Using statistical methods to improve care, introduce efficiencies, and reduce cost
Predictive modeling can help health systems decide how best to deploy limited resources and reach out to populations who might be at risk of a harmful event or negative health outcome. Predictive modeling uses data analysis, machine learning, and statistics to identify patterns in data and recognize the chance of specific events occurring. For example, a predictive model can help identify individuals with high medical need or disease burden who are at high risk of being hospitalized in the next 12 months.
Since the COVID-19 pandemic hit, predictive modeling has been helping guide the critical decisions that health systems have needed to make in the face of uncertainty. But it’s critical to do this work rigorously and to involve an interdisciplinary team that can guide health systems in avoiding pitfalls that lead to biased and misleading models.
The ACT Center’s advanced analytics team brings together data scientists and care delivery leaders to improve clinical and operational decision-making at Kaiser Permanente Washington by designing, testing, and implementing high-quality predictive models. We also advise on statistical design and methods and perform complex outcome analyses. Our team consists of scientists with deep knowledge and expertise in how to effectively use and interpret complex electronic health record and claims data.
- Use state-of-the-art machine learning and statistical methods to develop and validate risk prediction models — specifically focusing on methods to understand bias in the model and implementing strategies to mitigate bias.
- Advise on appropriate interpretation of risk model results and clinical implementation strategies.
- Leverage sophisticated data analytics to allow for accurate and meaningful conclusions to support decision-making.
Here are some examples of our work:
- When COVID-19 hit, we leveraged our advanced analytics capabilities to support Kaiser Permanente Washington's rapid response to the pandemic — using existing analytics infrastructure to rapidly deploy outreach efforts to medically vulnerable Medicare members to make sure their care needs were being addressed.
- We then quickly expanded this to include outreach efforts to non-Medicare members with complex medical and social needs and other chronic conditions.
- During the 2018/19 flu season, we validated a predictive risk model to identify members who were at high risk for influenza-related complications, which supported targeted outreach to encourage vaccination for those members.
- This brief, low-cost intervention was associated with increased vaccine uptake among high-risk patients. Based on these findings, Kaiser Permanente Washington adopted this intervention systemwide for the following flu season. (See story by KING 5 News, Dec. 2, 2019.)
Other analytics work
- Missed visits prediction model: Working with Kaiser Permanente Washington Patient Access, we developed a model for predicting visits likely to be no-show or same-day cancellation. Having pinpointed these appointments, we conducted a randomized quality improvement project to estimate the effectiveness of additional text message reminders for reducing no shows and same-day cancellations in primary care and mental health. We found that the intervention reduced no-shows and same-day cancellations, so we are currently building this model into our electronic health record to support systemwide adoption.
- Other prediction models: Working with leaders and care teams from across Kaiser Permanente Washington, we've developed or tested models for a wide range of health system challenges — including preventable hospitalizations, potential harm from opioid use, and risk for suicide attempt.
What we've learned
Applying a variety of advanced analytic methods helps Kaiser Permanente Washington health system leaders address a range of current challenges:
- Validation of risk models in our own population ensures that we avoid implementing models that do not perform well, including those that perpetuate bias and health disparities.
- Predictive models have been deployed in ways that more effectively target intensive interventions.
- Rigorously conducted randomized quality improvement projects have assessed the effectiveness of interventions in a way that reduces confounding effects.
Each project we conduct brings a new opportunity to improve our approach to evaluating data and presenting results to our partners in a meaningful way.
Leveraging the ACT Center’s capabilities in advanced analytics allows Kaiser Permanente Washington to make informed decisions about how to prioritize resources and improve care in equitable and sustainable ways. And as urgent questions arise — such as those related to the onset of COVID-19 — our data scientists act quickly to provide rigorous analyses that inform crucial decisions and help ensure patients are getting the care they need.
Ulloa-Pérez E, Blasi PR, Westbrook EO, Lozano P, Coleman KF, Coley RY. Pragmatic Randomized Study of Targeted Text Message Reminders to Reduce Missed Clinic Visits. Perm J 2022;26:21.078. E-pub 5 April 2022. doi.org/10.7812/TPP/21.078. Full text