Skip to Content
MDPIMDPI
DiagnosticsDiagnostics
PDF
  • Article
  • Open Access

7 June 2025

Evaluation of the Performance of Large Language Models in the Management of Axial Spondyloarthropathy: Analysis of EULAR 2022 Recommendations

and
1
Department of Physical Medicine and Rehabilitation, Medipol University, Istanbul 34810, Turkey
2
Department of Physical Medicine and Rehabilitation, Baskent University Alanya Hospital, Antalya 07400, Turkey
*
Author to whom correspondence should be addressed.
Diagnostics2025, 15(12), 1455;https://doi.org/10.3390/diagnostics15121455
This article belongs to the Special Issue Artificial Intelligence Revolution in Chronic Inflammatory Diseases: Diagnosis, Prognosis, and Future Perspectives
Version Notes
Order Reprints
Review Reports

Abstract

Introduction: Guidelines have great importance in revealing complex and chronic conditions such as axial spondyloarthropathy. The aim of this study is to compare the answers given by various large language models to open-ended questions created from ASAS–EULAR 2022 guidance. Materials and Methods: This was a cross-sectional and comparative study. A total of 15 recommendations in the ASAS–EULAR 2022 guideline were derived directly from their content into open-ended questions in a clinical context. Each question was asked to the ChatGPT-3.5, GPT-4o, and Gemini 2.0 Flash models, and the answers were evaluated with a seven-point Likert system in terms of usability, reliability, Flesch–Kincaid Reading Ease (FKRE) and Flesch–Kincaid Grade Level (FKGL) metrics for readability, Universal Sentence Encoder (USE) and ROUGE-L for semantic and surface-level similarity. The results of different large language models were statistically compared, and p < 0.05 was revealed as statistically significant. Results: Better FKRE and FKGL scores were obtained in the Google Gemini 2.0 program (p < 0.05). Reliability and usefulness scores were significantly higher for ChatGPT-4o and Gemini 2.0 (p < 0.05). ChatGPT-4o yielded significantly higher semantic similarity scores compared to ChatGPT-3.5 (p < 0.05). There was a negative correlation between FKRE and FKGL scores and a positive correlation between reliability and usefulness scores (p < 0.05). Conclusions: It was determined that ChatGPT-4o and Gemini 2.0 programs gave more reliable, useful, and readable answers to open-ended questions derived from the ASAS–EULAR 2022 guidelines. These programs may potentially assist in supporting treatment decision-making in rheumatology in the future.

1. Introduction

Axial spondyloarthropathy (axial SpA) is a chronic rheumatic disease characterized by inflammation of the spine, sacroiliac joint, and peripheral joints. It is generally detected in young adults, more frequently in the male gender [1]. The most frequent initial symptoms are lower back pain, morning stiffness, and limitation in spinal movements. The diagnosis of axial spondyloarthritis is often challenging and may be delayed by several years, with reported delays of up to ten years in some cases [2]. Accurate diagnosis and effective management of axial SpA can improve patients’ quality of life and prevent future damage. Therefore, effective management of the disease has great importance [3].
Guidelines have great importance in managing complex and chronic conditions such as axial SpA. Guidelines that were developed by international rheumatology societies, such as the European League Against Rheumatism (EULAR) and Assessment of SpondyloArthritis international Society (ASAS), play a critical role in making decisions by both patients and healthcare professionals. The EULAR guidelines recommend appropriate treatment strategies for different disease activity levels and manifestations. This helps take control of disease activity and improve patients’ quality of life. EULAR provides recommendations for controlling disease activity, regular follow-ups, and pharmacological and non-pharmacological alternatives for disease management. This provides a standard for monitoring progress and preventing side effects [4,5,6].
Chat Generative Pre-trained Transformer (ChatGPT) and Gemini are the most common large language models (LLMs) that utilize human feedback and reinforcement learning. They have been preferred by millions of people because of their fast answers and easy access and become more popular day by day. There are previous studies evaluating the accuracy and reliability of large language models’ answers to frequently asked questions about public health [7,8]. Artificial intelligence has recently started to be used in rheumatology. Individuals, including both medical staff and patients, have started using it to explore potential conditions by inputting their symptoms. Laboratory data have also been added to request diagnostic suggestions. Advancing further, users have begun uploading radiological scans, prompting the systems to provide a ranked list of likely diagnoses [9,10,11,12]. In this study, we wanted to evaluate how adequately and accurately different large language models (such as ChatGPT-3.5, ChatGPT-4o Plus, and Gemini) provide answers to questions about important issues in the management of axial spondyloarthropathy in light of valid and up-to-date guidance. The aim of this study is to compare the responses of artificial-intelligence-based large language models (ChatGPT-3.5, ChatGPT-4o, and Gemini) in terms of usability, reliability, readability, and semantic and surface-level similarity through open-ended questions derived from the axial spondyloarthritis management recommendations published by the EULAR in 2022. In addition, this study aimed to qualitatively examine the extent to which these models offer accurate guidance to patients, alongside the comprehensiveness and accuracy of the information they deliver.

