Ask any hospital’s CEO: how does your hospital do coding? 

Let’s say you have a chance to ask a hospital’s CEO – how does your hospital do coding? Very likely you won’t be able to get a clear answer or an answer at all.

That’s a crucial question to answer because a hospital’s sustainability and, even survivability, depends on having a medical coding practice that’s robust, ethical, precise, but most importantly efficient and effective.

That’s a very hard question for a human to answer because medical records are the worst case examples of big data, regardless of how you measure, be it by volume, velocity, veracity or variety. One of the hardest things is how every medical condition has nuanced contexts and can be correlated with other conditions in many different ways.

That’s an even harder question because every hospital’s medical coding practices are different too. Everything about the practice can vary, the amount of human resources, background and expertise level of these staff members, how workloads are categorized and distributed, how feedback are relayed, interpreted and digested …

But regardless, wouldn’t it be great if the CEO can say “I know everything about our coding practice, my whole team knows everything about our coding practice, and we all share exactly the same understanding about our coding practice too!”

Well, that’s the reality already.

At VisualizAI, we’ve been using AI methods to tackle the hardest problems in healthcare management. One of the models we’ve built is a Transformer that can learn the medical coding practices in a corpus of medical records. The Transformer model is called DgPg. DgPg is a generative model in that you can ask it a question like: in our hospital, what DX codes tend to get coded before Sepsis occurs and what DX codes tend to get coded after? You can further qualify your question by “over the past three months” or “in our Unit-A vs. Unit-B”. DgPg will literally draw a picture of how the codes would appear and how the codes fall into similarity or categorical groups. In this way, DgPg can show and explain the coding practice about any condition to anyone in that hospital.

The DgPg models trained on different corpus of medical records are different. The coolest result is that DgPg is only 2.5MB in size. Smaller than a photo you snap on your smart phone. 

We believe healthcare management can use a lot of help without getting into the forever debate of whether AI is replacing humans. In this regard, DgPg is just one of the many examples we can provide about how a key task that no one is doing (and everyone is suffering from) can now be done by a small and friendly AI agent.

We are presenting the transformer model DgPg at the ICPRAM conference in Portugal in the coming weeks. The DgPg paper’s title is “Visualizing Medical Coding Practices using Transformer Models.” We’ve also open sourced DgPg because that’s the best way to show an AI assistant can be a real friend!