Predicting the Risks of Using a Medical Device Through Machine Analysis of its Operational Documentation

Abstract

The study aimed to develop and implement a machine learning system for predicting the likelihood of incidents involving medical devices, using operating documentation (instructions for use) as a key data source. To this end, the authors designed a dedicated database architecture and curated open data sources on registered medical devices and related incidents, ensuring their effective parsing and preprocessing. A novel aspect of the research lies in the analysis of operating documentation contents: the study identifies textual patterns — potentially evolving over time — that correlate with past incidents linked to similar devices. This approach shifts the focus from modelling the physical interaction between the device and the human body to leveraging linguistic and procedural markers in technical documentation as risk indicators. The proposed machine learning framework is designed to be scalable and continuously retrainable as new data becomes available. An experimental model based on a recurrent neural network (RNN) was trained and validated, achieving an 89 % prediction accuracy on the validation set. These results confirm the hypothesis that specific features in medical device documentation are indicative of elevated risks to safety, quality, or effectiveness. The practical value of the developed solution lies in its potential for broad and cost effective deployment. It can support the identification of higher risk medical devices both at the stage of state registration and among those already circulating on the market. Economically, the approach offers benefits by streamlining the registration process and enabling manufacturers to proactively detect and address design or usage related deficiencies, thereby reducing long term costs.