2. Materials and Methods

This study was designed as a cross-sectional and comparative study. Ethical approval was obtained from the Medipol University Ethics Committee (approval number: E-10840098-202.3.02-2713, date: 17 April 2025) and this study was conducted in accordance with the Declaration of Helsinki.

2.1. Open-Ended Question Development Process

A total of 15 recommendations from the ASAS–EULAR 2022 guidelines were translated into open-ended questions in a clinical context, derived directly from their content, like in Liu et al.’s study [13]. This process was conducted by two physical medicine and rehabilitation specialists with clinical experience in axial spondyloarthritis. Care was taken to preserve the original clinical intent of each recommendation during the formulation of the questions. Each item was designed to reflect the core message of its corresponding recommendation. The anticipated responses are based on the original content of the guideline.
The questions address a range of clinical themes, including treatment individualization, disease monitoring, patient education, pharmacological and non-pharmacological interventions, treatment response scenarios, and surgical options. Accordingly, the questions were grouped into four thematic categories, as in the study by Sciberras et al. [14]: (1) general management and monitoring strategies, (2) pharmacological treatment, (3) treatment response and adaptation strategies, and (4) surgery and accompanying conditions. This classification is presented in Table 1, and open-ended questions created from the ASAS–EULAR 2022 guidelines are given in Table 2.
Table 1. Classification of the ASAS–EULAR 2022 recommendations.
Table 2. Open-ended questions created from the ASAS–EULAR 2022 guidelines.

2.2. Large Language Models Used and Response Collection Process

In the study, large language models, ChatGPT-3.5 (Google LLC, Alphabet Inc., America, California), ChatGPT-4o (Google LLC, Alphabet Inc., America, California), and Gemini 2.0 Flash (Google LLC, Alphabet Inc., California), were used. These are large language models trained on extensive textual data, allowing them to generate responses that are contextually relevant and coherent, resembling those of a skilled human reader [15]. ChatGPT-3.5 and ChatGPT-4o (plus) are recent popular LLMs created by OpenAI, like Gemini was formed by Google [16]. Each open-ended question was submitted to the models in separate and freshly initiated sessions, with no prior context. The standardized prompt used across all models was as follows: “Please answer without considering the previous conversation.” The generated responses were recorded as text and anonymized for subsequent evaluation.

2.3. Evaluation Process of Responses

The answer given by each model to each question was independently evaluated by two experienced physiatrists (AU, OK), according to the reliability and usefulness criteria using the 7-point Likert scale developed by Uz et al. [17]. In cases where the evaluators disagreed, a consensus was first sought. If a consensus could not be reached, a third referee, who was also an experienced physiatrist specialized in the field (AA), evaluated the response. The 7-point Likert scale for the evaluation of reliability and usefulness is presented in Table 3. It has been known that LLMs like ChatGPT and Gemini are updated frequently; thus, reproducibility may be affected over time. To mitigate this issue, each question was presented to the models only once. The responses were printed in duplicate and distributed to the specialists, who evaluated the models independently, without knowledge of which response corresponded to which model.
Table 3. The 7-point Likert system for evaluation of reliability and usefulness of Chat Generative Pre-trained Transformer, courtesy of Cuma Uz [13].
In this study, the semantic similarity between model-generated responses and the ASAS–EULAR 2022 recommendations was evaluated using the Universal Sentence Encoder (USE). USE encodes texts into 512-dimensional vectors; this dimensionality reflects an empirically optimized balance between semantic representation and computational efficiency [18]. The transformer-based version was applied to both model outputs and reference texts, followed by cosine similarity calculation. Scores range from 0 to 1, with values closer to 1 indicating higher semantic similarity. All computations were performed in Python (v3.10) using the tensorflow_hub and scikit-learn libraries [19]. The surface-level similarity between model responses and the recommendations was assessed using the ROUGE-L metric, which measures the longest common subsequence (LCS) to capture word order and lexical overlap. ROUGE-L F1 scores were computed via the rouge-score package in Python; higher scores indicate greater lexical alignment between the texts [20,21].
Additionally, to evaluate the readability of the model-generated responses, we used the Flesch–Kincaid Reading Ease (FKRE) and Flesch–Kincaid Grade Level (FKGL) metrics. FKRE reflects ease of reading, with higher values indicating more readable text, while FKGL estimates the U.S. school grade level required to understand the text [22]. The FKRE score is calculated based on the average sentence length and the number of syllables per word, where higher scores reflect easier-to-read text. The FKGL score represents the U.S. school grade level required to comprehend the content [23]. These metrics were calculated using the textstat package in Python. (FKRE = 206.835 − 1.015 × (total words/total sentences) − 84.6 × (total syllables/total words) FKGL = 0.39 × (total words/total sentences) + 11.8 × (total syllables/total words) − 15.59 ROUGE-L F1 = [(1 + β2) × P × R]/[R + β2 × P], where P = LCS/m, R = LCS/n, and β = P/R)

