From the October 2017 issue of HealthCare Business News magazine
By Neil Smiley
Upon admission, most hospitals utilize some form of risk stratification, either manual assessments or automated risk scoring running alongside their EHR.
LACE is a popular risk algorithm used by hospitals, which uses a combination of length of stay, acuity, comorbidities and history of emergency department visits to rank a patient’s readmission risk on a scale of 1 to 19. Patients with high readmission risk scores are then flagged for more intensive therapies and interventions.
However, just using a static risk prediction model to stratify patient populations at the time of admission is not enough. Hospitals with high populations of at-risk patients must leverage data analytics at every step along the care continuum, including:
• Identifying high-risk patients.
• Matching patients to interventions.
• Engaging patients through evidence-based programs.
• Evaluating clinical and economic outcomes.
Identifying high-risk patients
Hospitals have the opportunity to significantly improve on basic risk prediction models like LACE by integrating a more robust set of real-time data markers that are updated throughout a patient’s care journey. For example, LACE does not utilize data on social factors, medications, labs or unstructured data such as admitting complaint and clinical notes.
In addition to understanding what data is available, it is also essential to understand when data is available. For example, it’s not uncommon for a final diagnosis to arrive five to seven days after discharge. Prediction models usually have to rely on “proxy data” such as admitting complaint, medication lists, lab values and natural language processing of free form text to allow for early identification of patients with high-risk conditions.
Unless hospitals are part of a fully integrated health system, they will also need to collaborate with upstream and downstream care partners to fill in data gaps across the care continuum. Each hospital will vary in coding conventions and care practice patterns. Prediction models must be adaptive to these differences in order to fully leverage the most up-to-date information at each implementation site.
Once a prediction algorithm is deployed, hospitals should have a regular process to evaluate model performance. Risk prediction models will need to be periodically adjusted to reflect changes in underlying coding systems and care practices, and to capitalize on new data markers as they become available. Risk models should also be leveraging machine learning to steadily improve prediction accuracy as more outcome data becomes available.
Matching patients to interventions
Once high-risk patients have been flagged, data analytics should also be leveraged to match patients to the most appropriate interventions using individual patient risk characteristics and the inherent constraints associated with each intervention program. Patients with the highest risk scores may not always be the best candidates. For example, patients with 10 or more medications are typically less responsive to phone-based medication adherence programs and may need home visits instead.
Creating a data feedback loop for each program that analyzes which patients respond and which don’t enables a data-driven approach to refine program eligibility criteria and improve intervention methods.
