NLP on the fly in Capio Denmark

Bodil Skodborggaard is a Senior Analyst and debtor team leader at Capio and works with turning procedures into payment. “Our current way of working is not optimal. For example, we could see that when we returned after the holidays, there were notes everywhere. If we could improve it, this would be useful for many”.

Cross-fertilizing ideas

2021. In Denmark, Chief Transformation Officer Peter Baltzar Fredriksen and Data Manager Veniamin Tzingizis at Capio Private hospital implemented an administrative robot for invoicing after magnetic resonance imaging and mammography procedures.  

In a Swedish project at Capio Närsjukvård Stenungsund the same year, the team explored the use of voice-to-text engines and coding.

When the Innovation hub heard that Peter and Veniamin were looking at implementing a voice-to-text solution already widely used in the Danish healthcare sector, they called Peter to share their experience with testing voice-to-text and coding combined in Sweden. He liked the idea and that was the starting point for the project: Learning from both good and bad experiences in Sweden to get-go in Denmark. A project team was started!

Understanding the project design

The project team needed to understand the current flow - which tasks are done, and by whom. Then to see which systems are used to do these things, to know which parts and stakeholders are affected by the new ways of working. Finally, they explored different options of how the digital voice transcription and analysis could take place, and which puzzle parts needed to be fitted into each other.

With that puzzle laid, the design of the Proof-of-concept test (“POC”) was ready. The POC is trying to prove the feasibility of having a complete solution in place; from A to Z, where the chain is made of voice-to-text and coding algorithms, delivered via Microsoft Azure and PowerApp solutions, and integrated with the EMR.

They aimed for a small start with just a few involved healthcare workers, in a small segment of the Hellerup hospital, using historical data, not live ones. The diagnosis or treatment code had already been set when referred to Capio via the public hospital or the insurance company. The coding part was then to discover and add (if relevant) new diagnosis and treatment codes in addition to those already set. Lastly, it applied a set of billing rules to determine what billing codes to use under which customer regiment.

Two pieces of NLP

NLP is an umbrella term that spans a lot. In this project, two things that fall under that term are used:

1) “Voice-to-text” engine transcription. This means that algorithms are used to capture voice or sound files and transcribe those sounds into words in a text. This tech is well established and does not usually need that much training.

2) “Coding AI” engine analysis. These are algorithms to understand the text created by the voice-to-text engine. First, the algorithm needs to go through a period of “statistical training”, where a model uses lots of data to predict how a specific code usually shows up in connection with a specific set of words or phrases. After those relationships have been detected and established, the algorithm is further trained by directing and giving it human feedback to learn if its predictions were correct.

Nuance, considered the gold standard of the world, was used for voice-to-text transcription. It was a pretty safe bet.

For the analysis of the text, which was a bit more uncertain, the project started out with testing one vendor and quickly thereafter added another vendor to the POC, to see the pros and cons of their different approaches to coding. These two were trained and taught both by machine learning for the basic skills, and by humans to feedback and tell it when their first suggestions were wrong.

After testing the process as a whole and assessing the performance of both design approaches, the project team evaluated the first step.

Learnings made

They could confirm that Nuance is a good voice-to-text generator. Even in Danish.

When it came to the coding elements, they performed a thorough assessment of them.

Veniamin Tzingizis explains: “The algorithm needs to find the correct codes exactly, not too many nor too few. It has to be a medically certain procedure.”

Besides assessing the technical result, the provider’s agility and the business case were also analyzed. Corti seemed to be the best puzzle piece.

But to really make it more time efficient, there is a need to integrate these findings into the Electronic Medical Record (EMR). In this case X-Medicus. This is now in the making and will probably be ready before the end of the year. “It looks very promising”, says Niklas Sundler from the Innovation hub.

One key point is to focus on the purpose of the idea. When it comes to documentation, texts, and codes, the medical secretaries are the champions with knowledge about the total process. They are crucial players, and it is important to take part in their experiences and work closely with them in a project such as this one.

Bodil Skodborggaard gives an example: “We need to see that it’s possible for the computer to read and understand what’s been done, what’s booked on the invoice, and that we can improve this. Getting everything done to invoice is complex because there are different rules.”

The involved doctors have been positive so far.

What’s next?

Now that the technology providers have been chosen, a more thorough process for the Hellerup hospital’s operations will be designed. Starting with one surgical flow, the process will be tested, evaluated, and subjected to continuous improvement, and then maybe expanded to other parts of the operation. It’s up to the hospital to decide the order.

“I think it would help bort the secretaries and the algorithms if there was a mandatory step, in the beginning, to tell the relevant code for the visit”, says Bodil Skodborggaard.

The team believes that if they can find a way to automate 80 percent of the generated documentation good enough to read, code, and reimburse, it would definitely make an impact. The business case is about efficiency and saving time. But with help from what they learn from this, they can also get results in quality aspects that have not been reachable before:

“Then it will be really exciting – doing things we have not yet been able to do!” Michael Adler from the Innovation hub explains. “If this will be used the way it’s planned, we will have an option to record raw voice data, and coded textfiles, and might start to be able to start segmenting the data into different subgroups. We could for example be able to design visualization and highlights depending on the patient group, doctor, and their needs”, he exemplifies.

Niklas Sundler digs into the innovation hub’s interest in the project:

“Apart from the administrative side, there is a technical part of this, which we as an innovation hub will monitor closely. We can use learnings from this project and incorporate them into other needs, test, and learn more… We love to spread learning! For example, we will combine parts of this into the work on a Nurses’ portal or other needs we see from our Living Labs.  It’s like a flexible tool, a Swiss army knife.”

Working with innovation is stimulating

Interesting learnings have been made so far. The Nuance platform was able to be integrated into the Power app. It was low coding, which means pretty simple, but had not been done yet. Microsoft now uses this project as a reference case to inspire others globally.

“I really like doing something that is innovative in the field and putting our company in a very good position. I also like that it is complex and challenging, a big puzzle with many pieces” Veniamin Tzingizis comments.

Bodil Skodborggaard continues: “I like to be in a group and participate with my skills and to feed different things into this.

Michael Adler really appreciates the teamwork done: “The importance of having a dedicated team cannot be underestimated. Without Veniamin, Peter, Bodil, this would never have worked. But he also has another sweet spot in the project:

“As a Dane, I am pleased to see that we can participate in showing a clear value even for such a small language as Danish. Cool stuff is not exclusive to the speakers of world languages anymore, even other languages can benefit from this as well, and with quite small efforts.”

From talking to the project group, it’s easy to tell that this project can create a lot of value, and the momentum is building up. So much can be done, once the foundations are in place.