2.4. Statistical Analysis

All statistical analyses were performed using IBM SPSS Statistics for Windows, Version 21.0 (IBM Corp., Armonk, NY, USA). The normality of continuous variables was assessed using the Shapiro–Wilk test. For variables exhibiting a normal distribution, the results were expressed as the mean and standard deviation; in contrast, for those not conforming to a normal distribution, the data were summarized using the median and range (minimum–maximum). Pairwise comparisons between the three large language models (ChatGPT-3.5, ChatGPT-4o, and Gemini 2.0) were performed using the Paired Samples’ t-test or the Wilcoxon Signed-Rank Test, depending on the normality of the data. Associations between numerical variables were examined using Spearman’s rank-order correlation. A p-value of <0.05 was considered statistically significant, which is a widely accepted threshold in biomedical research to control for Type I error while maintaining statistical sensitivity.

3. Results

Descriptive statistics for each large language model are presented in Table 4. In terms of FKGL and FKRE values, the Gemini 2.0 model produced text that was statistically more readable compared to the other models. For FKGL, both ChatGPT-4o [17.1 (15–20.3)] and Gemini 2.0 [15.7 (13.6–17.9)] had significantly lower scores compared to ChatGPT-3.5 [17.8 (14.6–19)]. (p < 0.001 and p = 0.003, respectively). Regarding FKRE, ChatGPT-4o [4.17 (–2.94–16.73)] and Gemini 2.0 [15.31 (1.13–27.01)] produced significantly higher scores than ChatGPT-3.5 [6.74 (–0.1–20.48)] (both p < 0.001). The reliability scores were significantly higher for ChatGPT-4o [6 (5–7)] and Gemini 2.0 [6 (4–7)] compared to ChatGPT-3.5 [4 (3–6)] (p < 0.001 and p = 0.001, respectively). Similarly, the usefulness scores were significantly higher for ChatGPT-4o [7 (5–7)] and Gemini 2.0 [7 (4–7)] than for ChatGPT-3.5 [5 (4–7)] (p = 0.002 and p < 0.001, respectively). Additionally, ChatGPT-4o [68.89 (53.64–79.57)] yielded significantly higher semantic similarity scores compared to ChatGPT-3.5 [66.19 (50.83–76.61)]. (p = 0.038). In terms of ROUGE-L F1 scores, Gemini 2.0 [13.3 (8.3–17.1)] yielded significantly lower values compared to ChatGPT-3.5 [13.9 (10.7–18.6)]. (p = 0.025), while no significant difference was observed between ChatGPT-4o [14.4 (10.8–21.2)] and Gemini 2.0 (p = 0.074). No statistically significant differences were observed between ChatGPT-4o and Gemini 2.0 across any of the evaluated metrics (all p > 0.05) (Table 5).
Table 4. Median and range (Min–Max) values of evaluation metrics for large language models.
Table 5. Pairwise comparisons of Flesch–Kincaid Grade Level, Flesch–Kincaid Reading Ease, Reliability Score, Usefulness Score, and USE-based Semantic Similarity of large language models across evaluation metrics (p-values).
Spearman’s correlation analyses for each model are presented in Table 6. A significant negative correlation was observed between FKGL and FKRE across all models: r = −0.730 (p = 0.002) for ChatGPT-3.5, r = −0.764 (p = 0.001) for ChatGPT-4o, and r = −0.838 (p < 0.001) for Gemini 2.0. Additionally, a significant positive correlation between reliability and usefulness was found in all three models: r = 0.870 (p < 0.001) for ChatGPT-3.5, r = 0.546 (p = 0.035) for ChatGPT-4o, and r = 0.714 (p = 0.003) for Gemini 2.0. No other statistically significant correlations were detected among the remaining variable pairs (all p > 0.05).
Table 6. Spearman’s correlation coefficients (r) and p-values between evaluation metrics for each large language model.

