Sutter Health

Predicting hospital readmission risk for better interventions
Partner: Sutter Health Hospital Network
Area of Innovation: Predictive modeling to decrease hospital readmissions

The Opportunity

The United States spends more on health care than any other country on a per capita basis. Unfortunately, there are still huge gaps in quality of care. One specific and measurable gap is the patient readmission rate, the proportion of patients discharged from a hospital who are admitted again within a short amount of time. This figure stands at 15% for Medicare patients nationwide. The readmission of patients in the Medicare system alone costs taxpayers more than $26 billion annually, of which $17 billion is considered avoidable. The problem is most acute for socioeconomically disadvantaged patients and those requiring complex follow-up care.

Reducing readmissions requires evidence-based practices and data-driven solutions to understand the root causes of individual patients’ readmissions and design personalized interventions to proactively address them. Data science has the power to meet this challenge. Many recent academic studies have attempted to model patient readmission risk, but few have applied modern machine learning techniques. The resulting models have also rarely been fully integrated into hospitals’ technology stack or their care management workflows.

Solution

We partnered with Sutter Health, one of California’s largest hospital systems, to accurately identify patients at high risk of readmission in order to deliver appropriate follow-up care. Our team built an automated data-cleaning and ingestion pipeline for Sutter Health that automatically geocodes millions of patient addresses and combines the output with openly available demographic data. The pipeline eliminated the frequent complications caused by manual data ingestion and dramatically reduced overall processing time. We then built an artificial neural network trained on 1,500 features extracted from millions of patient records to identify patients at high risk of readmission at the point of care. The model has a 20% higher positive predictive value (PPV) than the industry standard and updates in real-time for a variety of medical conditions.

Results

The results of this project were published in PLOS ONE: Predicting all-cause risk of 30-day hospital readmission using artificial neural networks, 2017. The public GitHub repository can be found here.

This project was funded by a grant from Robert Wood Johnson Foundation and we are proud to publish our algorithms and methods under open source licenses for others to contribute to and use. This project is only the beginning. Further work in this direction has the potential to provide hospitals of all sizes with an advanced, adaptive, and integrated care management tool, which can improve the quality of care for millions of people across the country.

Interested in partnering with us?
let's connect