ChatGPT and reference intervals: a comparative analysis of repeatability in GPT-3.5 Turbo, GPT-4, and GPT-4o – Frontiers

Your new experience awaits. Try the new design now and help us make it even better
ORIGINAL RESEARCH article
Front. Artif. Intell., 12 December 2025
Sec. Medicine and Public Health
Volume 8 – 2025 | https://doi.org/10.3389/frai.2025.1681979
Background: Large language models such as ChatGPT hold promise as rapid “curbside consultation” tools in laboratory medicine. However, their ability to generate consistent and clinically reliable reference intervals—particularly in the absence of contextual clinical information—remains uncertain.
Method: This cross-sectional study evaluated whether three versions of ChatGPT (GPT-3.5-Turbo, GPT-4, GPT-4o) maintain repeatable reference-interval outputs when the prompt intentionally omits the interval, using reference interval variability as a stress-test for model consistency. Standardized prompts were submitted through 726,000 chatbot requests. A total of 246,842 reference intervals across 47 laboratory parameters were then analyzed for consistency using the coefficient of variation (CV) and regression models.
Results: On average, the chatbots exhibited a CV of 26.50% (IQR: 7.35–129.01%) for the lower limit and 15.82% (IQR: 4.50–45.30%) for the upper limit upon repetition. GPT-4 and GPT-4o demonstrated significantly lower CVs compared to GPT-3.5-Turbo. Reference intervals for poorly standardized parameters were particularly inconsistent across lower (β: 0.6; 95% CI: 0.35 to 0.86; p < 0.001) and upper limit (β: 0.5; 95% CI: 0.28 to 0.71; p < 0.001), while unit expressions also showed variability.
Conclusion: While the newer ChatGPT versions tested demonstrate improved repeatability, diagnostically unacceptable variability persists, particularly for poorly standardized analytes. Mitigating this requires thoughtful prompt design (e.g., mandatory inclusion of reference intervals), global harmonization of laboratory standards, further model refinement, and robust regulatory oversight. Until then, AI chatbots should be restricted to professional use and trained to refuse laboratory interpretation when reference intervals are not provided by the user.
The release of Chat Generative Pre-Trained Transformer (ChatGPT) in November 2022 has considerably altered the landscape of AI-based chatbots, as demonstrated by its remarkable user adoption rates (Hu, 2023). These AI-driven platforms provide rapid, interactive, and accessible responses, attracting considerable interest within laboratory medicine research for their potential to serve as “curbside consultations” in medical settings (Yang et al., 2023; Girton et al., 2024; Cadamuro et al., 2023). Concurrently, users on social media platforms advocate for employing ChatGPT to interpret personal laboratory results (Girton et al., 2024; O’Connor, 2023), mirroring a broader trend of self-diagnosis using ChatGPT (Shahsavar and Choudhury, 2023).
Despite ChatGPT’s remarkable performance in medical licensing exams (Meyer et al., 2024a; Liu et al., 2024), its reliability and applicability for addressing such post-analytical queries from laypersons remain debated (Girton et al., 2024; Meyer et al., 2024b). A noteworthy concern is the phenomenon of hallucinations, whereby the chatbot generates plausible yet erroneous information (Yang et al., 2023), while large-scale studies on the repeatability (use of the same prompts repeatedly) remain scarce. Specifically, when addressing laboratory medical questions, inconsistencies in reference intervals often lead to post-analytical errors and an overestimation tendency of 31% by ChatGPT—potentially further burdening patients and the healthcare system (Meyer et al., 2024b).
After all, reference intervals themselves are foundational to interpreting laboratory data (Coskun and Lippi, 2024) with recent years witnessing efforts to refine their estimation (Haeckel et al., 2017) and foster harmonization (Bohn et al., 2023). Although standardization is needed for the transferability of laboratory data (Guidi et al., 2006; Plebani and Lippi, 2023), the precise impact of existing initiatives such as by the American Association for Clinical Chemistry and the International Federation of Clinical Chemistry and Laboratory Medicine remains uncertain (Plebani, 2013).
This situation presents a critical dichotomy: reference interval standardization remains incomplete (Plebani, 2013) yet patients inconsistently provide reference intervals when seeking online interpretations (Meyer et al., 2024b). Understanding how ChatGPT responds to such variability is therefore crucial.
