Foundation Models for Emergency Medicine

industrial-doctorate
deep-learning
clinical-ai
Industrial doctorate building a multimodal foundation model trained on real-world Emergency Department data — a reusable AI infrastructure for clinical decision support and operational planning. In collaboration with Hospital Sant Joan de Déu.
Keywords

industrial doctorate, foundation model, emergency department, multimodal, EHR, clinical AI, Sant Joan de Déu, UPC, deep learning, decision support

Type Industrial Doctorate (Doctorat Industrial)
Doctoral student Blanca Aleajos
Supervisor Alexandre Perera Lluna
Company partner Hospital Sant Joan de Déu, Barcelona
University Universitat Politècnica de Catalunya (UPC)
Reference futur.upc.edu/42167932
Status Active

Context: The Emergency Department as a data challenge

The hospital Emergency Department (ED) is one of the most data-rich and decision-intensive environments in clinical medicine. Patients arrive with an enormous variety of presentations — from paediatric fever to polytrauma — and clinicians must triage, diagnose, and act quickly on incomplete information. At the same time, the ED generates vast streams of heterogeneous data: triage notes, vital signs, laboratory results, imaging reports, medication records, discharge codes, and flow metrics.

This data richness is largely untapped. Most clinical AI tools in emergency medicine are narrow — built for a single task (triage score prediction, sepsis alert, imaging interpretation) and a specific patient population. They are not designed to transfer across institutions, patient groups, or clinical questions, and they require fresh labelling effort and retraining for each new application.

Foundation models offer a different paradigm: train a large, general-purpose model on the full breadth of available data, then adapt it efficiently to specific downstream tasks. In language and vision, this approach has transformed what is possible with limited labelled data. The question this industrial doctorate asks is: can the same approach work for Emergency Department clinical data?

Research: a multimodal foundation model for the ED

Hospital Sant Joan de Déu is one of Europe’s leading paediatric centres, combining a large clinical volume with a strong tradition of translational research. Its Emergency Department handles a high-complexity, high-diversity paediatric population — an environment that generates the kind of rich, longitudinal, multimodal data that foundation model training requires.

This industrial doctorate, in collaboration between B2SLab (UPC) and Hospital Sant Joan de Déu, aims to build, evaluate, and deploy a multimodal foundation model trained on real-world ED data.

NoteWhat “multimodal” means in this context

Emergency Department data does not come in a single format. A complete patient encounter generates:

  • Structured data: triage codes, vital signs at presentation, laboratory values, imaging orders, medication administrations, length of stay
  • Free text: nursing notes, physician assessments, discharge summaries
  • Temporal sequences: the ordered chain of events from arrival through disposition
  • Categorical codes: ICD diagnoses, procedure codes, specialty referrals

A multimodal model learns joint representations across all of these simultaneously — capturing relationships between, say, a triage note and the sequence of lab tests that follows, or between a vital sign pattern and the eventual diagnosis.

The foundation model approach

Rather than building separate models for each clinical prediction task, the strategy is:

  1. Pre-train a large model on the full breadth of unlabelled (or self-supervised) ED data, learning general representations of patient state across all modalities.
  2. Fine-tune the pre-trained model with minimal labelled data for specific downstream tasks: disposition prediction, deterioration risk, length-of-stay estimation, resource demand forecasting.
  3. Evaluate reusability: the core claim of a foundation model is that the pre-trained representations transfer — a new task should require only a small additional training step, not a rebuild from scratch.
TipWhy this matters for hospitals

A reusable foundation model infrastructure changes the economics of clinical AI deployment. Instead of commissioning a separate AI project for each new clinical question, a hospital can maintain a single pre-trained model and rapidly adapt it. For a centre like Sant Joan de Déu — which serves a complex, high-acuity paediatric population across many specialties — this is a meaningful operational advantage.

Beyond prediction tasks, the shared representation space can support operational decision support: staffing planning, resource allocation, patient flow optimisation — areas where ED management teams face daily pressure.

Connection to B2SLab’s broader EHR research programme

This doctorate sits within B2SLab’s sustained programme of deep learning for longitudinal clinical data, which includes:

  • DARE: a transformer encoder for predicting Type 2 Diabetes outcomes from 200,000+ patient records (Expert Systems with Applications, 2025)
  • Deep survival analysis for COPD: joint prediction of hospitalisation and death in 150,000+ patients (arXiv, 2025)
  • BERT for EHR: language-model pretraining on clinical sequences (EMBC 2024)

The emergency medicine foundation model extends this programme into a new clinical domain, a new data modality mix, and a new institutional context — while building on the same architectural foundations (transformers, attention, sequential clinical representations) that have proved effective across the group’s prior work.

About the partners

Hospital Sant Joan de Déu is one of the most important paediatric hospitals in Europe, located in Esplugues de Llobregat (Barcelona). It combines a high-volume clinical operation with a large research institute, and is a recognised reference centre for rare diseases, oncology, neurology, and neonatology. The hospital’s commitment to clinical research and digital health makes it a natural partner for AI development grounded in real clinical data.

B2SLab / IRIS-UPC contributes deep learning methodology, transformer architectures, and experience building and evaluating clinical AI models on large-scale real-world health record data.

Industrial Doctorate programme funded by the Generalitat de Catalunya (Secretaria d’Universitats i Recerca).