A deep learning model that continuously analyses electronic health record data can predict inpatient hypoglycemia up to 24 hours before it occurs, potentially allowing hospitals to intervene before patients experience dangerous low blood glucose episodes.
Researchers at Cedars-Sinai Medical Center, Los Angeles, developed and prospectively evaluated the system using data from 143,124 adult hospital admissions across three hospitals collected between 2014 and 2025. The model combines routinely collected information, including laboratory test results, medication records, diet orders, meal consumption, and patient history, to generate updated risk predictions every four hours for patients receiving glucose-lowering treatment.
Unlike many previous approaches, the model analyses how clinical information changes over time rather than treating each measurement independently. This enabled it to outperform conventional machine learning methods, including logistic regression, dense neural networks, and gradient-boosted decision trees, while maintaining its performance during prospective testing using live hospital data.
The study, published in npj Digital Medicine, highlights the growing role of laboratory data within clinical decision support systems. Laboratory results alone were among the strongest individual contributors to prediction performance, second only to medication data, while combining laboratory information with other routinely collected clinical data produced the best overall results.
Rather than replacing existing glucose monitoring, the researchers propose that the model could operate in the background of electronic health record systems, identifying patients who may benefit from earlier review. By flagging high-risk patients before hypoglycemia develops, laboratory findings could be interpreted alongside medication history and nutritional status to support proactive adjustments to treatment.
The model also provides information about which recent clinical factors contributed most to each prediction. Recent insulin administration and previous episodes of hypoglycemia were among the most influential drivers of risk, offering clinicians greater transparency than is typical for artificial intelligence systems.
Importantly, performance remained broadly consistent across demographic groups and during daily prospective validation, suggesting the approach could be suitable for real-world deployment. However, the authors caution that the model was developed within a single health system and requires external validation before widespread adoption. They also note that future studies should determine whether integrating the system into routine hospital workflows ultimately reduces hypoglycemia, shortens hospital stays, and improves patient outcomes.
Newsletters
Receive the latest pathologist news, personalities, education, and career development – weekly to your inbox.
