Radiologists, particularly radiologists on-call, have to manage numerous non-interpretive tasks in the reading room that interrupt image interpretation. Originating in the airline industry, it has now been well established that interruptions often lead to error. This has resulted in widespread implementation of safety measures across industries. In radiology, a large number of the interruptions result from phone calls, with up to 15% of a 12 hour call shift being spent on the phone. It has been shown that on-call radiologist have a 59% percent chance of being interrupted by a phone call while interpreting a study that takes 10 minutes to complete.
Meanwhile, unnecessary imaging has been a major target for innovation by Centers for Medicare and Medicaid Services under its “Caring Wisely” campaign. Annually, it is estimated that $12 billion dollars are wasted on imaging that is not clinically indicated. This has most recently been manifested in the Medicare Access and CHIP Reauthorization Act of 2015 (MACRA) Merit-based Incentive Payment System, in which Medicare is requiring all advanced imaging studies to have been screened through approved clinical decision support.
The goal of this project will be to develop tool that both reduces phone-call burden on radiologists and provides a more natural clinical decision support for ordering physicians. To this end, we are in the process of creating Radlexa, a voice-user interface that automates routine questions fielded by radiologists that can be generally answered without clinical decision making.
The first step in this project will be to characterize the current state of phone calls to radiology reading rooms. Following IRB approval and with caller notification, we will record phone calls to the reading rooms over a 3 month period and then use voice-to-text technologies to generate transcripts of these calls. After some manual annotation, we will use natural language processing to train a model that can recognize which calls fall into which categories. Once we have generated this model, tasks that can be automatically answered will have automated responses generated. For example, inquiries about where a patient is in the imaging queue could generate automated responses. In this way, we will enrich calls to the radiologist primarily for urgent clinical questions.
As a financial incentive to drive adoption of Radlexa, clinical decision support algorithms will be built into the voice assistant, specifically around MACRA requirements. The goal of this is to both provide an alternate means of satisfying MACRA and to protect radiologists from becoming primary points of support for MACRA and electronic health record (EHR) integration. The interface could correctly route questions about how to place orders, reducing the need of radiologists to act as de-facto IT help desk for clinical decision support integrated into the EHR.
Radlexa has the potential to significantly reduce the non-clinical calls that are currently being fielded by radiologists while also being a clinical decision support tool for ordering providers. Once implemented, this will improve quality and patient safety and reduce unnecessary medical spending.
ROBOTIC TECHNOLOGY & VIRTUAL ASSISTANTS
Author: Lindsay Palmer Busby
Coauthor(s): Michael D. Wang, MD and David E. Avrin, MD, PhD
Status: Project Concept
Funding Acknowledgment: Dr. Busby was supported by the National Institutes of Health (NIBIB) T32 Training Grant, T32EB001631.