Cross-Continental Generalizability of EHR Foundation Model: Predicting Length of Stay in an Italian Hospital �Using US-Trained Embeddings

Giovanna Nicora¹, Francesca Negri¹, Valentina Tibollo², Alberto Malovini², Arianna Dagliati¹,

Lucia Sacchi¹, Federico Sottotetti², Paola Baiardi², Riccardi Bellazzi^1,2

​

¹Department of Electrical, Computer and Biomedical Engineering, University of Pavia, Pavia, Italy

²IRCCS ICS Maugeri Hospital, Pavia, Italy

October 7th, 2026

Paper published (soon) in Frontiers in Digital Health

Foundation Models and Electronic Health Records

• Clinical FMs, pretrained self-supervised on large-scale EHR data, encode a patient's history as a dense embedding reusable across many clinical tasks

Patient’s

medical history

Task-specific Learning

• Length of stay
• Next event
• Mortality
• …

Embedding

EHR

Pretrained

Clinical FM

Wornow, M., Xu, Y., Thapa, R. et al. The shaky foundations of large language models and foundation models for electronic health records. npj Digit. Med. 6, 135 (2023). https://doi.org/10.1038/s41746-023-00879-8

Foundation Models and Electronic Health Records

RQ: Do FMs trained on a specific patient population generalize effectively to fundamentally different healthcare systems?

Case study: length of stay at Maugeri Hospital (Pavia)

• 32 departments spanning rehabilitation, oncology, cardiology and general medicine
• EHR stored in an OMOP database since 2014

Clinical FM: CLMBR-T-Base

• 141M-parameter autoregressive FM
• Pretrained on 2.57M EHRs from Stanford Medicine
• Vocabulary: 26,249 clinical codes

*MEDS: Medical Event Data Standard, github.com/Medical-Event-Data-Standard/meds

{

"subject_id": 3,

"events": [

{ "time":

"2010-02-03",

"code": "LOINC/…",

"value": 7.2 },

…

]

}

US-pretrained embedding vs count-based feature for LoS prediction

Feature engineering

OMOP to MEDS schema and Vocabulary Mapping

• OMOP data is converted to the MEDS schema automatically with an existing ETL
• Issue: some clinical codes have no match in the CLMBR-T-Base vocabulary

Maugeri OMOP database

concept_code

In CLMBR‑T‑Base

vocabulary?

yes

Keep code

no

Find alternative in OMOP

concept_relationship

up to 3

attempts

Relationships tried

• same_as from
• poss_eq_to
• is a
• mapped from
• value mapped from
• maps to

Vocabulary coverage by OMOP table

Latent space analysis

Embedding clusters separate hospitalisations by length of stay

HDBSCAN finds these clusters without LoS labels, yet median LoS triples from cluster 0 to cluster 2

Length of stay – Predictive Performance

• 80% training, 10% validation and 10% test, patient-level splitting to avoid data leakage
• Comparison with baseline model whose covariates are: department, age, gender

Robustness analysis

Pandemic shift

Train on admissions before 2020, test on 2020 onwards

Number of admissions

R² by patient admission history

Embeddings hold up better after 2020 but lose their lead on repeat admissions

Best embedding model

0.65

R² on 2020+ test

ΔR² −0.05

after the shift

Best count-based model

0.608

R² on 2020+ test

ΔR² −0.08

after the shift

Discussion and conclusions

A US-pretrained FM matches count-based models on Italian data, with less feature engineering

What we found

• A US-pretrained FM transfers to Italian data
• Performance similar to count-based models, with less feature engineering
• Shared standards (OMOP, MEDS) make the transfer possible

Limitation

• Patient history covers Maugeri admissions only; stays and visits elsewhere are missing
• Past history seems to add little for the current stay

Next steps

• Predict new clinical tasks
• Explainability (XAI) of embedding-based predictions

Biomedical Informatics Labs

giovanna.nicora@unipv.it