Development and Validation of Deep Learning Transformer Models for Building a Comprehensive and Real-time Trauma Monitoring (Preprint)

Abstract

BACKGROUND In order to study the feasibility of setting up a national trauma observatory in France, OBJECTIVE we compared the performance of several automatic language processing methods on a multi-class classification task of unstructured clinical notes. METHODS A total of 69,110 free-text clinical notes related to visits to the emergency departments of the University Hospital of Bordeaux, France, between 2012 and 2019 were manually annotated. Among those clinical notes 22,481 were traumas. We trained 4 transformer models (deep learning models that encompass attention mechanism) and compared them with the TF-IDF (Term- Frequency - Inverse Document Frequency) associated with SVM (Support Vector Machine) method. RESULTS The transformer models consistently performed better than TF-IDF/SVM. Among the transformers, the GPTanam model pre-trained with a French corpus with an additional auto-supervised learning step on 306,368 unlabeled clinical notes showed the best performance with a micro F1-score of 0.969. CONCLUSIONS The transformers proved efficient multi-class classification task on narrative and medical data. Further steps for improvement should focus on abbreviations expansion and multiple outputs multi-class classification.

Noisy and Unbalanced Multimodal Document Classification: Textbook Exercises as a Use Case

“…For French, two models based on RoBERTa [25] are trained and optimized: CamemBERT [27], trained on the French part of OS-CAR [33] and FlauBERT [20], trained on 24 corpora of various styles collected from the internet. These state-of-the-art French pretrained language models have been fine-tuned on various data for text classification tasks, such as tweets classification [19] or clinical notes classification [4]. Finally, studies on automatic speech transcriptions [14] and digitized texts with optical character recognition (OCR) [18] analyzed the impact of noisy inputs on contextualized word embeddings.…”

“…Most documents are 1-10 sentences and 5-91 tokens long. The dataset is, therefore, composed of short documents compared to the reference datasets 4 . Moreover, depending on the type of exercise, the word and sentence lengths are highly variable.…”