MeDeT: Medical Device Digital Twins Creation with Few-shot Meta-learning

TL;DR

This paper tackles the challenge of rigorously testing healthcare IoT systems when medical devices continuously evolve, by introducing MeDeT, a meta-learning-based pipeline for generating and adapting digital twins (DTs) of medical devices. MeDeT trains device-specific DTs with few-shot learning (1-, 2-, or 5-shot) to rapidly adapt to new device variants, achieving high fidelity (>96%) to real devices and enabling large-scale testing of up to 1000 DTs concurrently with minimal adaptation time (~1 minute). The approach combines data generation from device OpenAPI schemas, lightweight meta-learning models (MAML), and scalable DT deployment with a DT request handler and device communication layer. Empirical evaluation on five real-world devices (three dispensers and two measurement devices) demonstrates strong fidelity, scalable performance, and practical time costs, and yields insights and guidelines for practitioners deploying MeDeT in industrial healthcare IoT test infrastructures.

Abstract

Testing healthcare Internet of Things (IoT) applications at system and integration levels necessitates integrating numerous medical devices of various types. Challenges of incorporating medical devices are: (i) their continuous evolution, making it infeasible to include all device variants, and (ii) rigorous testing at scale requires multiple devices and their variants, which is time-intensive, costly, and impractical. Our collaborator, Oslo City's health department, faced these challenges in developing automated test infrastructure, which our research aims to address. In this context, we propose a meta-learning-based approach (MeDeT) to generate digital twins (DTs) of medical devices and adapt DTs to evolving devices. We evaluate MeDeT in OsloCity's context using five widely-used medical devices integrated with a real-world healthcare IoT application. Our evaluation assesses MeDeT's ability to generate and adapt DTs across various devices and versions using different few-shot methods, the fidelity of these DTs, the scalability of operating 1000 DTs concurrently, and the associated time costs. Results show that MeDeT can generate DTs with over 96% fidelity, adapt DTs to different devices and newer versions with reduced time cost (around one minute), and operate 1000 DTs in a scalable manner while maintaining the fidelity level, thus serving in place of physical devices for testing.

Figures (6)

• Figure 1: Overview of an IoT-based healthcare system
• Figure 2: High-level overview of the approach, showing main components and their interactions. The detailed workflow is shown in \ref{['fig:app']}
• Figure 4: An example of meta-dataset and meta-taskset creation using preprocessed data. The data corresponding to class C1 is highlighted in green, while the data for class C2 is indicated in red.
• Figure 5: DT adaptations employed in our experiment: it demonstrates adaptations across devices (Karie, Medido, and Pilly) using their based version (v1), and depicts the progression of version adaptations for each device ranging from version v1 to v4.
• Figure 6: RQ3 results showing fidelity of 1000 DTs in different batch sizes