Gastroenterology and Artificial Intelligence: 4th Annual Global Artificial Intelligence Summit (On-Demand | August 2022-Learning from Other Disciplines

Pdf Summary

In this document, Dr. Charles E. Kahn, Jr. discusses the topic of learning from other specialties in the context of artificial intelligence (AI) in healthcare. He highlights the importance of understanding the "ground truth" of AI algorithms, including who annotated the data and how inter-rater and intra-rater variability were assessed. He also mentions the need for blinding and adjudication of discrepancies in the annotation process.<br /><br />Dr. Kahn also references the use of the DICOM standard, which is an international standard for medical images and related data. He points out that there are thousands of imaging devices and billions of DICOM images in radiology.<br /><br />The document also references a study that found a decrease in algorithm performance when validated on external datasets. According to the study, 81% of algorithms reported a decrease in external performance, with 24% reporting a substantial decrease.<br /><br />Dr. Kahn concludes the document by providing a list of ten questions to consider when evaluating AI applications in healthcare. These questions cover topics such as the intended problem and target audience of the application, potential benefits and risks, validation of the algorithm, integration into clinical workflow, IT infrastructure requirements, compliance with regulations, return on investment analyses, maintenance and user training, and handling of potential malfunctions or erroneous results.<br /><br />Overall, the document highlights the importance of understanding the limitations and considerations related to AI in healthcare and emphasizes the need for thorough validation and careful implementation of AI algorithms in clinical practice.

Asset Subtitle

Charles Kahn, Jr., MD, MS

Keywords

learning from other specialties

artificial intelligence in healthcare

ground truth of AI algorithms

data annotation

DICOM standard

algorithm performance

external datasets

evaluation of AI applications

clinical workflow integration

malfunctions in AI algorithms