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

to predict impact of otology manuscripts. Koroleva et al. cre ated an NLP algorithm with a 97.8% F-measure for detecting distorted research interpretations. Safranek et al. demonstrated ChatGPT's utility in automating ethical compliance checks, and Carnino et al. reported increased AI-generated text in 143 JAMA Otolaryngology papers after ChatGPT's release. 4 | Discussion To the author's knowledge, this study represents the most com prehensive review of NLP in the otolaryngology literature. Our findings highlight a surge of NLP in OHNS, with 94% of publi cations between 2020 and 2024 and the earliest in 1982 by Sagar et al. The majority of studies were US-based, focused on H&N oncology, and were in clinical settings. 4.1 | Patient Education and Monitoring NLP tools can enhance patient education by tailoring infor mation to factors like culture, language, and health literacy [179, 180]. With OHNS patients forgetting up to 80% of pro vider education, chatbots could provide ongoing support, but they must ensure accuracy, safety, and appropriate knowledge levels [181–183]. Our study found LLM response accuracy var ied, with some as low as 69%, underscoring their limitations in patient education [98]. While most errors were safe, up to 5% were potentially harmful, and most responses exceeded the NIH-recommended grade 6 reading level, even with prompt en gineering [34, 51, 76, 95, 98, 113, 126, 159, 160]. However, LLM responses improved significantly when prompted and demon strated utility in improving the readability of existing patient material [51, 76, 95, 98, 113, 126, 159, 160]. Patient interactions with chatbots can be utilized in an outpa tient setting during or post treatment to clarify concerns and questions and can reduce unnecessary hospital visits and detect more complications than physicians, as suggested by the find ings of Ma et al. [56]. Additionally, chatbots have the potential to detect complications earlier, provide more timely advice, and reduce in-person wait times, thereby improving outcomes. Accuracy, reliability, and user confidence must continue to be improved and validated with ongoing research in utilizing NLP for patient education and monitoring. More research is needed to compare LLMs for patient education. 4.2 | Electronic Medical Record Improvement Time-consuming documentation is a major source of physician burnout [184]. While NLP tools are widely studied as scribes in patient care, evidence of their effectiveness in otolaryngology re mains limited [185]. Using NLP for surgical charting is unique, requiring adherence to routine steps, procedure-specific lexicon, and intraoperative charting to enhance efficiency and accuracy. Our findings suggest further refinements are needed before implementation. Integrating NLP into EMR systems not only promises to alleviate the burden of documentation but also to enhance the accessibility of critical patient information. The few studies on this topic demonstrated mixed reliability of NLP for

this task. The literature also shows that NLP can aid in labeling results with impressive accuracy, useful for auditing reports and highlighting critical findings to prevent oversight. However, this application should complement, not replace, clinical judgment. 4.3 | Triaging and Patient Classification NLP has the potential to effectively and objectively triage pa tients, enabling centralized systems to efficiently match patients to physicians and resources, thereby reducing wait times. While promising, the results for both triaging and classifying patients' symptoms to the appropriate medical departments demon strated the need for significant improvement before use in a clinical setting. 4.4 | Clinician Education and Decision Support NLP tools utilizing LLMs have the potential to efficiently tai lor information and provide precise training-level adjusted answers in time-constrained clinical or educational settings. Overall, the accuracies of LLM treatment recommendations, diagnostic workups, or responses to OHNS questions were found to be varied between studies. This may be explained by inter-study methodological heterogeneity and the rapid evolu tion of LLM models. Additionally, the language used to interact with LLMs may dramatically change the reliability, as demon strated by Noda et al. 2024, where ChatGPT 4 had a 2-fold in crease in accuracy when prompted in English versus Japanese [141]. Limitations around language can have far-reaching im plications for resource-limited and non-English settings. Only a small subset of studies compared different LLMs in this domain, with the vast majority of studies utilizing ChatGPT [36, 47, 49, 52, 84, 103, 148, 158]. NLP tools may streamline the selection process for postgraduate training programs and enable more comprehensive review by predicting the most suitable applicants. Studies have shown that NLP can identify applicants suited for training programs, but further research is needed to validate its use and prevent bias, as noted by Halagur et al. [91, 134]. Moreover, while NLP has been assessed for clinical education and selection processes, studies specifically for teaching medical students and residents are cur rently limited. 4.5 | Data Extraction and Analysis for Research NLP can extract vast sets of data that were previously too labor-­ intensive by manual methods, enabling research previously unfeasible. Our results indicate that NLP tools demonstrate extraction accuracies comparable to humans over large data sets [38, 55, 87, 142, 145, 147, 155, 170, 175, 186, 187]. This is high lighted by Ali et al. [175], where 2,184,309 items of information were extracted from histopathology reports in 22.7 min, with an estimated accuracy of 99%. It is important to note that accuracies may vary by letter type, and it is prudent to utilize a verification set when automatically extracting information using NLP tools [188]. These results highlight the potential of integrating NLP into diverse databases, including EMRs, academic repositories,

3055