SOLUZIONI DI DEEP LEARNING PER IL RICONOSCIMENTO E LA NORMALIZZAZIONE DI DIAGNOSI DA REFERTI CLINICI

This thesis work dives into the dynamic field of Natural Language Processing (NLP), with a particular focus on its applications within the health care sector. The advent of deep learning technologies has catalyzed the development of advanced language models, including BioBERT, that promise unprecedented understanding of human language. This research focuses on adapting and optimizing BioBERT to overcome the limitations of the previous ARGO 2.0 model, specifically targeting the analysis and identification of pathologies in medical reports, with a focus on different types of cancers. Through an in-depth look at the foundations of NLP and the revolutionary transformer architecture, the paper illustrates how these technologies are transforming data management in the healthcare setting, significantly improving the effectiveness and accuracy of diagnoses. The thesis evaluates the efficiency of BioBERT in recognizing named entities (NERs) specific to the medical domain, comparing its performance with that of the ARGO 2.0 model and demonstrating advantages in terms of F1-score, Recall and Precision. Despite its successes, the model has shown some limitations in generalization across a wide range of diseases, suggesting the need for additional data and more extensive training to increase its understanding capability. In response to these challenges, the research proposes future developments focused on enriching the training dataset and expanding the model’s recognition capabilities, with the ultimate goal of implementing a solution capable of accurately identifying any type of cancer described in medical reports. This thesis not only testifies to the transformative impact of advanced NLP in the field of medicine but also opens new perspectives for the use of these technologies in clinical decision support, highlighting the potential for significant improvements in the personalization and efficacy of disease treatment.