Human-Centered Explainable AI for TinyML Edge Devices: A Pareto-Based Selection Framework with LLM-Guided Design

XAI methods involve inherent trade-offs: The framework reveals that explanation quality, stability, and computational cost cannot all be maximized simultaneously on resource-constrained devices. Healthcare organizations must consciously choose which aspects matter most for their specific use case rather than assuming one solution fits all.

Stakeholder preferences should guide XAI selection: Rather than letting technical constraints alone determine which explanation method is used, the framework uses an LLM-guided interface to first capture what healthcare professionals and patients actually need to understand about AI predictions, then maps those preferences to appropriate technical solutions.

Edge AI deployment in healthcare requires careful planning: The paper addresses a real challenge in clinical settings—deploying AI models directly on local devices for immediate, private inference—but emphasizes this creates competing demands between explanation quality and the computational resources available on tiny devices, requiring deliberate trade-off decisions.

Current evaluation is incomplete for real-world deployment: The authors acknowledge their proof-of-concept evaluation covers only the computational selection stages, with "physical MCU deployment and empirical human-expert validation" remaining outside the scope, suggesting organizations should be cautious about implementing such frameworks without additional real-world testing with actual clinicians and patients.