Bringing structure to clinical data: Supporting pediatric emergency care with Tune Insight’s methodological LightLM approach

The European TI4PEC project is part of a broader dynamic driven by PERN (Pediatric Emergency Research Networks). This international association brings together dozens of hospitals across the globe involved in pediatric emergency research.

The project’s goal is clear: to allow facilities in different countries (starting with hospitals in Padua, Bologna, and Trieste in Italy, and Lausanne in Switzerland) to share their clinical data in complete compliance with the strict regulatory, ethical and technical constraints that define the healthcare sector as a whole, and pediatrics in particular.

But before any federated, interhospital analysis can be carried out, data must first be exploitable locally.

This is where Tune Insight, in collaboration with the University of Padua, comes in, contributing its expertise in bringing structure to unstructured clinical data. Thierry Bossy, Machine Learning Engineer at Tune Insight, tells us more about the model developed to meet this initial need.

Hidden data: the free-text challenge

In pediatric emergency departments, physicians write their notes in narrative form. Symptoms, clinical observations, suspected diagnoses… All in free text.

This information holds significant medical value, but leveraging it effectively in analytical systems remains challenging. “We need to structure this data in a common language if we want to be able to carry out advanced, distributed analyses,” explains Thierry.

It’s a well-known problem in healthcare: the information exists, but there is no standard codification to facilitate statistical processing or the training of models.

In the TI4PEC project, diagnoses are documented in Italian—adding a whole new layer of complexity to the mix.

The Light Language Model methodology: a specialized, modular approach

The Tune Insight team developed a local process chain based on Light Language Models (LightLM). These lighter versions of externally-hosted large language models focus on specific tasks, and are directly integrated into software infrastructure. The approach is methodical and sequential.

“The diagnoses were in Italian, so the first step was to translate them into English using a locally-run open-source light model. The translation is a means, not an end in itself. It serves to make the text compatible with the medical classification models at our disposal. Once translated, the diagnoses are processed by a model that can match the text to a standardized diagnostic code, the ICD-10 (International Classification of Diseases, 10th Revision). It’s like building blocks. We put together the blocks we need. The first one is translation, then classification,” continues Thierry.

The final result? A standardized medical code that can be used for:

• Statistical analyses,
• Comparisons across departments,
• Training machine learning models.

An integrated element of Tune Insight, ensuring traceability and security for healthcare facilities

Light Language Models are not external tools. They are integrated into the Tune Insight platform so that health data does not need to be transferred.

This means:

• Processing is run locally,
• Facilities maintain ownership of their data,
• Blocks can be combined,
• And no data science expertise is required to use them.
  

“The hospital processes its own data locally. Tune Insight provides the infrastructure and the modules, but never has access to the clinical data,” clarifies Thierry.

The decision to use a modular infrastructure is not simply a technical one. This approach also meets a key regulatory and operational requirement: making processing auditable. As Thierry reminds us, “The person who establishes the data preparation chain is not always the person who runs it. But they need to know exactly what has been done to the data.”

In reality, this means that:

• Each transformation is traceable: which modules were run, in what order, and with what settings.
• Each brick is replaceable: a module can be improved or updated without impacting the entire chain.
• The method is reproducible: the same pipeline can be deployed in another facility, or even adapted to another language.

In the medical sector, this transparency is not incidental—it’s essential.

* In accordance with security and compliance requirements, the pipeline development and validation phases were not carried out with sensitive clinical data directly. This experimental phase allowed the modules to be finely tuned to the specific needs of the University of Padua, in particular with regard to the typology of pediatric diagnoses processed.