Amazon Alexa skill RAD-Assistant successfully retrieves pulmonary nodule and ovarian cyst clinical follow-up guidelines.
Amazon Alexa could have a job in a radiology office soon – and not as the office DJ responsible for background music.
RAD-Assistant, an Amazon Alexa skill, developed by researchers at the University of Pittsburgh Medical Center, can take direction from radiologists and accurately return follow-up guidelines for pulmonary nodules and ovarian cysts, making it a strong candidate to be a point-of-care clinical decision support tool.
The research team made their results available in the Society of Imaging Informatics in Medicine 2020 virtual meeting.
“This free, voice-enabled assistant is capable of relieving radiologists from constant reference of guidelines via web-browser on PACS workstations with expensive screen real-estate,” they said. “We demonstrate the Alexa service is capable of understanding the medical lexicon with accurate recommendations based on pre-programmed user intents.”
According to the research team, RAD-Assistant is comprised of a voice user interface (VUI) that listens for pre-defined user “utterances” to decipher the user’s intent. Visual menus and re-prompts can supplement audio queues. To confirm the basis for the recommendation, the VUI reads back the submitted information.
To determine the skill’s efficacy, the team had two native English speakers verbalize their requests to receive ovarian cyst and pulmonary nodule recommendations to three devices: Echo Spot, Samsung phone, and iPhone. Overall, 143 of 154 (92.9 percent) of requests were successfully completed, and all completed intents resulted in correct Society of Radiologist in Ultrasound or Fleischner Society recommendations. Broken down by subgroup, the ovarian cyst group had 73 of 77 (94.8 percent) successful attempts, and the pulmonary nodule group had 70 of 77 (90.9 percent) success.
The investigators did note two obstacles for RAD-Assistant. Currently, the Alexa platform is not HIPAA-compliant, and further validation with additional accents, dialects, and languages is necessary. However, they noted, expanding to other published algorithms, such as TI-RADS and LI-RADS is a possibility.