• Research

AI: a new French algorithm inspired by GPT improves trauma surveillance

Updated on:

In France, one third of emergency room visits are due to trauma. In order to better understand the mechanisms involved and improve treatment, a research team has developed an algorithm capable of classifying emergency room visits due to trauma by analysing clinical reports using artificial intelligence (GPT).

Photo : Scientists have developed an algorithm to better understand trauma, which accounts for one third of emergency room visits. © Unsplash
Scientists have developed an algorithm to better understand trauma, which accounts for one third of emergency room visits. © Unsplash

In France, one third of emergency room visits are due to trauma. In order to better understand the mechanisms involved and improve treatment, researchers from Inserm and the University of Bordeaux at the Bordeaux Population Health research centre, together with teams from Bordeaux University Hospital, have developed an algorithm capable of classifying emergency room visits due to trauma by analysing clinical reports using artificial intelligence (GPT). The results of this project, known as TARPON [1], which has achieved 97% accuracy, were published in the Journal of Medical Internet Research Artificial Intelligence. The results point to the possibility of setting up a national trauma observatory in the near future.

Trauma accounts for 9% of deaths in France, and often affects young people. More than one third of the 21 million emergency room visits each year are due to trauma. It is therefore a major public health concern, with significant health, social and economic implications, for which scientists are working to find solutions.

The idea for the TARPON project, carried out by researchers from Inserm and the University of Bordeaux in collaboration with Bordeaux University Hospital, arose from the observation that for every visit to the emergency room, a report is written by medical staff. These reports contain a wealth of information, including a description of symptoms, the patient's condition and many details about the circumstances in which the trauma occurred.

However, these data sets remain unexploited today, and few statistics on victims of everyday accidents, violence or attempted suicide are available. An observatory exists for road accidents, but it is only complete for fatalities, and most accidents involving cycling, walking or scootering are not included. An analysis of anonymised information from emergency room reports could form the basis of a nearly exhaustive trauma monitoring system.

These reports are unstructured texts written using a mixture of common terms, but also medical and technical terms and abbreviations. To extract interesting information without having to read them all, the research teams developed an automatic language processing technique based on artificial neural networks.

The researchers adapted the GPT artificial intelligence model and trained it with a sample of over 500,000 reports from the adult emergency room at Bordeaux University Hospital [2]. The result is a clinical French-language processing tool that complies with GDPR rules.

With the support of the Health Data Hub, Bpifrance, the Nouvelle Aquitaine region, the French National Agency for the Safety of Medicines and Health Products (ANSM) and the French road safety delegation, the researchers were able to finance the purchase of a powerful server, dedicated to artificial intelligence and installed within the hospital itself. The server was used to implement the algorithm developed by the scientists, which automatically classifies trauma according to its types, with surprising accuracy.

In fact, the method developed by the researchers enables 97% of reports to be classified correctly (compared with 86% using previous methods), as the researchers detail in their scientific article. Thanks to this first step, the study of the data will be able to begin on the Health Data Hub's technological platform between now and the summer.

These results pave the way for the introduction of a national trauma surveillance system, as well as epidemiological analyses of the impact of drug consumption on accident risk, for example. This work should shed new light on important public health issues. In the immediate future, the TARPON project will be extended to around fifteen emergency departments across France.

[1] TARPON: Traitement Automatique des Résumés de Passages aux urgences pour un Observatoire National (Automatic processing of emergency room summaries for a national observatory).

[2] This research meets the requirements of the General Data Protection Regulation (RGPD).