Ziekenhuis Aan De Stroom (2 of 2)

The foundation for the benefits of semantic interoperability and meaningful primary and secondary use of SNOMED CT-coded clinical data lies in high-quality, up-to-date documentation at the source by clinicians. In reality, however, there are multiple reasons why this proves challenging in many hospitals, even when SNOMED CT has been technically implemented in their electronic health record (EHR) systems. Contributing factors include lack of awareness about the capabilities and benefits of SNOMED CT, and the still underrealized value of semantically structured data resulting in limited clinical buy-in ("what's in it for me?"). A particularly important barrier is the lack of solutions that both empower clinicians to manage registration effectively and simultaneously reduce their administrative burden through user-friendly interfaces. Given that clinicians spend up to 40% of their time on administrative and registration tasks, there is a pressing need for real-time, AI-powered Clinical Documentation Improvement (AI-CDI) tools that seamlessly integrate into natural workflow points, enabling more efficient and accurate documentation.

Clinicians often work in multidisciplinary teams around a patient, making the maintenance and updating of (often complex) hospital-wide problem lists a continuous and coordination-intensive task. In practice, this process frequently falters.

The scope of the current project (funded by the Belgian Federal Public Service of Health) is to develop an EHR-agnostic AI-CDI tool that helps to maintain the accuracy and completeness of SNOMED CT-coded problem lists. The tool is trained to detect diagnoses, history elements, and procedures in free-text conclusions of discharge or referral letters. It compares this information with the existing problem list to identify deltas‚ relevant changes which are then presented to the clinician for validation. Once confirmed, the tool automatically adds, updates, or replaces entries in the problem list (e.g., when a diagnosis has been refined and a more granular concept should be used).

A key strength of this design lies in its integration into the clinical workflow: the moment the clinician is prompted to validate the delta coincides with the moment they have just mentally consolidated the patient's condition, as they've just written the conclusion of the letter.

The project includes a manual annotation phase (where clinicians identify deltas between conclusions and problem lists), a Large Language Model (LLM) training phase, and a proof-of-concept phase in which clinicians assess the model's performance in identifying deltas in new cases. The project also delivers a user interface that supports validation, enabling the clinician to accept, reject, or replace proposed delta entries. Ultimately, the goal is to improve the quality, completeness, and currentnessof a centrally maintained, SNOMED CT-coded problem list at the hospital level.