Taking medical coding to the next level with Artificial Intelligence

Early results show that artificial intelligence (AI) can help medical coders find accurate codes for unstructured data with fewer searches.

By | 2 minute read | March 19, 2021

Woman with a mask on seated at desk and working on a laptop.

AI cannot solve every healthcare data challenge, but strategic applications of AI can help teams improve performance and experience. For example, in life sciences companies, we’re seeing early signs of the potential of AI to help medical coders improve efficiency of clinical trial research by more quickly being able to find accurate codes.

Medical coding is one of many disciplines that must adjust to the increasing volume, variation and complexity of healthcare data. For starters, medical coding is becoming more specific and granular over time. MedDRA is the standard medical dictionary resource for regulatory communication and pharmaceutical medical coders. At the most specific level in its hierarchy, there are more than 70,000 terms to communicate information.1 Medical coders must use these vast lists to search for and select the most appropriate code(s) for each clinical trial participant.

At the same time, incoming data is also increasing. More widespread use of decentralized trials will generate even larger quantities of unstructured data. And, in an aging population where as many as four in ten adults have two or more chronic diseases2, it can be increasingly complex to apply medical codes accurately and efficiently.

Accuracy in medical coding is an important factor in clinical trial operations. Medical codes for things like medical history or adverse events in clinical trial participants can affect researchers’ decisions about trial revisions, as well as gather more accurate information. It also has the potential to help clinicians improve the patient experience by delivering the best quality care throughout the clinical trial. These codes directly impact patient care, for example, capturing medication allergies can help clinicians prevent potential adverse events.

CROs share early findings from AI-assisted medical coding

One of the most time-consuming tasks for medical coders is to identify the right codes for unstructured data in free-form text, also known as verbatims. A verbatim for a patient might say “cannot taste anything,” and the medical coder will conduct searches in dictionaries (such as MedDRA) to associate it with the right code.

Two CROs compared how using AI affected the number of dictionary searches medical coders used to find the correct MedDRA code. When comparing the number of verbatims that required two or more searches (instead of zero or one search), they found that AI assistance helped medical coders find the appropriate code with fewer searches:

Percentage of verbatims that required 2 or more searches3

Without AI With AI
Biostata 90% 19%
Prosciento 96% 16%

AI designed to support the expertise of a medical coder

With the exponential growth of healthcare data and complexity expected to continue, the expertise of medical coders remains in high demand. The US Bureau of Labor Statistics projects 8% growth for the profession between 2019 and 2029, which is faster than the estimated growth rate for other professions.4

AI can support the work of medical coders. Not only can this relieve medical coding fatigue and help improve efficiency and accuracy, but it has the potential to enable medical coders to focus on more meaningful work – such as spotting trends of unusual numbers of adverse effects being reported. AI can help reduce manual tasks in medical coding workflows and enable teams to apply those resources where most needed.

  1. “MedDRA Hierarchy” accessed on the MedDRA website Feb. 1, 2021
  2. www.cdc.gov
  3. Reported by the firms, based on 800+ unstructured verbatims at Biostata and 1,500+ verbatims at ProSciento, and the number of dictionary searches to find the appropriate MedDRA code.
  4. https://www.bls.gov/ooh/Healthcare/Medical-records-and-health-information-technicians.htm#tab-1