Imaging experts have developed an artificial intelligence tool that can help predict delays in radiology turnaround times during nights and weekends, key info for quality improvement efforts.

University of California, San Francisco, researchers created the machine learning model utilizing more than 15,000 CT scans. Testing the tool out, they produced solid early results predicting delays greater than 245 minutes (area under the curve of 0.85) and interpretation setbacks of 57 minutes or longer (AUC 0.71).

“As delays in radiology are an important measure of patient safety and hospital efficiency, having the ability to predict such potential delays has important benefits,” Jae Ho Sohn, MD, a cardiothoracic radiology fellow at UCSF, and colleagues wrote June 27 in Academic Radiology. “Furthermore, prediction of delays in radiology can improve the referrer and radiologist relationship and help clinicians to prepare alternative options in case a delay is expected.”

For their study, San Francisco scientists gathered retrospective CT data from two hospitals within the same organization, logged between 2018 and 2019. The original set included nearly 30,000 inpatient and emergency cases, whittled down to about half that for their analysis. They tracked order and scan time, first communication by radiologist, free-text indications and more.

Sohn et al. used 85% of this data to train their ensemble machine learning model and the remaining 15% for testing. AI was tasked with predicting delays between when the exam was ordered to the first communication, along with delays between scan completion and interpretation.

The team discovered that CT study description, time of day and year in training were much more predictive features than body part imaged, inpatient status and hospital campus. In addition, some protocols were associated with delayed turnaround time because of the complexity of cases, including CT of the neck with contrast, were associated with delayed turnaround times

Future studies could potentially add additional variables, such as hospital and ED patient census, number of providers, transportation and average technology operating time. Sohn and colleagues see their work as an important starting point for quality improvement projects.

“Given the complexity of real-world radiology workflow, no algorithm can make perfect predictions on which cases will be delayed. However, attaining a reasonable prediction of such cases can be relevant,” the authors advised.

Read more about their analysis in the Association of University Radiologists' official journal here.