Landscape: AI for administrative load reduction in the care sector

Patient files
Depending on who you ask, the administrative load for care professionals ranges from 30% to as much as 70% of their time. Given the growing pressure on healthcare and the risk of burnout, errors and less attention being paid to people as a result, it is important to reduce the administrative load.

One of the many administrative tasks is assigning DBC and ICD codes to patient records, largely for statistical purposes. At the moment, there are trained specialist coders who manually assign ICD-10 codes to all the discharge letters written in a hospital. This means reading the letters and deducing what the condition and treatment were. This takes a lot of time and not only is the number of letters not dropping, but the number of coders is.

Using artificial intelligence

Artificial Intelligence (AI) is now being used to assist coders, and in future also the treating physicians, in coding patient records by automating a large portion of the task and making suggestions in cases where the AI model itself cannot assign a code with enough certainty.
This is done by training an AI model (using natural language processing) with previously coded records from multiple hospitals so that it can learn from as many different examples as possible and get the optimum results, making it more widely deployable and reducing the load more. This saves a substantial amount of time for coders or doctors.

What challenges does it solve?

To be able to develop an AI model based on privacy-sensitive data such as patient records and use that model understandably and transparently, Landscape has developed several essential building blocks:

  • Pseudonymisation of patient data to be able to process it without violating privacy.
  • Federated learning (an element of the Personal Health Train) allows the system to learn from multiple healthcare institutions without them having to exchange sensitive data.
  • Explainable AI gives end users a clear picture and confidence in the choices made by a model, by explaining what these choices were based on. This can be used for cases in which tasks cannot be fully automated and the model is used for support.
  • Active learning is used when there is insufficient high-quality labelled data (as is often the case). Active learning allows existing domain knowledge to be used efficiently and in a minimally invasive way to train a model during regular work activities.
  • Question answering for extracting (structured) data from free text.
  • Domain-specific language models that are tailored to healthcare, data from electronic patient records and/or client dossiers, so that they can be used easily for various purposes in the domain without taking extra time and with better performance.

The benefits and results

At present, a model has been developed that can automatically assign an ICD-10 code to about half of the records as accurately as human coders can. For the remainder, the coder is given suggestions, simplifying the coding work and further training the model. More hospitals will join so that the model can be trained on more representative data for even wider application and a better representation of reality and thereby even better performance. In 2022, it will be possible for the model to be linked to patient records (EPRs) for use in the National Basic Hospital registration.

Future applications

The building blocks mentioned are also being used for load reduction in reporting in elderly care. To that end, a pilot is ongoing for automatic data extraction from patient records to prevent double registration work.

Landscape is in discussions with various general practitioners, doctors and care professionals who are looking for solutions to reduce the administrative load in their daily practice (reporting, coding, diagnosing, reading, data structuring, etc.).

“We’re still looking for more narratives, experiences and needs in this area,” says Erwin Haas from Landscape. “These building blocks are easy to implement for other uses.”

Parties involved:

The collaboration for this application of AI-supported ICD-10 diagnosis coding was with Dutch Hospital Data (DHD), Haga Hospital, Zorgsaam, Maasstad Hospital, Koningin Beatrix Regional Hospital (SKB), LUMC and Slingeland Hospital.

More information:

If you are interested, you can visit the Landscape website (www.wearelandscape.nl) or contact:
• Erwin Haas, erwin@wearelandscape.nl, +31 (0)6-24412942
• Frédérique de Paus, frederique@wearelandscape.nl, +31 (0)6-26270522

Share with

More information

Building blocks

The NL AIC collaborates on the necessary common knowledge and expertise, resulting in five themes, also called building blocks. Those are important for a robust impact in economic and social sectors.

Sectors

AI is a generic technology that is ultimately applicable in all sectors. For the development of knowledge and experience in the use of AI in the Netherlands, it is essential to focus on specific industries that are relevant to our country. These industries can achieve excellent results, and knowledge and experience that can be leveraged for application in other sectors.

    Become a participant

    The Netherlands AI Coalition is convinced that active collaboration with a wide range of stakeholders is essential to stimulate and connect initiatives in Artificial Intelligence. Within fields of expertise and with other stakeholders in the ecosystem to achieve the most significant result possible in the development and application of AI in the Netherlands. Representatives from the business community (large, small, start-up), government, research and educational institutions and civil society organisations can participate.
    Interested? For more information, see the page about participation.