Thus, this study aimed to determine whether GPT-3.5-Turbo, GPT-4, and GPT-4o provide internally consistent reference intervals within the context of laboratory medicine. Treating reference interval variability itself as the stress-test, it could provide insight into the potential utility of prompt engineering, through the mandatory inclusion of reference values, for post-analytical questions, thereby responding to the call in the literature for large-scale analysis into this matter (Girton et al., 2024).
Building upon previous research on AI-based chatbots in laboratory medicine using both simulated (Cadamuro et al., 2023) and real patient cases (Girton et al., 2024; Meyer et al., 2024b), ChatGPT was selected as the subject of this study. Due to the cap of 50,000 requests per batch, we decided to submit 100 prompts per laboratory parameter per session for both male and female subjects, generating a batch of 24,200 per session. We conducted these sessions twice daily over five days across three different versions of ChatGPT (GPT-3.5-Turbo, GPT-4, GPT-4o), resulting in a total of 726,000 requests, with 242,000 per version. To submit the prompts, we utilized the ChatGPT API provided through OpenAI’s Researcher Access Program and API. Access to the API was achieved via the interfaces made available on the ChatGPT website.
To ensure methodological consistency, we programmatically generated uniform prompts for a hypothetical 30-year-old patient (height 170 cm, weight 70 kg) of female and male gender. Each prompt followed the exact template: “Please give me the lower and upper reference values for [parameter] in [medium] for a 30-year-old [gender] patient with a height of 170 cm and a weight of 70 kg. The output should only include the lower and upper reference limit with the unit in the format ‘lower_reference_limit;upper_reference_limit;unit’”. By rigidly specifying both the clinical context and the required output schema (lower_reference_limit;upper_reference_limit;unit), and by periodizing requests across twice-daily sessions over five consecutive days, we aimed to minimize natural-language variability while assessing temporal stability and potential drift. We intentionally used a standardized prompt with a single, hypothetical patient profile to isolate the intrinsic repeatability of model outputs. Thus, the aim of this study design was to prioritize internal validity over external generalizability.
All parameters outlined in the table for “Permissible Measurement Uncertainty” from the German Society for Clinical Chemistry and Laboratory Medicine were initially included [Deutsche Gesellschaft für Klinische Chemie und Laboratoriumsmedizin (DGKL), n.d.], in order to encompass a broad array of high-volume assays—clinical-chemistry parameters (electrolytes, enzymes), hematology indices (hemoglobin, cell counts), immunology markers (immunoglobulins), therapeutic-drug concentrations, and tumor markers.
Excluded from analysis were any ChatGPT responses that failed to adhere to the output format, those in which the reported lower reference limit exceeded the upper reference limit, and any parameter for which fewer than 80% of the 100 repeated queries per analyte used a consistent unit notation. This approach aimed to ensure data consistency and processability, as well as to avoid distortions due to unit conversions, given that ChatGPT is based on a Large Language Model (LLM) that inherently struggles with mathematics (Frieder et al., 2024). Moreover, to ensure data quality, we conducted a random manual review of extreme values within the predefined upper and lower limits for each parameter. This allowed us to select 246,842 chatbot outputs from the 726,000 requests.
Each laboratory parameter was classified based on its harmonization/standardization status (“none/needed”, “active/incomplete” or “adequate/maintained”), as defined by the International Consortium for Harmonization of Clinical Laboratory Results (2025) (Appendix 1). This classification enabled stratification of model performance by the degree of harmonization/standardization inherent to each analyte.
Statistical analysis was performed using the programming language R (Team RC, 2022), mainly utilizing the packages “tidyverse” (Wickham et al., 2019), “rio” (Chan et al., 2023), “gtsummary” (Sjoberg et al., 2022), and “viridis” (Garnier et al., 2024). After computing the coefficient of variation (CV) for each variable, the assumption of normality was refuted using both the Kolmogorov–Smirnov and Shapiro–Wilk tests. Thus, non-parametric methods were employed for subsequent analyses, with paired data being compared using the Wilcoxon signed-rank test and the Friedman test, and unpaired data being assessed via the Kruskal–Wallis H test. For graphical presentation and regression modeling, CV values were log₁₀-transformed. Finally, we summarized the distribution of CV and explored its determinants by fitting a linear regression model with log₁₀(CV) as the dependent variable.
