Explainable Human Activity Recognition: A Unified Review of Concepts and Mechanisms

A comprehensive review examines explainable AI methods for human activity recognition (HAR) systems across wearable, ambient, and physiological sensors. The paper addresses the critical gap between deep learning's performance improvements and the opacity that limits real-world deployment, proposing a unified framework for understanding XAI mechanisms in HAR applications.

This academic survey addresses a fundamental challenge in AI deployment: the tension between model performance and interpretability. Deep learning has significantly advanced human activity recognition across healthcare, smart homes, and assistive technologies, yet the black-box nature of these systems creates barriers to clinical adoption and user trust. The paper's contribution lies in establishing a unified taxonomy that separates conceptual dimensions of explainability from algorithmic mechanisms, providing clarity that prior research lacked.

The research emerges from growing regulatory and practical pressures for AI transparency. Healthcare systems, assistive devices, and smart environments increasingly require models that physicians, engineers, and users can understand and verify. This shift reflects broader industry recognition that accuracy alone is insufficient for mission-critical applications where failures could impact human safety or autonomy.

For developers and organizations deploying HAR systems, this review provides a structured framework for evaluating explainability trade-offs. The mechanism-centric taxonomy enables practitioners to match explanation approaches to their specific sensing modalities—whether wearable accelerometers, ambient microphones, or multimodal sensor fusion—and to specific stakeholder needs. Understanding how XAI methods handle HAR's temporal and semantic complexities is particularly relevant for healthcare IoT and eldercare applications.

Looking ahead, the gap between explainability research and production deployment remains substantial. Organizations implementing HAR should prioritize evaluation practices beyond accuracy metrics and establish clear interpretability objectives aligned with regulatory requirements and end-user needs. The review's emphasis on trustworthy systems that support human decision-making rather than mere automation suggests the field is maturing toward more responsible AI deployment practices.

• →Explainability has become critical for HAR deployment in healthcare and assistive living, bridging the gap between deep learning performance and real-world trust requirements.
• →The paper introduces a unified framework separating explainability concepts from algorithmic mechanisms, reducing ambiguity in prior XAI-HAR research.
• →XAI methods must address HAR's unique challenges including temporal dependencies, multimodal sensor fusion, and semantic complexity of human activities.
• →Current evaluation practices for XAI-HAR remain underdeveloped, creating deployment challenges for mission-critical healthcare and assistive applications.
• →Trustworthy activity recognition requires aligning interpretability objectives with specific stakeholder needs and regulatory requirements, not just accuracy optimization.