03:45 PM Abstract No. 38 The development of a machine learning smart speaker application for device sizing in interventional radiology

Abstract

Smart speakers such as the Amazon Echo and Google Home offer a fundamentally new method for interacting with software. These tools offer unique value in interventional radiology (IR), as a conversational voice interface allows IR physicians to interact with software and retrieve information without breaking sterile scrub. We present a smart speaker application developed specifically for use during IR procedures, using natural language processing (NLP) and machine learning to rapidly provide information about device sizing and compatibility in IR. We have developed a device sizing application for the Google Home smart speaker device. This application processes a human voice query and provides sizing recommendations. It can, for example, specify the minimum sheath size needed for a particular Amplatzer plug, among many other sizing tasks. Size data was acquired in an extensive literature review, with data compiled for 475 IR devices (catheters, sheaths, stents, vascular plugs, etc.). Natural language processing was implemented using Dialogflow, which extracted information of interest from an input query. Logic operations and other data processing were performed using a Python script deployed to the cloud. The conversational application is found to offer an effective tool for device sizing in interventional radiology, achieving excellent performance for a wide range of tasks. Technically satisfactory outcomes were seen using a training data set of 10 samples for each information category and a classification threshold of 0.3. Smart speakers have the potential to offer immense value in interventional radiology, giving proceduralists the ability to access modern information technology while maintaining procedural sterility. This work represents an early step in developing a suite of smart speaker IR tools for intra-procedural support. Future applications may focus on sterile approaches to querying the inventory database, performing dictations during procedural downtime, and accessing key data from the electronic medical record. These tools offer significant value to interventional radiologists and allow us to benefit from the power of machine learning.