xRead - Incorporating Artificial Intelligence into Clinical Practice (March 2026)

2.3 | Selection and Appraisal of Studies Search results were uploaded to the Covidence (Melbourne, Australia) platform, which automatically removed duplicates. Two reviewers (NB, BM) screened studies using Covidence, with disagreements resolved by AA (Figure 1). NB and BM then performed a full-text review. Discrepancies were settled through consensus by all three reviewers. 2.4 | Data Extraction Data were extracted in duplicate from the included studies by NB and BM. Discrepancies were settled by consensus. The ex traction template included: Study type, level of evidence, coun try, objective, AI used, application, subspecialty, sample size, methodology, and outcomes. 2.5 | Level of Evidence The 2011 Oxford Centre for Evidence-Based Medicine's levels of evidence were assigned to each study as an indicator of the potential bias in each study [12]. 3 | Results A total of 3224 studies were initially identified (Figure 1). After duplicate removal ( n = 2219) and abstract screening of 1005 studies, 274 studies were included in the full-text review. One hundred sixty-six studies were ultimately included [13–178]. Study characteristics are summarized in Table S1. The majority of studies were performance evaluations ( n = 118), case series ( n = 14), or cross-sectional ( n =9). The median publication year was 2024 (range 1982, 2024) (Figure 2). The majority of stud ies were from the USA ( n =71), China ( n = 16), Italy ( n = 12), or Germany ( n = 9) (Table 1). H&N oncology was the most common subspecialty ( n = 54), followed by rhinology ( n =15) and neuro tology ( n =9) (Table 2). Many studies did not focus on a subspe cialty of OHNS ( n = 78). One hundred twenty studies utilized NLP in a clinical context, while 46 applied it for research purposes. Figure 3 shows the breakdown of the specific applications. Table S2 contains a brief description of study methods and outcomes. No meta-analysis was performed in this study due to the hetero geneity of the methodologies of the included studies. 3.1 | Large Language Models Sixty-one percent (101/166) of studies included the use of ChatGPT, all of which were published in 2023 or 2024. ChatGPT-4 was used in 33% (55/166) of studies, ChatGPT-3.5 in 31% (52/166), ChatGPT-3 in 1% (2/166), and 4% (6/166) did not specify the ver sion. Twenty-five studies utilized other large language models [28, 36, 47, 49, 52, 60, 61, 82, 84, 88, 90, 91, 97, 107–109, 115, 118, 124, 144, 146, 148, 155, 158, 168]. BERT (Google Research, CA, USA)

training on massive datasets [4]. LLMs have recently been used to power NLP applications, leading to significant improvements. Different LLMs optimize NLP for different tasks. For example, Generative Pre-Trained Transformer (GPT) is optimized for gen erating coherent and relevant text, and Bidirectional Encoder Representations from Transformers (BERT) for understanding and interpreting language. The release of ChatGPT (OpenAI, November 2022) marked a significant NLP milestone, achieving the fastest growth in consumer app history by reaching 100 mil lion users in 2months [5]. NLP is still only emerging in health care and promises a wide range of future applications. Otolaryngologists must often balance patient-centered care with research and teaching [6, 7]. As clinicians, they must navigate triaging, ensuring timely diagnosis and treatment, patient education, documentation, monitoring, and follow-up. Scholarly activities include grant writing, data extraction and analysis, as well as manuscript preparation. As educators, oto laryngologists give lectures and provide student evaluations on top of clinical supervision and teaching. NLP technology emerges as a promising adjunct to enhance efficiency across these domains to aid otolaryngologists in fulfilling their varied responsibilities. Interest in the application of NLP to otolaryngology has surged since the introduction of large language models [8]. This study aims to review the current literature, shedding light on the exist ing applications of NLP within otolaryngology. 2 | Materials and Methods This study explores the application of NLP tools in otolaryngol ogy for enhancing traditional workflows. Reporting adhered to the Preferred Reporting Items for Systematic Reviews and Meta-analyzes Extension for Scoping Reviews (PRISMA-ScR) statement, and the methodology was published on Open Science Framework (DOI: 10.17605/OSF.IO/QR26G) [9]. 2.1 | Search Strategy The MEDLINE, EMBASE, SCOPUS, Cochrane Library, Web of Science, and CINAHL databases were searched on December 26th, 2023 using a search strategy peer-reviewed by a medical librarian under the Peer Review of Electronic Search Strategies criteria [10]. The medical subject headings “otolaryngology”, “natural language processing”, “otorhinolaryngology”, “neu rotology”, and “head and neck cancer” with relevant keywords were used (Supporting Information S1). 2.2 | Study Inclusion and Exclusion Criteria Original full-text articles utilizing NLP for otolaryngology or head and neck (H&N) cancer patient care and research, regard less of publication date, were included. Reviews on the topic of NLP in otolaryngology as well as basic science, text and opinion, and supplement articles were excluded. Irrelevant, non-English articles, and articles where the full text could not be found were also excluded.

3050

The Laryngoscope, 2025