From 3M Health Information Systems
Machine learning in computer-assisted coding: It’s just not human
Artificial Intelligence (AI) has creeped its way into so many facets of our lives. Video games, self-driving cars and smart home devices are just a few, and I’m sure if you asked Siri or Alexa, they could name several more. Machine Learning (ML), an application of AI, has been used in computer-assisted coding (CAC) for many years. ML working in conjunction with rule-based coding systems has improved CAC output and increased the speed of code development during a time when the medical coding industry has gone through tremendous change.
Unlike the daunting task of writing rules based on language patterns for each new or revised ICD-10 or CPT code (there were over 700 ICD-10 changes in 2018 alone!), ML utilizes algorithms to identify language patterns within documents to assign codes, and it “learns” from the coder’s interaction with the assigned codes (it is way more complicated than that, but this is a blog, not a dissertation!). Like all technology, however, there are pros and cons to the application of Machine Learning in CAC.
Let’s start with what is great about ML. Well, it’s fast—I mean, really fast. It can learn from millions of documents in a couple of days. It would take a group of expert coders and developers months to write rules to assign codes for the same identified patterns. ML can also process multiple code systems simultaneously. Which leads to the next advantage of ML: cost. It is a lot less expensive to build and maintain than rule-based coding, yet when the two are combined, it makes for a prolific hybrid CAC product. ML also has the ability to identify complex patterns that may not be easily apparent to a human coder, as well as identify inconsistency in human coding patterns more readily than a human audit.
Well, geez, its sounds like we are all out of a job, right? Nah. Artificial Intelligence is just that, artificial. It does not really “think” like a human. Sometimes the math of the ML algorithms can add up to inappropriate use of language evidence, and lead to some pretty silly mistakes (like assigning the code for pulmonary effusion to the word “effusion” in an x-ray of the knee). The ML application and its rule-based counterpart also will make the same mistake over and over again, and will continue to do so until a human goes in and makes adjustments to the algorithm or rule. The ML application faces challenges when coding across patient encounters and when assigning an appropriate code based on multiple documents. It lacks the ability to understand all of the nuances of medical coding—the gray area that coders deal with on a consistent basis.
Despite its shortcomings, ML is an extremely beneficial application of AI and will continue to improve CAC coding output and continue to assist coders in working more efficiently. At the same time, Machine Learning relies on coder input to continue to “learn” and to identify errors. So, put away your applications to that Information Technology School because human coders are irreplaceable!
Gail Barackman is a coding analyst at 3M Health Information Systems.