Risk Analysis Applied to Edge Artificial Intelligence Devices in Healthcare

Johnson, David M.; Malen, Lydia H.; Clemens, Erik; Flor, Bryce

Permalink

https://hdl.handle.net/2142/121830 Copy

Description

Title

Risk Analysis Applied to Edge Artificial Intelligence Devices in Healthcare

Author(s)

• Johnson, David M.
• Malen, Lydia H.
• Clemens, Erik
• Flor, Bryce

Issue Date

2023

Keyword(s)

• Artificial intelligence (AI)
• Edge devices
• Risk analysis
• Real-time data

Abstract

Unique approaches are required to further risk management for Edge devices containing Artificial Intelligence (Edge AI) in healthcare. Regulatory guidance on Edge AI is emerging from the Association for the Advancement of Medical Instrumentation (AAMI), the National Institute of Standards and Technology (NIST), and the U.S. Food and Drug Administration (FDA). These approaches provide the Edge developers with a process to identify hazards and hazardous situations, estimate and evaluate risks, and determine risk mitigations. This research proposes a risk management process for Edge AI devices that builds upon principles of the application of risk management to medical devices standard (ISO 14971:2019) and the guidance on the application of ISO 14971 to artificial intelligence and machine learning consensus report (AAMI CR34971:2022), in combination with the software development for medical devices standard (IEC 62304:2006). The proposed risk analysis method will close the gap in identifying risk in the functionalities and capabilities present in an Edge AI device. The technique demonstrates using an Edge AI design concept consisting of data acquisition, data processing, and AI inference model components. The design concept utilizes the real-time data acquisition of biological signals to provide input to an Edge AI model. The data acquisition component incorporates functionality to digitize sensor data, and the data processing component conditions the signals to match the input characteristics of the Edge AI model. The Edge AI model may be created (trained, tested, validated) on the Edge device or transformed into a more compact model from an existing cloud AI model to ensure that the model performance is unchanged related to processing time, accuracy, sensitivity, specificity, bias, and other quality metrics. This research implements the ISO 14971 process for performing risk management of medical devices and IEC 62304 to classify the software development associated with system risks in the data acquisition and processing components. In addition, this research proposes extending the characteristics related to safety present in artificial intelligence and machine learning models identified in CR34971 to the safety risks presented by the artificial intelligence model residing in the Edge AI component. Finally, we will evaluate Edge AI devices using risk analysis techniques for data management, bias, data storage/security/privacy, overtrust, adaptive or learning systems, post-release monitoring, and other risks inherent in AI systems. This research will create a list of criteria and questions to identify the risks that lead to hazards and hazardous situations resulting from the manifestation of an artificial intelligence risk. We will employ an existing safety risk analysis tool to record the risks, hazards, and hazardous situations. This research will also present an approach to estimate the probability of a hazardous situation using principles adapted from the medical device software development standard (IEC 62304).

Type of Resource

text

Language

eng

Handle URL

https://hdl.handle.net/2142/121830

Owning Collections