Diego Mellado

Extracting structured information from free-text radiology reports is essential for downstream clinical analysis and decision support. In mammography, this

Background: The application of text mining in radiological reports is crucial for many tasks but principally analyzing and projecting trends to enhance diagnostic accuracy. This is especially important in mammography reports, where early detection is crucial. Large language models (LLMs) provide a viable alternative. They can generate accurate results from a limited set of examples compared with the traditional state-of-the-art models. Methods: This work presents a framework utilizing a general proposed retrieval-augmented generation (RAG) and large language model LLM to create a replicable model capable of structuring mammography radiology reports and extracting relevant concepts associated with the findings. The resulting model is applied to a real-world scenario, using a dataset of mammography radiology reports provided by a hospital. These reports, written by radiologists, are in free text and Spanish. The application of the designed framework is evaluated across several LLMs, and its results are compared to a conventional and specialized NER-type model based on BERT, using a dataset labeled by radiologists. Results: Several models have been implemented and evaluated with the proposed LLM framework. In Named Entity Recognition (NER) tasks using GPT-4, the zero-shot learning scenario achieved an F1-score of 0.80, while the five-shot scenario reached an F1-score of 0.96. This is comparable to the specific-context NER-BERT model, which achieved an F1-score of 0.97. Similarly, in Relation Extraction tasks, we achieved an F1-score of 0.93, a task for which a specialized model was not available. Conclusion: The results demonstrate that large language models (LLMs) can benefit from additional in-prompt examples and achieve results comparable to those of specialized models like NER-BERT. Additionally, this study shows that through a well-defined framework, it is possible to effectively leverage the capabilities of LLMs for specific purposes such as NER and Relation Extraction over clinical text and mammography reports.

Multimodal emotion recognition implies the use of different resources and techniques for identifying and recognizing human emotions. A variety of data sources such as faces, speeches, voices, texts and others have to be processed simultaneously for this recognition task. However, most of the techniques, which are based mainly on Deep Learning, are trained using datasets designed and built in controlled conditions, making their applicability in real contexts with real conditions more difficult. For this reason, the aim of this work is to assess a set of in-the-wild datasets to show their strengths and weaknesses for multimodal emotion recognition. Four in-the-wild datasets are evaluated: AFEW, SFEW, MELD and AffWild2. A multimodal architecture previously designed is used to perform the evaluation and classical metrics such as accuracy and F1-Score are used to measure performance in training and to validate quantitative results. However, strengths and weaknesses of these datasets for various uses indicate that by themselves they are not appropriate for multimodal recognition due to their original purpose, e.g., face or speech recognition. Therefore, we recommend a combination of multiple datasets in order to obtain better results when new samples are being processed and a good balance in the number of samples by class.