4. Discussion

In the current study, the Google Gemini 2.0 model demonstrated a significantly better FKRE score compared to ChatGPT-3.5 and ChatGPT-4o. Additionally, the model achieved significantly better FKGL scores than ChatGPT-3.5 and ChatGPT-4o, indicating enhanced readability. Furthermore, both reliability and usefulness ratings were significantly higher for ChatGPT-4o and Gemini 2.0 than for ChatGPT-3.5. ChatGPT-4o also exhibited greater semantic similarity compared to ChatGPT-3.5. No significant differences in semantic similarity were observed among the other models. A negative correlation was found between the FKRE and FKGL scores across all models. In addition, a positive correlation was observed between reliability and usefulness ratings in all three models, while no significant correlations were identified among the other parameters.
Ankylosing spondylitis is an autoimmune chronic inflammatory disease often seen in young men [24]. Environmental factors such as genetic and epigenetic factors, autoimmune antibodies, abnormal inflammation, stress, and gut dysbiosis are the most frequent causes of this disease. Mechanical stress, vitamin D deficiency, low estradiol levels, infections like Klebsiella pneumoniae, and oxidative stress have also been found to be among the risk factors [25,26,27,28]. In addition to spinal symptoms, the disease may also be accompanied by findings such as peripheral findings, psoriasis, and inflammatory bowel disease [29].
It has been determined that there are delays in diagnosing ankylosing spondylitis over time. This disease is grouped under the umbrella of spondylarthropathies, and the definition is divided into two: Axial spondyloarthropathy and non-radiographic spondyloarthropathy [30,31]. This distinction has contributed greatly to early diagnosis and initiation of treatment.
In addition to these developments, the diagnostic criteria and treatment guidelines have been constantly updated by the EULAR and ASAS [4,5,6]. This has allowed the patient and doctor to make decisions on treatment in light of scientific updates. The disease activity has been controlled easily, deformities have been prevented, and the patients’ quality of life has been improved significantly [3]. In addition to the renewal of diagnostic criteria and treatment guidelines, many new TNF-blocking agents have come into use in pharmacological treatment. Also, more emphasis has been placed on non-pharmacological treatments and rehabilitation methods, resulting in more success in treatment [32,33].
The developments are not limited to this. Artificial intelligence has been started to be used in the field of rheumatology [9]. Possible diagnoses began to be questioned by both patients and healthcare professionals by entering symptoms [10]. The test results were entered, and a diagnosis was requested. Furthermore, radiology images were uploaded to these programs, and these programs were asked to rank the possible diagnoses [10,11,12].
Artificial intelligence programs such as ChatGPT and Gemini were developed by Open AI and Google as large language models. These programs can find the appropriate answer by connecting to various databases when a question is asked of them, and moreover, when feedback is given to these answers, they can memorize it and produce a higher-level answer. This feature has made these programs more and more used day by day. New versions of the programs have also become available [34,35,36]. However, the increasing use of these programs has raised questions about how accurate, safe, and scientific the information they provide is [37]. In light of these questions, studies have been carried out on the use of artificial intelligence in rheumatology by healthcare professionals [38,39].
Nevertheless, model behavior is also influenced by technical parameters such as prompt formulation, temperature settings, and context window size. While these factors were standardized and held constant in our study to ensure methodological uniformity, their potential to impact the structure, accuracy, and creativity of model outputs has been demonstrated in prior research [40]. Future studies may benefit from systematically evaluating how such parameters affect clinical usefulness and information fidelity in medical domains.
Liu et al. examined 36 suggestions generated by ChatGPT and 29 suggestions made by humans for 7 different emergency situations. They underlined that ChatGPT had significant potential in learning from human feedback and giving the right advice in emergency situations [31]. Ren et al. evaluated the ChatGPT-4o and Kimi programs in a cross-sectional study involving 178 patients with ankylosing spondylitis and 4 rheumatologists. They prepared 15 questions related to AS from the literature and analyzed the programs’ answers and revealed these programs as understandable and useful [38]. Chen et al. performed a study on graduate students and checked ChatGPT’s usefulness, novelty, trust, and risk about medical and non-medical issues, and they determined that the program was found to be useful, but there was concern about its usage in clinical practice [37]. Kurshe et al. entered patients’ data into the ChatGPT-4o program, and they found that the preliminary diagnosis predictions of the ChatGPT-4o version were more accurate and cost-effective than the rheumatologists’ [41]. In the current study, the responses of ChatGPT-4o and Google Gemini were significantly more reliable and useful than those of ChatGPT-3.5.
In Coskun et al.’s study, twenty-three questions from an earlier study related to MTX were directed to ChatGPT-3.5 and ChatGPT-4. The accuracy of GPT-4 was found to be 100%, while GPT 3.5’s accuracy was revealed as 86.96%. It was concluded that the ChatGPT-4 program provided more accurate information about methotrexate use [42]. In a cross-sectional study conducted by Efe et al. on ChatGPT-4o and Google Gemini with 420 Board Vitals question bank questions, the accuracy rate of ChatGPT was found to be 86.9%, and that of Gemini was found to be 60.2%. When feedback was given, it was found that these rates increased to 86.7% and 60.5% [43]. Tong et al. asked ChatGPT-3.5, ChatGPT-4o, and Google Gemini 25 open-ended questions derived from the 2022 American College of Rheumatology Guideline prepared for glucocorticoid-induced osteoporosis. Google Gemini was more accurate than the ChatGPT programs, but when feedback was given, the ChatGPT programs were better than Gemini [12]. In the current study, ChatGPT-4o and Google Gemini programs’ answers given to open-ended questions created from the ASAS–EULAR 2022 guideline showed better reliability and usefulness results. In addition to high usefulness and reliability scores, better FKRE and FKGL scores showed that the ChatGPT-4o (p = 0.001) and Google Gemini 2.0 programs (p = 0.001) were significantly more readable than ChatGPT-3.5. However, in this study, no feedback was given to the programs, and statistical analysis was not performed on new answers. Although post-feedback analysis was not performed in this study, semantic and superficial similarity with ROUGE-L and USE were examined. ChatGPT-4o demonstrated significantly higher semantic similarity scores than ChatGPT-3.5. No significant semantic similarity was observed between the other programs.
Although these programs provide accurate information, educate patients, and suggest treatment options to clinicians, the biggest risk of these programs is undoubtedly the possibility of providing false information [30]. Moreover, inherent limitations of LLMs such as output variability, susceptibility to generating hallucinated or factually incorrect content, dependency on potentially outdated or biased training data, and a general lack of explainability should be carefully considered, especially in clinical contexts [44]. Ethical issues, misuse, responsibility of decisions, patient confidentiality and trust are among other reported risks [33]. During the evaluation, ChatGPT-3.5 failed to mention the ASDAS index alongside ASAS-HI in response to the second question and omitted the BASDAI and ASAS-HI thresholds required in the third question. This oversight contributed to a lower overall score. The primary reason for the significant point deduction was the failure to address the treat-to-target strategy. Furthermore, both ChatGPT-4o and Gemini lost points in the fourth question due to the omission of physiotherapy, a key component of management. Between the sixth and tenth questions, there were several critical errors in ChatGPT-3.5’s responses. Although it mentioned the use of opioids, it neglected to address the associated risk of addiction. It also recommended combining glucocorticoids with other medications, including conventional disease-modifying antirheumatic drugs (DMARDs) such as methotrexate, in the treatment approach and suggested their use in combination with biologic DMARDs, practices that are inconsistent with the current guidelines. Additionally, it recommended IL-23 inhibitors, despite their absence from the approved treatment algorithm. Although the combination of bDMARDs and DMARDs may be preferred in rheumatoid arthritis, such strategies are not supported for this disease, making these recommendations substantial inaccuracies and resulting in major point deductions. In response to the eleventh question, ChatGPT-3.5 did not consider the possibility of misdiagnosis, which also led to a considerable loss of points. Moreover, all three AI programs were found to provide inadequate guidance in the context of comorbid inflammatory bowel disease and uveitis. They incorrectly recommended switching to an IL-17 inhibitor following the failure of a first bDMARD, despite evidence supporting the use of an alternative, established bDMARD as a safer option. Additionally, none of the programs acknowledged the lack of sufficient evidence supporting dose reduction strategies for IL-17 inhibitors and Janus kinase inhibitors.
The strengths of the study are that it compared the latest versions of the programs and objective evaluations were made in terms of reliability and usability, as well as length of answers, understandability, and superficial and semantic similarity. The main limitation of this study is that no feedback was given when the programs responded, and a second analysis was not performed. In addition, this study was designed with only 15 open-ended questions created from the ASAS-EULAR 2022 guidelines. In the future, further studies may be designed with more questions, feedback may be given, LLM–user dialogue scenarios may be designed, and step-by-step diagnosis and symptom clarification may be supplied. Another important limitation is that usefulness and reliability could vary significantly depending on the evaluator’s clinical background. To minimize this issue, we performed this study with two specialists who had experience of more than 10 years in the specialty. Also, it has been known that LLMs like ChatGPT and Gemini are updated frequently; thus, reproducibility may be affected over time and this issue may affect the statistical results. Although the statistical differences were significant, their clinical implications should be interpreted with caution. For example, while higher reliability and readability scores may facilitate patient understanding and support clinical decision-making, the current findings should be validated in real-world clinical settings before drawing definitive conclusions. As has been known, the biggest risk of these programs is that they might give wrong answers. Although wrong answers were detailed qualitatively, new parameters that objectively show the risks are needed. Additionally, it should be kept in mind that LLMs like ChatGPT and Gemini are updated frequently; thus, reproducibility may be affected over time.

