Autonomous AI-Driven FMEA Optimization For Safety And Reliability In Healthcare

SUMMARY & CONCLUSIONSIn the quest to enhance design safety and reliability throughout the product lifecycle, this paper explores the infusion of Failure Modes and Effects Analysis (FMEA) with intelligent autonomy, transforming it into a dynamic, self-sustaining system. In this scenario, a vital feedback loop connects post-market surveillance, field data, and proactive design adjustments, with Failure Modes and Effects Analysis (FMEA) traditionally serving as a key tool to identify and mitigate risks within this cycle. As product complexity grows, it becomes increasingly important to enhance FMEA processes by incorporating dynamic, ongoing feedback from real-world data sources.We propose to explore the integration of FMEA with autonomous capabilities, employing advanced techniques to automatically identify and assess failure modes, thereby improving design safety and reliability throughout the product lifecycle. The process incorporates item parsing and AI-driven feature extraction to determine traditional and emerging failure modes and causes. Aiming at transforming raw data into a set of numerical features that can be used for modeling. To work with such an intelligent system, it is crucial to define the minimal domain knowledge while ensuring optimal contextual input for efficiency and accuracy.Within the domain of the healthcare industry, we address optimization challenges such as noise reduction and differentiation of similar failure modes, ensuring the AI system reliably avoids inaccurate mode approximations. This in turn demonstrates the capability of the system to autonomously learn and adapt over time while accounting for the emerging threats arising from the use of LLM models such as hallucinations and sycophancy. Our findings suggest a future where models autonomously identify their knowledge needs, optimizing the integration of data analytics with AI-driven FMEA to create a safer and more reliable product ecosystem.In closing, we are demonstrating how this approach can be enhanced by utilizing domain-specific languages (DSLs) to link failure modes with system definitions.