Machine learning for non-technical skills assessment in emergency simulation: a scoping review and research agenda

Machine learning is increasingly applied to the assessment of non-technical skills in high-stakes simulation, yet the alignment between model selection and the construct being assessed remains underexamined. This study presents a scoping review of 114 peer-reviewed publications, conducted in accordance with the PRISMA extension for scoping reviews. Studies were identified through Scopus, IEEE Xplore and PubMed, screened manually in two stages without automated relevance filtering, and synthesised along five dimensions: targeted non-technical skills, machine learning model family, operational context, data modality, and reported methodological practice. Communication is the most frequently targeted skill, addressed by 58.8% of studies, followed by decision-making at 57.9% and teamwork at 48.2%. Leadership, at 15.8%, and coping with fatigue, at 5.3%, are the least addressed domains. Transformer, large language model and natural language processing methods constitute the largest model family, accounting for 34.1% of studies coded from full text. This represents a rapid transition: such methods appear in 3.6% of studies published before 2024 and in 47.6% of those published from 2024 onward, while non-transformer neural networks fall from 42.9% to 17.5% across the same boundary. The pairing of language-capable models with communication assessment is now the densest in the literature, and situational awareness is the one substantial skill domain the transition has not reached. Methodological practice has not kept pace with methodological capability: 62.6% of studies report no validation strategy, 65.9% do not address explainability, approximately half report no discriminative evidence of performance, and only 6.6% report operational deployment. The review advances a Model-Skill-Context Alignment Framework organised around construct modality, instrumentation capacity and model capability, and argues that skill coverage in this field follows the availability of computational capability rather than the operational importance of the construct.