Abstract: SA-PO967
Reducing Hospitalizations with Artificial Intelligence and Clinical Decision Support: Lessons Learned
Session Information
- Hemodialysis and Frequent Dialysis - V
November 09, 2019 | Location: Exhibit Hall, Walter E. Washington Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Dialysis
- 701 Dialysis: Hemodialysis and Frequent Dialysis
Authors
- Long, Andrew, Fresenius Medical Care, Waltham, Massachusetts, United States
- Lindsey, Adriana M., Fresenius Health Partners, Leander, Texas, United States
- Willetts, Joanna, Fresenius Medical Care, Waltham, Massachusetts, United States
- Han, Hao, Fresenius Medical Care, Waltham, Massachusetts, United States
- Chaudhuri, Sheetal, Fresenius Medical Care, Waltham, Massachusetts, United States
- Gallagher, Cara S., Fresenius Health Partners, Leander, Texas, United States
- Usvyat, Len A., Fresenius Medical Care, Waltham, Massachusetts, United States
- Ketchersid, Terry L., Fresenius Medical Care, Waltham, Massachusetts, United States
- Aronson, Andrew, Fresenius Health Partners, Leander, Texas, United States
- Maddux, Franklin W., Fresenius Medical Care, Waltham, Massachusetts, United States
Background
Artificial intelligence has the potential to improve healthcare. Previously, we created a model to predict which patients with kidney failure treated in outpatient dialysis clinics were at risk for all cause hospital admission in the next week. This model has an area under the receiver operating characteristic curve (AUROC) of 0.78 with a sensitivity of 69% and specificity 72%. Here we discuss lessons learned from integrating this model in a telephonic intervention.
Methods
Starting in December 2018, our analytics team partnered with a team of nurses who perform chart reviews and telephonic outreach to manage a large population of patients distributed across the United States. The goal of the outreach is to reduce the number of hospitalizations. The workflow consists of pulling patients from a worklist, reviewing a chart, documenting any needs identified, and then calling the patient if warranted. Through the last few months, we have utilized agile techniques to iteratively improve each step of this process using observations, surveys, and data analytics.
Results
We deployed the predictive model as a worklist ranked by risk score through an excel sheet format. Initial review demonstrated that 1) excel sheets are difficult to use to display individual patient data, and 2) significant time was spent digging through electronic charts. To help mitigate these issues, we built a dashboard that showed the prediction-based prioritization worklist as well as an integrated patient view. Testing of the dashboard with 3 nurses increased the number of chart reviews by over 50%. We believe this is likely due to aggregating information from multiple electronic health records, reducing the time spent searching for information. Further, nearly every nurse who uses this new system has reported an increase in job satisfaction. To date, the existing workflow results in 70% increase in chart reviews per day with 250% increase in calls per day. Investigation of hospitalization rate is still underway.
Conclusion
Building healthcare predictive models is only part of the story for artificial intelligence to improve healthcare. Additional work must be conducted to understand how to fully integrate predictive models into existing and newly designed clinical workflows.
Funding
- Commercial Support –