Machine-learning model predicts interpretation delays

Thursday, December 5 | 11:00 a.m.-11:10 a.m. | SSQ11-04 | Room N229
A machine-learning algorithm can predict delays in performing and interpreting on-call radiology exams, a team of California researchers will report in this presentation.

One of the most common reasons for clinicians to contact radiologists during off hours is to obtain a wet read for an imaging study, according to senior study author Dr. Jae Ho Sohn of the University of California, San Francisco.

"Delays in radiology interpretation outside of business hours can lead to poor provider-radiologist relationship, worsened patient care, and increased costs for the hospital," he told

The researchers had two goals for this project: First, they wanted to be able to predict whether there would be a delay in the scanning and/or interpretation of a particular study based on factors such as body part, inpatient status, time of day, modality, clinical indications, the individual radiologist, and others. Next, they wanted to determine the most important factor that influences whether a radiology study will be delayed, Sohn said.

Using 12,525 cross-sectional imaging studies that were performed after hours at their institution, the researchers trained a random-forest algorithm to predict delays that were above the median. The model yielded an area under the curve of 0.76 on a test set of cases.

By predicting which radiological studies will be delayed during off-hours, clinicians can act accordingly and improve care for patients, Sohn said. The model is specific to their hospital, but the approach to training the algorithm can be widely applicable to other hospitals, he noted.

"Further work is being done to precisely identify and confirm the factors that best predict delays, and these results can be used in quality improvement studies to optimize radiology turnaround time during off-hours," Sohn said.

Stop by this presentation by lead author Vaibhavi Shah for more details.

Page 1 of 603
Next Page