Epidemiological surveillance of the covid-19 pandemic period by real-time automatic classification of clinical notes from the emergency call centres using Transformer-type artificial neural networks. – COSAM

The emergency call regulation centres of the SAMU centre 15 are a source of information that should be used to set up real-time epidemiological surveillance. We propose to take advantage of our work to develop and validate an automatic classification tool (AI model GPT-2) that we have adapted to classify clinical observation reports available in free text during emergency room visits. We propose here to apply it to the classification of calls to the centre 15 in order to monitor general mental and physical health indicators, as well as calls suggesting SARS-CoV-2 infections in the pre- and post-confinement periods, so as to participate in epidemiological surveillance in the post-confinement period.
- A first phase implemented very quickly (15 days, submitted on the CARE platform) consists of the creation of a learning sample of the model by selecting clinical notes by keywords.
However, the keyword search is a first step which does not allow the identification of complex clinical concepts. Our current work shows that this will be possible using a model of natural language processing by artificial neural network (GPT-2 Transformer model, see our first publication Xu et al. 2019). The results of this first phase will be used for temporal and geographical analysis for the public health surveillance and the early detection of epidemic clusters.

- A second phase corresponds to the manual coding of clinical notes allowing on the one hand the validation of the first phase, and on the other hand the constitution of a second training sample based on human coding.
There is an urgent need to validate this tool to allow relevant real-time monitoring based on the contents of these clinical notes. This tool could then be integrated via the Health Data Hub to a large sample of emergency call centres in the country in order to build a surveillance system capable of monitoring the post-confinement period.