Closing the Loop on Unexpected Imaging Findings: Augmenting the Process with Artificial Intelligence and Deep Learning (CLU)

The Opportunity

The use of imaging for patient care increased over the past 20 years and resulted in the ability to identify pathology and facilitate accurate and timely diagnosis. Imaging may uncover abnormalities unrelated to the focus of the imaging, but nonetheless may have large clinical significance. Unfortunately, because these findings may not be relevant to the patient’s current issue, results of the imaging study are often “lost” due to lack of provider follow-up. The problem this team set out to solve was how to implement an automated tool to analyze radiology reports and identify incidental findings that should have follow-up clinical evaluation.

Our Approach

The team created a workflow to automatically identify unexpected imaging findings that should have follow-up imaging and/or follow-up clinical evaluation by using natural language processing, AI, and Deep Learning (DL). These fields then populated a worklist that is monitored by our quality and safety office. If patients did not have follow-up imaging or a clinic visit in the prescribed time frame, the office contacted the patient’s physician, or in some cases, the patient directly, to make certain they were aware of the recommendation for follow-up.

The Solution

A novel generative-discriminative model was created that was able to obtain a high classification accuracy across the board, demonstrating that it is possible to automatically identify follow-up recommendations. At the end of each day, a list of missed incidental findings was sent to the quality and safety staff, who then agreed or disagreed with the AI solution, which triggered an email to the reporting physician. If the physician concurred with the model findings, they adhered to the current follow-up workflow, resulting in appropriate communication to care team members and follow-up care for the patient.