Depending on who you ask, the administrative burden for healthcare professionals ranges from about 30% to even 70% of their time. Given the growing pressure on healthcare—leading to risks of burnout, mistakes, and less attention for patients—it is important to reduce this administrative load. How can we use AI to achieve this?
One of the many administrative tasks is assigning DBC and ICD codes to patient records, primarily for statistical purposes. Currently, specially trained coders manually assign ICD-10 codes to all discharge letters written in a hospital. This means reading the letters and deducing the condition and treatment. It is time-consuming work, and while the number of letters is not likely to decrease, the number of coders is.
Use of Artificial Intelligence
Artificial Intelligence (AI) is now being used to support coders—and in the future, potentially treating physicians—in coding patient records by automating much of the process and providing suggestions in cases where the AI model cannot confidently assign a code.
This is done by training an AI model (using Natural Language Processing) on previously coded records from multiple hospitals, enabling it to learn from as many diverse examples as possible for optimal results, broader applicability, and greater burden reduction. In this way, coders or physicians can save substantial time.
What techniques do we use?
To develop an AI model based on privacy-sensitive data such as patient records and deploy it in an understandable and transparent way, Landscape has developed several essential building blocks:
- Pseudonymization of patient data to work with the data without violating privacy.
- Federated Learning (an element of the Personal Health Train) to learn from multiple healthcare institutions without the need to exchange sensitive data.
- Explainable AI provides users with insight and trust in the model's decisions by showing on what basis these decisions were made. This is used in cases where tasks cannot be fully automated and the model serves as support.
- Active Learning is used when there is a lack of (quality) labeled data (which is often the case). With Active Learning, existing domain knowledge can be efficiently and minimally invasively used to train a model during routine tasks.
- Question Answering for extracting (structured) data from free text.
- Domain-specific language models tailored to healthcare, EHR and/or ECD data, so they can be easily deployed for various applications in this domain with better performance and without additional time investment.
The benefits and results
A model has now been developed that can automatically assign an ICD-10 code as accurately as human coders for about half of the records. For the rest, the coder is provided with suggestions, simplifying the coding task and enabling further training of the model. More hospitals will join, allowing the model to be trained on more representative data for broader applicability, better real-world representation, and improved performance. In 2022, the model can be linked to patient records (EHRs) for use in the National Basic Hospital Registry.
Future applications
The mentioned building blocks are also being used to reduce reporting burdens in elderly care. A pilot is underway for automatic data extraction from patient records to prevent duplicate registration efforts.
Landscape is in conversation with various (general) practitioners and healthcare professionals who are seeking solutions to reduce their administrative burden (through reporting, coding, diagnosing, reading, data structuring, etc.) in daily practice or research. “We are always looking for more stories, experiences, and needs in this area. These building blocks are easily deployable for further applications,” says Erwin Haas of Landscape.
Made possible by:
For this application of AI-assisted ICD-10 diagnosis coding, collaborations are in place with Dutch Hospital Data (DHD), Haga Hospital, Zorgsaam, Maasstad Hospital, Queen Beatrix Regional Hospital (SKB), LUMC, and Slingeland Hospital.
More information:
Interested? For more information, visit the Landscape website (www.wearelandscape.nl). Or contact Erwin Haas.
Also follow our LinkedIn network for data specialists!