Figure 5
in computer sciences and would never call myself an expert . But the next place I look is to see if any of the authors have a degree in the computer sciences , preferably with prior citations in the space . I always check to make sure that they have never tested their algorithm using the same data that was used for training . That is to say , you cannot show a computer a million images of a cat labeled as such and then show a thousand images from that same data set to confirm that the machine has learned what a cat is . Computers excel at just such tasks ; the images that you show to confirm the machine understands “ cat ” must be new images the computer has never evaluated to confirm it can apply what it has learned .
Following this , I check for sensitivity and specificity and issues with their assessment . Currently , one of the safeguards against the concerns surrounding machine learning is that humans are frequently expected to review findings . However , a possible error we could be blind to can be illustrated by another example . Suppose your job is to find animals in images . There is a data set containing images of zebras , lions , elephants , trees , lakes and mountains . Now
ARTIFICIAL INTELLIGENCE suppose the animal images were all labeled as zebras . To me , that ’ s no big deal . As long as the machine can show me , without fail , images of every zebra , lion or elephant , I can re-label them correctly . But what if someone forgot to label the lions as animals . No matter how good the computer scientist or the machine is , the sensitivity for lions would be exactly zero . Moreover , as a researcher I might never even know that there were lions in the data set at all . In this context , when I read machine learning manuscripts , I try to ask myself what issues might affect the sensitivity .
Lastly , as with anything we must ask ourselves what benefit adding a machine into the assessment provides . No single practitioner can answer that question . Low hanging fruit is likely wherever there is a time-consuming process that people hate doing and where errors make small differences . What defines time consuming , hatred of an activity , and the size of an error are things that we will have to decide as a profession . In an ideal future , physicians will help guide technology to paradoxically decrease the amount of time we spend in front of our computer screens . Until then , medicine can keep watch for errors from any corner , regardless of the intelligence of the source .
References
1 . Schober P , Vetter TR . Linear Regression in Medical Research . Anesth Analg 2021 ; 132:108-109 .
2 . Ding Z , Shi H , Zhang H , et al . Gastroenterologist-Level Identification of Small-Bowel Diseases and Normal Variants by Capsule Endoscopy Using a Deep-Learning Model . Gastroenterology 2019 ; 157:1044-1054 e5 .
Dr . Rogers is an Assistant Professor in the Division of Gastroenterology , Hepatology , and Nutrition at the University of Louisville .
Did you know ? GLMS recently launched a text messaging service to keep you updated while you ’ re on the go . This will provide information , announcements and opportunities exclusive to GLMS members .
To sign up to receive these messages , please text the word GLMS to ( 833 ) 509-1891 .
Note standard texting rates may apply .
APRIL 2022 17