Global Models, Local Friction : Interoperability, Adaptation, and Deployment of Clinical Foundation Models in Swedish Healthcare

Sammanfattning

What does it cost to bring an international clinical foundation model into Swedish emergency care? Foundation models are large AI systems that learn from the time-ordered records of millions of patients to anticipate what is likely to happen to a patient next. The clinical foundation models most relevant to this project have largely been trained on American records and expect each record in one international format — which Swedish hospitals, with their own code systems, do not use. This project approaches that gap as a feasibility study, using one such model (CLMBR-T-Base) as its anchor. The adaptation cost takes three forms. Terminological cost arises when Swedish code systems (ICD-10-SE, KVÅ, ATC, NPU) are mapped onto international standard vocabularies (SNOMED CT, RxNorm, LOINC). Distributional cost arises because Swedish care records a different mix and concentration of events than the American data the model learned from, so the model interprets even faithfully translated codes through expectations tuned elsewhere. Implementation cost is the organisational and regulatory distance to using the model in practice. To measure that cost, a data-processing chain was built to translate Swedish emergency department records into the format the model expects, recording at each step what survives the translation and what is lost. A decisive finding appears when the transformed Swedish records enter the model: fewer than 20% of distinct codes are recognised by the model's vocabulary. The model can still be trained further on Swedish data, a step that would make it more adapted to Swedish care. However, further training on real Swedish records was deferred because the severely limited vocabulary recognition made a full run unlikely to produce scientifically reliable performance estimates. Moreover, a proposal for integration with the standard hospital-record interface (FHIR) is specified but not tested. Our assessment is that, for this model, the total cost of adapting it was too large relative to what the model could offer. What this work leaves behind, however, is not tied to this particular model. Its main contributions are an infrastructure for standardising Swedish emergency department records, a vocabulary mapping with broad coverage, a proposal for integrating the model with the hospital-record interface, and a three-part framework for adaptation cost.