A convolutional neural network has proven better than traditional equations, indexes and scoring systems at predicting which hospitalized patients will face readmission not long after being sent home.
The advance, tested in three separate hospitals, stands to help improve discharge planning, in the process helping patients avoid repeated exposures to sedation-related falls, hospital-borne germs and other health risks associated with inpatient stays.
The study behind the findings was conducted at the University of Maryland and is running in JAMA Network Open.
Daniel Morgan, MD, MS, and colleagues worked with medical records from 14,062 adult patients who were discharged from a tertiary care center, a suburban community hospital and an urban critical-access hospital over the last four months of 2016.
A data scientist on the team created a neural network to comb the records for thousands of variables before settling on 382 that were most closely associated with subsequent readmissions.
These included not only demographics, test results, need for breathing assistance and body mass index but also religious affiliation, marital status, employment and substance abuse.
Comparing their automated machine-learning system against several traditional means of predicting readmissions, the researchers found their approach 25.5% to 54.9% more efficient than comparison scores.
In their discussion, the authors point out that predicting readmissions is not the same as preventing them.
“Although we used the [Medicare & Medicaid] definition of potentially preventable readmission, the literature would suggest that most of these readmissions are not preventable,” they write. “Interventions to prevent readmissions are often labor intensive and costly, including discharge clinics, transitional care and telemonitoring.”
Nevertheless, more accurate readmission predictions facilitate better pre-discharge interventions and care decisions, they add.
E. Albert Reece, MD, PhD, MBA, dean of the University of Maryland School of Medicine, commented on the study to the school’s news division.
“The widespread use of electronic health records has enhanced information flow from all clinicians involved in a patient’s treatment,” Reece said. “This study underscores how patient data may also help solve the readmission puzzle and, ultimately, improve the quality of patient care.”