5. Conclusions

When open-ended questions derived from the EULAR 2022 recommendations were asked to the ChatGPT-3.5, ChatGPT-4o, and Gemini 2.0 programs, it was determined that ChatGPT-4o and Gemini 2.0 gave more reliable, useful, and readable answers. Although there is a risk of obtaining false information, these programs may be useful in patient education and clinician determination of a road map in the future. Future research may explore additional model iterations. Beyond merely adhering to standardized treatment guidelines, personalized therapeutic strategies may be formulated through the integration of patient-specific feedback and patient scenarios. In the longer term, LLMs could also be integrated into clinical decision support systems to assist healthcare professionals in interpreting guidelines and synthesizing complex data, provided that appropriate validation, oversight, and ethical safeguards are implemented. Further studies are needed.

Author Contributions

Study design (A.U.), collecting data (A.U., O.K.), statistical analysis (A.U., O.K.), detailed review (A.U.), writing the article (O.K., A.U.). All authors have read and agreed to the published version of the manuscript.

Funding

The researchers did not receive any funding for this study.

Institutional Review Board Statement

This study was approved by Medipol University Non-Interventional Clinical Research Ethics Committee (approval number: E-10840098-202.3.02-2713, date: 17 April 2025). All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki Declaration and its later amendments or comparable ethical standards.

Data Availability Statement

The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.

Acknowledgments

We are grateful to Ahmet Akyol for being a referee for disagreements of seven Likert points given to open-ended questions by the artificial intelligence programs in this study.

Conflicts of Interest

The authors declare that they have no conflicts of interest.

