"The king died, and then the queen died of grief"
The Power of Generative Models for Medical Summaries
In 1927, E. M. Forester was the first to distinguish between a mere succession of information and plot. Essentially, a list of information is a sequence of events in their temporal sequence, while a plot is a recitation of those events with a causal emphasis. The principle is illustrated by Forster's famous example:
“The king died and then the queen died,” is a succession of information. “The king died, and then the queen died of grief” is a plot. The time sequence is preserved, but the sense of causality overshadows it.
In the case of medical records summarization, we encounter the same concept. One way to summarize medical information is by using a list:
● 1/1/1988 The king died
● 1/3/1988 The queen died
The second is using a short narrative: "On 1/1/1988 the King died and two days later the queen died of grief".
The famous mathematical statistician and risk analyst Nassim Nicholas Taleb was referring in his book The Black Swan to Forster’s example.
“...But notice the hitch here: although we added information to the second statement [the queen died of grief], we effectively reduced the dimension of the total. The second sentence is, in a way, much lighter to carry and easier to remember; we now have one single piece of information in place of two. As we can remember it with less effort… This, in a nutshell, is the definition and function of a narrative.”
Consider a more realistic example of a medical record.
The list:
● Pneumonia
● Gastrointestinal problems
● X-ray
● Amoxicillin
● Azithromycin
The narrative:
"Pneumonia was suspected because of a chest x-ray and therefore amoxicillin was given. Because the patient experienced gastrointestinal side effects from the amoxicillin, azithromycin was given instead, resulting in an improvement of the pneumonia."
While the narrative is a bit longer than the list, it is easier to possess, understand and remember. And there is a good reason why.
Lists and data points can be useful, particularly when it comes to normalizing and automating processes. However, we humans are wired to understand, remember and communicate with narratives, not data points. Contexts and casualties not only enrich the data but also mold it to be compatible with the way our brains work.
For an underwriter, claim adjuster or legal entity, the narrative can be more helpful for understanding the medical situation and decision-making.
Until recently, it was impossible to generate narrative-like summaries by AI and these summaries had to be generated manually, which is time-consuming and expensive.
Today, Generative Models allow the creation of a "human language" summary of the data instead of the usual list of data points. This summary includes not only the facts and figures but also the context and relationships between the elements to paint a more complete picture.
By using machine learning algorithms, a summary can be generated from scratch based on the original document's content to empower users to review medical records in a fast, reliable, accurate and efficient way.
To learn more about generative models for medical record summarization, watch our latest webinar. To learn how you can use generative models in your line of business, contact us.