Toward Safer Integration of Active Back-Support Exoskeletons in Construction through Domain-Specific Language Models

Abstract

Active back-support exoskeletons have shown promise in reducing back-related musculoskeletal disorders in labor-intensive occupations. Yet, adoption in construction remains constrained by fragmented evidence on their biomechanical benefits, human factors risks, and practical usability. This study presents the development of a domain-adapted large language model fine-tuned with the Retrieval-Augmented Fine-Tuning framework to systematically synthesize and communicate evidence relevant to active back-support exoskeleton use in construction. The training dataset was developed from a review of twenty peer-reviewed studies spanning nine construction tasks and seven active back-support exoskeleton models, covering outcomes related to muscle activity, range of motion, perceived discomfort, exertion, usability, cognitive load, fall risk, and sociotechnical adoption factors. These studies identified both facilitators (e.g., productivity gains, posture correction, and device durability) and barriers (e.g., restricted mobility, thermal discomfort, and task–device incompatibility) shaping adoption in construction workflows. A total of 3,650 question–answer pairs were generated with distractors and Chain-of-Thought reasoning and used in a teacher–student distillation process with GPT-4o and GPT-4o-mini. The fine-tuned model achieved a validation accuracy of 88% and demonstrated stable generalization without overfitting, supported by low validation loss. In head-to-head evaluation against the baseline, the fine-tuned model achieved reliable scores in coherence, relevance, and harmlessness, a 10% improvement in response completeness (96% vs. 86%), and a 2% increase in factual accuracy (82% vs. 80%). The results demonstrate the feasibility of deploying fine-tuned large language models as interactive decision-support tools for exoskeleton adoption in construction, advancing the intersection of artificial intelligence, biomechanics, and occupational safety.