Work-related musculoskeletal disorders remain common among office and information technology workers. The observational instruments used to assess them sample posture at intervals of weeks or months, while the exposure they aim to characterise is continuous. This paper specifies an architecture intended to narrow that sampling gap and sets out how it should be tested. The AI-Driven Ergonomic Digital Twin (AI-EDT) organises six layers: multimodal sensing, edge pre-processing, a physics-informed Vision Transformer for risk estimation, a two-tier federated protocol with differential privacy, a persistent per-worker twin state, and a reinforcement learning engine for workstation configuration. The contribution is formal rather than empirical. We give the twin state as a quintuple with an update law, derive a composite training objective in which a Newton-Euler residual penalises postures that are kinematically admissible but dynamically implausible, state the aggregation and privacy accounting relations, and formulate workstation adjustment as a continuous-action Markov decision process. Dimensional consistency is checked for each governing equation in Appendix A. Because no component has been validated, Section 10 specifies a five-phase protocol running from bench verification to a randomised intervention trial, with primary endpoints, sample size logic and pre-registered stopping criteria at every phase boundary. Section 11 states the analysis that will be applied once data exist, and identifies the measurements required before any performance claim can be made.

Keywords

Ergonomic risk assessment; digital twin; physics-informed neural networks; federated learning; differential privacy; musculoskeletal disorders