Knowledge Gaps and Methodology for Real-Time Deep Learning-Based Ergonomic Posture Evaluation

Work-related musculoskeletal disorders (MSDs) have become one of the more pressing concerns in occupational health, particularly among those whose jobs revolve around sustained computer use. While tools like RULA and REBA have served the ergonomics community well, their dependence on trained observers and periodic assessment windows means they routinely miss the gradual postural deterioration that accumulates during a typical workday. This paper charts a path toward addressing that limitation. Drawing on a structured review of the existing literature, five gaps are pinpointed the absence of systems that adapt in real time to individual users, the disconnect between AI-generated labels and clinically grounded risk categories, an underinvestment in temporal modelling of prolonged strain, a scarcity of datasets that reflect genuine office environments, and the persistent mismatch between model complexity and deployment-ready hardware. Against each gap, a corresponding methodological response is proposed: a pipeline built around YOLOv8n-Pose for keypoint detection, an LSTM or Temporal Transformer module for sequence analysis, Elastic Weight Consolidation for continual user adaptation, and an output scoring layer calibrated to RULA and REBA thresholds. Initial backbone testing returned 93% PCK@0.2 at 28 frames per second, which is taken as a feasibility signal.