Abbreviations

axSpaAxial spondyloarthropathy
EULAREuropean League Against Rheumatism
ASASAssessment of SpondyloArthritis international Society
ChatGPTChat Generative Pre-trained Transformer
LLMLarge language models
LCSLongest common subsequence
USEUniversal Sentence Encoder
LCSLongest common subsequence
FKREFlesch–Kincaid Reading Ease
FKGLFlesch–Kincaid Grade Level
csDMARDsConventional disease-modifying antirheumatic drugs
bDMARDsBiologic disease-modifying antirheumatic drugs

References

  1. Smith, J.A. Update on ankylosing spondylitis: Current concepts in pathogenesis. Curr. Allergy Asthma Rep. 2015, 15, 489. [Google Scholar] [CrossRef] [PubMed]
  2. Walsh, J.A.; Magrey, M. Clinical Manifestations and Diagnosis of Axial Spondyloarthritis. J. Clin. Rheumatol. 2021, 27, e547–e560. [Google Scholar] [CrossRef]
  3. Fattorini, F.; Gentileschi, S.; Cigolini, C.; Terenzi, R.; Pata, A.P.; Esti, L.; Carli, L. Axial spondyloarthritis: One year in review 2023. Clin. Exp. Rheumatol. 2023, 41, 2142–2150. [Google Scholar] [CrossRef] [PubMed]
  4. Ramiro, S.; Nikiphorou, E.; Sepriano, A.; Ortolan, A.; Webers, C.; Baraliakos, X.; Landewé, R.B.M.; Van den Bosch, F.E.; Boteva, B.; Bremander, A.; et al. ASAS-EULAR recommendations for the management of axial spondyloarthritis: 2022 update. Ann. Rheum. Dis. 2023, 82, 19–34. [Google Scholar] [CrossRef]
  5. Ortolan, A.; Webers, C.; Sepriano, A.; Falzon, L.; Baraliakos, X.; Landewé, R.B.; Ramiro, S.; van der Heijde, D.; Nikiphorou, E. Efficacy and safety of non-pharmacological and non-biological interventions: A systematic literature review informing the 2022 update of the ASAS/EULAR recommendations for the management of axial spondyloarthritis. Ann. Rheum. Dis. 2023, 82, 142–152. [Google Scholar] [CrossRef]
  6. Webers, C.; Ortolan, A.; Sepriano, A.; Falzon, L.; Baraliakos, X.; Landewé, R.B.M.; Ramiro, S.; van der Heijde, D.; Nikiphorou, E. Efficacy and safety of biological DMARDs: A systematic literature review informing the 2022 update of the ASAS-EULAR recommendations for the management of axial spondyloarthritis. Ann. Rheum. Dis. 2023, 82, 130–141. [Google Scholar] [CrossRef] [PubMed]
  7. Singhal, K.; Azizi, S.; Tu, T.; Mahdavi, S.S.; Wei, J.; Chung, H.W.; Scales, N.; Tanwani, A.; Cole-Lewis, H.; Pfohl, S.; et al. Large language models encode clinical knowledge. Nature 2023, 620, 172–180. [Google Scholar] [CrossRef]
  8. Park, Y.J.; Pillai, A.; Deng, J.; Guo, E.; Gupta, M.; Paget, M.; Naugler, C. Assessing the research landscape and clinical utility of large language models: A scoping review. BMC Med. Inform. Decis. Mak. 2024, 24, 72. [Google Scholar] [CrossRef]
  9. Hügle, T. The wide range of opportunities for large language models such as ChatGPT in rheumatology. RMD Open 2023, 9, e003105. [Google Scholar] [CrossRef]
  10. Kelly, B.S.; Judge, C.; Bollard, S.M.; Clifford, S.M.; Healy, G.M.; Aziz, A.; Mathur, P.; Islam, S.; Yeom, K.W.; Lawlor, A.; et al. Radiology artificial intelligence: A systematic review and evaluation of methods (RAISE). Eur. Radiol. 2022, 32, 7998–8007, Erratum in Eur. Radiol. 2022, 32, 8054. https://doi.org/10.1007/s00330-022-08832-1. [Google Scholar] [CrossRef]
  11. Lee, S.; Jeon, U.; Lee, J.H.; Kang, S.; Kim, H.; Lee, J.; Chung, M.J.; Cha, H.S. Artificial intelligence for the detection of sacroiliitis on magnetic resonance imaging in patients with axial spondyloarthritis. Front. Immunol. 2023, 14, 1278247. [Google Scholar] [CrossRef] [PubMed]
  12. Chandwar, K.; Prasanna Misra, D. What does artificial intelligence mean in rheumatology? Arch. Rheumatol. 2024, 39, 1–9. [Google Scholar] [CrossRef] [PubMed]
  13. Liu, S.; Wright, A.P.; Patterson, B.L.; Wanderer, J.P.; Turer, R.W.; Nelson, S.D.; McCoy, A.B.; Sittig, D.F.; Wright, A. Using AI-generated suggestions from ChatGPT to optimize clinical decision support. J. Am. Med. Inform. Assoc. 2023, 30, 1237–1245. [Google Scholar] [CrossRef] [PubMed]
