page contents

Linkous Health

header photo

AI’s Problem with History

NOTE: While I am a proponent in the use of advanced technology, such as AI, automation, and robotics, in the delivery of healthcare, effectively supporting such change requires identifying areas where we can and must make improvements.

In 2018, Amazon announced that it disbanded an AI tool used to recruit new employees in the company.  It turns out the application showed bias in scoring applicants.  According to the article: “That is because Amazon's computer models were trained to vet applicants by observing patterns in resumes submitted to the company over a 10-year period. Most came from men, a reflection of male dominance across the tech industry…In effect, Amazon's system taught itself that male candidates were preferable. It penalized resumes that included the word "women", as in "women's chess club captain.”Anchor

This reflects a problem with some approaches to machine learning.  Many applications are based on absorbing and manipulating huge volumes of data.  The program’s intelligence derives from what it learns.  If it learns only from skewed data, it will reflect that in its results.  Of course, emerging areas of artificial intelligence includes more than just rote learning.  For example, the use of neural networks could allow the application to discover new ways to approach a problem heretofore unknown.

How is this applicable to healthcare?  Let’s look at ulcers.  For many years the medical profession taught and understood that ulcers were caused by stress, spicy foods and too much acid.  Then, in the mid 1980’s a gastroenterologist discovered that a bacterium, Helicobacter pylori, is often the culprit and its long-term effect could even lead to cancer.  The discovery met with widespread opposition as common wisdom, based on historical thought, did not include reference to bacterial infection.  Similar controversy surrounded the discovery of the cause of mad cow’s disease (prions) and the idea of recessive and dominant genes in heredity, which overturned a substantial history regarding simple blended inheritance.  Because the discoveries were new and not drawn from existing studies, they were ignored.

Now suppose a machine learning diagnostic-and-treatment application tackled one of these issues at a time prior to the discoveries.  The application could churn out lots of intelligence based on current thought and prescribe commonly reasoned treatment (e.g. reduce stress for an ulcer) which may not have any affect at all.

This is not a lesson to distrust AI.  But to remind us that must make way for breakthroughs and new discoveries into the algorithms, even if they challenge the status quo.  We are still in the initial stages of this new revolution which needs constant monitoring, quality control of results and continuous input of new data, which are critical to success.

Healthcare professionals have strict rules about continuing education.  The same should be true for medical computer algorithms of the future.

Go Back


Blog Search


There are currently no blog comments.