Self-supervised learning for standardizing medical software failure classification

The growing reliance on Health Information Technologies (HITs) in healthcare has increased the demand for accurate failure classifications to improve maintenance and reduce adverse events. This study explores a self-supervised learning approach to standardize HIT failure classification using real-world data (RWD). Leveraging the US Manufacturer and User Device Experience (MAUDE) spontaneous reporting system, over 18,000 unique HIT-related failure records were processed using Natural Language Processing. Afterwards, a classification of failures was derived through prompt engineering and clustering methods. Preliminary results show the coherence and applicability of the approach, highlighting its potential to standardize failure reporting and benchmarking in health technology management. Future refinements in dataset selection, model optimization, and validation are expected to strengthen the classification framework, accelerating the implementation of evidence-based maintenance in clinical engineering.