Transforming Health Care With Machine Learning
By Sridhar R. Tayur, Ford Distinguished Research Chair; University Professor of Operations Management
The Affordable Care Act has radically altered how health care systems operate. In one of the biggest shifts, hospitals are no longer reimbursed for patients who are readmitted within 30 days for the same condition. The goal is to give hospitals an incentive to reduce costly admissions and improve the overall quality of patient care.
The stakes are high. If the number of people who come back to a hospital within 30 days over the course of a year is higher than the national average for this disease, a hospital may face fines in the order of millions of dollars. It also is rewarded if it does better than the average.
Machine Learning Algorithms With a Higher Predictive Power
We partnered with University of Pittsburgh Medical Center physicians and researchers to apply machine learning techniques to predict hospital readmissions of high-risk patients with sickle cell disease.
Patients with sickle cell disease are not routinely evaluated for their readmission risk. But their complications cause frequent hospitalizations and readmissions. This increases costs and patient deaths.
Data is the cornerstone of decisions that drive health care today. Hospitals use two benchmark models to evaluate the readmission risk among patients. These were developed using data based on patient populations that weren’t a match for the patients at UPMC or the hospital’s ecosystem. So we developed machine learning algorithms with a higher predictive power using a broader set of factors applicable to sickle cell disease.
We studied anonymous health records of the hospital’s sickle-cell patients over five years, noting which patients returned within 30 days and which ones didn’t. Using both machine learning techniques and domain knowledge, we input variables that might explain the difference, such as demographics, vital signs, comorbidities, medications, and lab data.
Customizing Analytics to the Hospital’s Unique Needs
The machine learning methods we used are well known and widely available. But by inputting variables tailored to the high-risk sickle cell disease patients at UPMC, we customized the analytics to the hospital’s unique needs. This fared better than standard risk prediction tools.
This issue is top of mind for every health care system in America. Our approach shows how individual health care systems can use predictive analytics on their own data to craft decisions that are best for them.