Walking Backward to Ensure Risk Management of Large Language Models in Medicine

Abstract

This paper examines in what way providers of specialized Large Language Models (LLM) pre-trained and/or fine-tuned on medical data, conduct risk management, define, estimate, mitigate and monitor safety risks under the EU Medical Device Regulation (MDR). Using the example of an Artificial Intelligence (AI)-based medical device for lung cancer detection, we review the current risk management process in the MDR entailing a "forward-walking" approach for providers articulating the medical device's clear intended use, and moving on sequentially along the definition, mitigation, and monitoring of risks. We note that the forward-walking approach clashes with the MDR requirement for articulating an intended use, as well as circumvents providers reasoning around the risks of specialised LLMs. The forward-walking approach inadvertently introduces different intended users, new hazards for risk control and use cases, producing unclear and incomplete risk management for the safety of LLMs. Our contribution is that the MDR risk management framework requires a backward-walking logic. This concept, similar to the notion of "backward-reasoning" in computer science, entails sub-goals for providers to examine a system's intended user(s), risks of new hazards and different use cases and then reason around the task-specific options, inherent risks at scale and trade-offs for risk management.

Ähnliche Arbeiten