  14. Sciberras, M.; Farrugia, Y.; Gordon, H.; Furfaro, F.; Allocca, M.; Torres, J.; Arebi, N.; Fiorino, G.; Iacucci, M.; Verstockt, B.; et al. Accuracy of Information given by ChatGPT for Patients with Inflammatory Bowel Disease in Relation to ECCO Guidelines. J. Crohns. Colitis. 2024, 18, 1215–1221. [Google Scholar] [CrossRef] [PubMed]
  15. Alhur, A. Redefining Healthcare With Artificial Intelligence (AI): The Contributions of ChatGPT, Gemini, and Co-pilot. Cureus 2024, 16, e57795. [Google Scholar] [CrossRef]
  16. Tong, L.; Zhang, C.; Liu, R.; Yang, J.; Sun, Z. Comparative performance analysis of large language models: ChatGPT-3.5, ChatGPT-4 and Google Gemini in glucocorticoid-induced osteoporosis. J. Orthop. Surg. Res. 2024, 19, 574. [Google Scholar] [CrossRef]
  17. Uz, C.; Umay, E. “Dr ChatGPT”: Is it a reliable and useful source for common rheumatic diseases? Int. J. Rheum. Dis. 2023, 26, 1343–1349. [Google Scholar] [CrossRef]
  18. Fattoh, I.E.; Kamal Alsheref, F.; Ead, W.M.; Youssef, A.M. Semantic Sentiment Classification for COVID-19 Tweets Using Universal Sentence Encoder. Comput. Intell. Neurosci. 2022, 2022, 6354543. [Google Scholar] [CrossRef]
  19. Izzidien, A.; Fitz, S.; Romero, P.; Loe, B.S.; Stillwell, D. Developing a sentence level fairness metric using word embeddings. Int J. Digit Humanit. 2022, 1–36. [Google Scholar] [CrossRef]
  20. Tang, L.; Sun, Z.; Idnay, B.; Nestor, J.G.; Soroush, A.; Elias, P.A.; Xu, Z.; Ding, Y.; Durrett, G.; Rousseau, J.F.; et al. Evaluating large language models on medical evidence summarization. NPJ Digit. Med. 2023, 6, 158. [Google Scholar] [CrossRef]
  21. Gulden, C.; Kirchner, M.; Schüttler, C.; Hinderer, M.; Kampf, M.; Prokosch, H.U.; Toddenroth, D. Extractive summarization of clinical trial descriptions. Int. J. Med. Inform. 2019, 129, 114–121. [Google Scholar] [CrossRef] [PubMed]
  22. Kılıcoglu, V.S.; Yurdakul, O.V.; Aydın, T. Evaluation of Chat Generative Pre-trained Transformer’s responses to frequently asked questions about psoriatic arthritis: A study on quality and readability. Ann. Med. Res. 2025, 32, 79–84. [Google Scholar]
  23. Solnyshkina, M.; Zamaletdinov, R.; Gorodetskaya, L.A.; Azad, I. Evaluating text complexity and Flesch-Kincaid grade level. J. Soc. Stud. Educ. Res. 2017, 8, 238–248. [Google Scholar]
  24. Mauro, D.; Thomas, R.; Guggino, G.; Lories, R.; Brown, M.A.; Ciccia, F. Ankylosing spondylitis: An autoimmune or autoinflammatory disease? Nat. Rev. Rheumatol. 2021, 387–404. [Google Scholar] [CrossRef] [PubMed]
  25. Păsăran, E.D.; Diaconu, A.E.; Oancea, C.; Bălănescu, A.R.; Aurelian, S.M.; Homentcovschi, C. An Actual Insight into the Pathogenic Pathways of Ankylosing Spondylitis. Curr. Issues Mol. Biol. 2024, 46, 12800–12812. [Google Scholar] [CrossRef]
  26. Hwang, M.C.; Ridley, L.; Reveille, J.D. Ankylosing spondylitis risk factors: A systematic literature review. Clin. Rheumatol. 2021, 40, 3079–3093. [Google Scholar] [CrossRef]
  27. Bilski, R.; Kamiński, P.; Kupczyk, D.; Jeka, S.; Baszyński, J.; Tkaczenko, H.; Kurhaluk, N. Environmental and Genetic Determinants of Ankylosing Spondylitis. Int. J. Mol. Sci. 2024, 25, 7814. [Google Scholar] [CrossRef] [PubMed]
  28. Yang, H.; Chen, Y.; Xu, W.; Shao, M.; Deng, J.; Xu, S.; Gao, X.; Guan, S.; Wang, J.; Xu, S.; et al. Epigenetics of ankylosing spondylitis: Recent developments. Int. J. Rheum. Dis. 2021, 24, 487–493. [Google Scholar] [CrossRef]
  29. Wei, Y.; Zhang, S.; Shao, F.; Sun, Y. Ankylosing spondylitis: From pathogenesis to therapy. Int. Immunopharmacol. 2025, 145, 113709. [Google Scholar] [CrossRef]
