Artificial intelligence (AI) is to be used in various areas to provide support in the shock room. Intelligent information management is intended to relieve hospital staff so that they do not have to concentrate on information management in time-critical situations. To this end, researchers from the Fraunhofer Institute for Intelligent Analysis and Information Systems IAIS have teamed up with partners such as the University of Witten/Herdecke, the University Hospital of Aachen and the Cologne-Merheim Hospital as part of the “TraumAIInterfaces” project and developed two AI prototypes. The project was funded by the Federal Ministry of Health.
Advertisement
Live display in the shock room
A prototype of the Fraunhofer IAIS displays live information on treatment support in the shock room via a front end. Information includes the ABCDE scheme – a strategy for assessing seriously injured patients – emergency medical handover and other treatment information. In another screen, further information can be recorded manually, but training data can also be recorded. The voice recognition software Whisper from OpenAI is used.
The small number of voice recordings was compensated for by using GPT-4. “An LLM agent was developed based on GPT-4 and LangChain, which as an environment receives the current state of the frontend and the output from the speech transcription,” says the white paper. The agent has access to the REST interface from the shock room frontend.
The AI should take over the organizational tasks and automatically record the information exchanged orally and measured electronically.
Documentation tasks
Documentation tasks that arise during treatment should also be taken over. To this end, Fraunhofer IAIS has developed another prototype for a form assistant. Based on the discussions in the shock room, relevant information should be included in the “TraumaRegisterDGU” standard form, an instrument of the “TraumaNetwork” of the German Society for Trauma Surgery to improve care for seriously injured people.
“A particular amount of work was invested in designing the “prompts” for the agent so that the model provides reasons for the individual answers and only refers to factually correct information,” say the researchers.
The researchers explained details in their white paper “Artificial Intelligence in the Shock Room: How Agents and Foundation Models Help in the Care of the Seriously Injured.”
challenges
However, there are also challenges in developing and implementing AI systems in the trauma room, such as the generation and use of data. “The acquisition of a sufficiently large amount of training data for AI models from the machine learning spectrum that are developed for patient-related processes (e.g. in the shock room) is generally confronted with considerable ethical concerns. Established data generation strategies such as generation,” says about this in the white paper. There are also challenges regarding interoperability and practicality. Therefore, according to Fraunhofer IAIS, the systems should be tested and further developed in the clinic at an early stage.
(mack)