Applications of transformer-based pretrained language models for analyzing nursing records in adult care: A scoping...

Nursing records contain clinically meaningful information but remain underutilized due to their unstructured narrative format. Transformer-based pretrained language models, including encoder-based pretrained language models (PLMs) and decoder-based generative large language models (LLMs), are increasingly being used to analyze these records, yet the landscape of their applications remains unclear.

To map the application of these models to nursing records in adult care, including the records analyzed, analytical tasks and clinical purposes, methodological approaches, model performance, and reported limitations.

This scoping review followed Joanna Briggs Institute methodology and PRISMA-ScR guidelines. PubMed, Embase, CINAHL, and Scopus were searched from inception through May 2026 for studies applying these models to unstructured nursing records in adult care. Two reviewers independently screened titles/abstracts and assessed full texts. Data were extracted by one reviewer and independently verified by two reviewers. Findings were synthesized descriptively.

Thirty-two studies published between 2021 and 2026 were included. Encoder-based PLMs (n = 19) were adapted mainly by supervised fine-tuning, whereas decoder-based generative LLMs (n = 13) were used mainly through prompting. Both architectures were applied primarily to classification and information extraction, most often to identify patient conditions or predict clinical outcomes. Summarization was conducted in two studies. Model performance was not comparable across studies because of heterogeneous evaluation conditions and limited reporting of uncertainty. Only one study reported prospective silent evaluation; none evaluated clinical deployment or effects on clinical outcomes. Generalizability was the most frequently reported limitation (n = 14).

Transformer-based pretrained language models have been applied to nursing records mainly to transform narrative text into predefined categories, with little use of approaches that leverage the narrative characteristics of these records. Validation has largely remained confined to model development settings, limiting conclusions about clinical applicability. Future research should prioritize external and prospective validation and investigate tasks and adaptation strategies suited to nursing records.