Now that most health care organizations have moved from pen and paper to electronic health records (EHRs), health systems face a new opportunity: incorporating machine learning to improve patient care. However, figuring out the best ways to leverage billions of data points can be difficult. For the last several years, Children’s Hospital of Orange County (CHOC Children’s) applied machine learning to data with on-premise resources and now has taken the next step of leveraging Cerner’s cloud-based environment for the work.
By using machine learning data to more accurately predict which patients are more likely to be readmitted to the hospital, CHOC gains an opportunity for providers to proactively take measures to reduce the likelihood of readmission.
“When Cerner approached us to be the development partner with HealtheDataLab™, we quickly said ‘yes,’ because we knew what its power would likely do,” said William Feaster, MD, chief health information officer. “Cerner did all the heavy lifting. While it took several months of a lot of hard work, we’re now reaping the benefits.”
Once staff began using cloud-based HealtheDataLab in mid-2018, staff could dig deeper into the data with more complex datasets to fine-tune the algorithm that predicts the accuracy of the number of readmissions. The more detailed data provides more accurate results.
“Now that we have the database in HealtheDataLab, we are no longer limited by resources, and we can analyze billions of data points, which could not previously be done,” said Louis Ehwerhemuepha, PhD, data scientist.
Feaster agrees the solution helped CHOC Children’s staff further improve their predictive model.
“HealtheDataLab is an environment not limited by storage resources, not limited by computing power, not limited by access to manipulating big data and running artificial intelligence-type transactions,” he said.
The AUC (area under the curve) is a measure used to assess the performance of a predictive model. Good models have an AUC closer to one and incremental improvements, even by one-hundredth is difficult to attain. Before implementation of HealtheDataLab, CHOC Children’s AUC for 30-day readmission rate stood at 0.79. After implementation and the pilot program, the AUC rose to 0.82 as the CHOC data scientist made numerous tweaks to the dataset.1
“It took several iterations of adding more and more data before we got it to a point now where we’re comfortable with the accuracy,” said Feaster. “The corresponding models built in HealtheDataLab currently ranks among the highest in literature and utilizes de-identified data from 48 hospitals and 1.4 million encounters. The wide breadth of data helped fine-tune the model.”
“Our data science team gained tremendous efficiency,” said Ehwerhemuepha. “We no longer require assistance from multiple groups for all our activities, and we are not constrained by hardware limits.”
CHOC Children’s explored 125 variables including metrics like gender, age, race, previous number of visits and chronic conditions. The algorithm also looks at social determinants of health variables in patients’ ZIP codes including medium income, vacant homes and the number of single parent households.
The clinical readmissions team adopted the readmission model which is now implemented into the EHR, allowing care-takers and providers to see the readmission risk for all hospitalized patients. By seeing which patients are at a higher risk to return, it helps them decide if additional patient care is necessary to prevent readmission.
CHOC Children’s leaders say predicting readmissions is only a beginning. The health system is looking to utilize HealtheDataLab to predict which patients will not show up for appointments. Other planned future projects include developing models to better predict emergency department sepsis cases and a rising risk model (patients who begin to return to the hospital more frequently) for patients with chronic diseases.
“We’re really excited to be working in this environment. I think it’s going to advance our ability to care for patients,” said Feaster.
1 Utilizing admissions data between 2004 – 2018.