approaches, incorporated both automation and human process in RS, educated staff on the importance of RS, and had readily accessible risk scores.

Health information technology to improve care for people with multiple chronic conditions

“…Currently, health IT applications are limited by problems such as poor usability, poor workflow integration, fragmentation, lack of interoperability, and the uneven distribution of technology (the so‐called “digital divide”) 24,25 . However, advances in our ability to exchange data, to use advanced algorithms, to parse complex situations or provide tailored support, and broad changes in policy and uptake using health IT present opportunities to leverage new capabilities to improve care and outcomes for PLWMCC 26–28 …”

Socially situated risk: challenges and strategies for implementing algorithmic risk scoring for care management

“…Our findings on the need to combine physician input and claims-based risk stratification to improve patient eligibility determination build on previous analysis demonstrating providers’ increased confidence in risk stratification that incorporates human review. 1 , 16 , 19 Because providers often incorporate psychosocial risks in their assessments of complexity, they provide key insight into patient needs that is not available in claims data or the patient record. 14 Furthermore, missing data on patient needs introduces bias, despite growing efforts to incorporate social determinants of health in the EHR and risk assessments.…”

“…Thus, until these issues can be addressed, algorithmic risk scores may need to be supplemented by provider input. 15 , 16 …”

Of Screening, Stratification, and Scores

“…In developing policies and assessment methodologies for the provision of risk-stratified healthcare, it is necessary to define the concerned population and to categorise population members according to their individuated health risk in accordance with a defined risk-stratification methodology. Methodologies for assessing individual health risk include algorithmic methodologies that entail the calculation of a risk score based on input data and human-initiated methodologies that are reliant on clinical judgment to assess health risk [ 41 ]. In practice, the most common are hybrid approaches, which involve the application of human interpretation to algorithm-derived scores [ 42 ].…”