  30. Navarro-Compán, V.; Benavent, D.; Capelusnik, D.; van der Heijde, D.; Landewé, R.B.; Poddubnyy, D.; van Tubergen, A.; Baraliakos, X.; Van den Bosch, F.E.; van Gaalen, F.A.; et al. ASAS consensus definition of early axial spondyloarthritis. Ann. Rheum. Dis. 2024, 83, 1093–1099. [Google Scholar] [CrossRef]
  31. van der Heijde, D.; Molto, A.; Ramiro, S.; Braun, J.; Dougados, M.; van Gaalen, F.A.; Gensler, L.S.; Inman, R.D.; Landewé, R.B.M.; Marzo-Ortega, H.; et al. Goodbye to the term ‘ankylosing spondylitis’, hello ‘axial spondyloarthritis’: Time to embrace the ASAS-defined nomenclature. Ann. Rheum. Dis. 2024, 83, 547–549. [Google Scholar] [CrossRef] [PubMed]
  32. Zimba, O.; Kocyigit, B.F.; Korkosz, M. Diagnosis, monitoring, and management of axial spondyloarthritis. Rheumatol. Int. 2024, 44, 1395–1407. [Google Scholar] [CrossRef] [PubMed]
  33. Zouris, G.; Evangelopoulos, D.S.; Benetos, I.S.; Vlamis, J. The Use of TNF-α Inhibitors in Active Ankylosing Spondylitis Treatment. Cureus 2024, 16, e61500. [Google Scholar] [CrossRef]
  34. Fatani, B. ChatGPT for Future Medical and Dental Research. Cureus 2023, 15, e37285. [Google Scholar] [CrossRef]
  35. Mihalache, A.; Huang, R.S.; Popovic, M.M.; Muni, R.H. ChatGPT-4: An assessment of an upgraded artificial intelligence chatbot in the United States Medical Licensing Examination. Med. Teach. 2024, 46, 366–372. [Google Scholar] [CrossRef]
  36. Waisberg, E.; Ong, J.; Masalkhi, M.; Zaman, N.; Sarker, P.; Lee, A.G.; Tavakkoli, A. Google’s AI chatbot “Bard”: A side-by-side comparison with ChatGPT and its utilization in ophthalmology. Eye 2024, 38, 642–645. [Google Scholar] [CrossRef]
  37. Chen, S.Y.; Kuo, H.Y.; Chang, S.H. Perceptions of ChatGPT in healthcare: Usefulness, trust, and risk. Front. Public Health. 2024, 12, 1457131. [Google Scholar] [CrossRef]
  38. Ren, Y.; Kang, Y.N.; Cao, S.Y.; Meng, F.; Zhang, J.; Liao, R.; Li, X.; Chen, Y.; Wen, Y.; Wu, J.; et al. Evaluating the performance of large language models in health education for patients with ankylosing spondylitis/spondyloarthritis: A cross-sectional, single-blind study in China. BMJ Open 2025, 15, e097528. [Google Scholar] [CrossRef] [PubMed]
  39. Daungsupawong, H.; Wiwanitkit, V. Comparative performance of artificial intelligence models in rheumatology board-level questions: Evaluating Google Gemini and ChatGPT-4o: Correspondence. Clin. Rheumatol. 2024, 43, 4015–4016. [Google Scholar] [CrossRef]
  40. Currie, G.M. Academic integrity and artificial intelligence: Is ChatGPT hype, hero or heresy? Semin. Nucl. Med. 2023, 53, 719–730. [Google Scholar] [CrossRef]
  41. Krusche, M.; Callhoff, J.; Knitza, J.; Ruffer, N. Diagnostic accuracy of a large language model in rheumatology: Comparison of physician and ChatGPT-4. Rheumatol. Int. 2024, 44, 303–306. [Google Scholar] [CrossRef] [PubMed]
  42. Coskun, B.N.; Yagiz, B.; Ocakoglu, G.; Dalkilic, E.; Pehlivan, Y. Assessing the accuracy and completeness of artificial intelligence language models in providing information on methotrexate use. Rheumatol. Int. 2024, 44, 509–515. [Google Scholar] [CrossRef] [PubMed]
  43. Is, E.E.; Menekseoglu, A.K. Comparative performance of artificial intelligence models in rheumatology board-level questions: Evaluating Google Gemini and ChatGPT-4o. Clin. Rheumatol. 2024, 43, 3507–3513. [Google Scholar] [CrossRef] [PubMed]
  44. Azamfirei, R.; Kudchadkar, S.R.; Fackler, J. Large language models and the perils of their hallucinations. Crit. Care 2023, 27, 120. [Google Scholar] [CrossRef]
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Article Metrics

Citations

Article Access Statistics

Journal Statistics
Multiple requests from the same IP address are counted as one view.
Editorial for “Advances in Oral Diseases Diagnosis and Management: 2nd Edition”
Previous
Atypical Manifestations of Cowden Syndrome in Pediatric Patients
Next