Health Information Technology Adverse Events Identification and Classification with Natural Language Processing and Deep Learning

The main topic of this work is to develop a framework to extract Real-World Evidence through Natural Language Processing (NLP) and Neural Networks. An initial literature analysis has been performed, from which it clearly emerges that adverse events concerning Health Information Technology (HIT) are gradually growing over time. The goal of the proposed framework is to automatically identify adverse event reports related to HIT, aiming to support Health Technology Assessment and Post Market Surveillance as outlined in European Regulation 2017/745 on Medical Devices. The designed model uses a pre-trained version of ClinicalBERT, additionally fine-tuned on 3, 705 adverse events reports extracted from the FDA MAUDE database, which had been previously manually labelled by experts. Results show better metrics than other existing HIT adverse events reports text classifiers based on non-BERT models, performing with an accuracy of 0.9906, precision of 0.9840, recall of 0.9973, and F1 score of 0.9906.