Examining patient health records with artificial intelligence could give hospitals advance notice of next-day discharges, providing hospital staff with a more efficient method of optimizing space and services.
Vanderbilt University Medical Center researchers led by bioinformatician You Chen, Ph.D., developed and published a strategy to manage hospital capacity with help from the EHR. Their report in the Journal of the American Medical Informatics Association identified key factors that hospital administrators might use to forecast next-day discharges.
“Without data, you’re just another person with an opinion,” Chen said. “We used a large collection of clinicians’ opinions, reflected in audit log data, to predict next-day discharge.”
Harnessing the Dataset
A major strength of the study is that the researchers used the EHR audit log instead of subjective clinician judgement or manual predictions. This is in contrast to other initiatives, such as the AM bed huddle, endorsed by the Institute for Healthcare Improvement. The researchers’ automated approach was designed to maximize objectivity and scalability.
Their approach also incorporated general patient data, such as past diagnoses, age, heart rate, body mass index, admission day of the week and more into the predictions.
The full dataset included 26,283 inpatient stays and 133,398 patient days that involved 819 types of interactions between users and EHRs. Chen and colleagues incorporated EHR interactions until 2 p.m. to predict the following day’s discharge schedule. A machine learning model then analyzed interactions among nearly 3,000 unique features – far outweighing manual approaches.
High Degree of Accuracy
After training, the machine learning model predicted who would and would not be discharged the next day with 87.7 percent accuracy.
In total, it predicted 77.8 percent of actual discharges.
“We also observed 29.4 percent of the false positives corresponded to a discharge within 30 hours from the prediction time point,” the authors noted.
Accuracy further improved when the researchers focused on specific hospital units.
The model also identified 20 predictive factors that most influence next-day patient discharge. Half came directly from the EHR audit log. For example, a patient with a connective-tissue disease diagnosis was not likely to be discharged. Similarly, a patient involved in a frequent barcode scan, such as from a medical device, or with access to their EHR from a nurse’s monitoring station was not likely to be discharged the next day.
“The innovation is that we are the first to find that audit-log data – documented activities of clinicians in EHRs – can be leveraged to indicate a patient’s discharge status,” Chen said.
Improvements Over Time
As with most machine learning models, performance improves over time. Chen says EHR analyses might therefore be used to predict patient discharge on an ever-improving, rolling basis.
“Documented activities of clinicians in EHRs – can be leveraged to indicate a patient’s discharge status.”
“The audit log data is widely used to measure clinicians’ time spent in EHR systems, user-EHR interaction flows, and asynchronized user-user interactions. Our study assigns the audit log data a new function – predicting the next 24 hours of discharge with high accuracy.”
Support for Patients
Efficient capacity management of hospitals sets the stage for effective patient care, the researchers emphasize.
“Poor operation of hospital capacity management often prolongs a patient’s waiting time until admission and discharge,” they wrote. “Both of these delays lead to inefficient planning and utilization of care resources, increased health care costs, and consequently, a higher risk of in-hospital complications and mortality.”
The new model provides hospital administrators an analysis tool to predict discharges one day ahead – a boon for resource allocation, particularly during a persistent pandemic.
Said Chen, “Using artificial intelligence, we’ve hit upon a predictive model that can be built in EHRs. With comprehensive interactions between users and EHRs to work with, there’s every reason to imagine that a thorough scan of this sort can yield important insights and new solutions to improve the management of patient flows and the allocation of clinical resources.”