Out of 726,000 outputs, a match of less than 80% was found between the units for 75 parameters, accumulated for 410,645 queries. Inconsistencies were noted not only in the units themselves but also their notations. For instance, both “liter” (American English) and “litre” (British English) were used interchangeably for the same analyte. Non-standard units, such as “pictogram” instead of “picogram,” and spelling errors like “millilittre” were also observed.
Beyond these issues, some reference intervals were affected by unnecessary weight-adjusted calculations, further contributing to deviations from the reference interval.
Moreover, 8% (58,271/726,000) of ChatGPT’s outputs were not in the desired format. Of these, 76% (44,485/58,271) originated from the GPT-3.5-Turbo model, 20% (11,581/58,271) from GPT-4, and 3.8% (2,205/58,271) from GPT-4o. In total, 58% (33,903/58,271) of these format issues were associated with requests for female reference values. For eight variables (angiotensin-converting enzyme, carbohydrate-deficient transferrin, cholinesterase, immunoglobulin M, immunoglobulin G, methotrexate, prostate-specific antigen, and reticulocytes) more than 2% of the ChatGPT outputs deviated from the desired format. The chatbot primarily attributed these discrepancies to the lack of standardization in reference intervals, the irrelevance of prostate-specific antigen in women, and the absence of clinical context (Table 1).
Table 1. Format, unit and extreme values of reference ranges provided by ChatGPT with regard to laboratory medicine, 2024.
The upper and lower limits also varied widely. The median CV for the analytes was 26.50% (IQR: 7.35–129.01%) at the lower limit and 15.82% (IQR: 4.50–45.30%) at the upper limit with no significant gender differences at either boundary (lower limit p = 0.862; upper limit p = 0.542). GPT-3.5-Turbo generated markedly broader dispersion, with median CVs more than twice those observed for GPT-4 and nearly four times those for GPT-4o at the upper limit (all model comparisons p < 0.001).
Analytical harmonization had a strong effect on variability. Analytes classified as “adequate/maintained” exhibited the lowest CVs, whereas those classified as ‘None/Needed’ showed CVs that were three times higher at the lower limit and over twice as high at the upper limit (p < 0.001; Figure 1; Appendix 2). Notably, measurands standardized to a limited extent like tumor markers such as free PSA or Ca 19–9 nearly consistently displayed such elevated CVs in contrary to well standardized analytes like pH (Figures 1, 2; Appendix 3).
Figure 1. Coefficient of variation (%) for upper and lower limit supplied by GPT-3.5-Turbo (blue) GPT-4 (green), GPT-4o (light green) in regard to harmonization/standardization status, 2024. Panels present results for (A) the combined-sex cohort, (B) females only, and (C) males only.
Figure 2. Coefficient of Variation (%) on a Log₁₀ Scale for Upper and Lower Limit of the measurands supplied by GPT-3.5-Turbo (blue) GPT-4 (green), GPT-4o (light green), 2024.
Univariate regression confirmed these patterns. Compared with GPT-3.5-Turbo, both GPT-4 and GPT-4o were independently associated with significantly lower CVs for the lower limit, upper limit, and reference range (all p < 0.001). Gender again showed no predictive value, whereas analytes lacking harmonization remained strong positive predictors of high variability (all p < 0.001) (Figure 3; Appendix 4).
Figure 3. Univariate regression of the coefficient of variation for ChatGPT-derived laboratory reference intervals (β, log₁₀ scale; 95% CI). Marker shapes denote harmonization status (●), gender (▲), and model version (■), while lower limit is colored in blue, reference range in green, and upper limit in light green.
Despite their promising performance in medical licensing and certification exams globally (Meyer et al., 2024a; Liu et al., 2024), current AI-based chatbots, such as ChatGPT, demonstrate notable limitations in addressing laboratory medicine queries under real-world conditions (Meyer et al., 2024b). This issue is particularly pressing given the rising prevalence of self-diagnosis among ChatGPT users (Shahsavar and Choudhury, 2023). In such scenarios, patients often seek interpretations of laboratory results online, yet commonly neglect to provide the corresponding reference intervals (Meyer et al., 2024b)—a critical omission that compromises post-analytical accuracy (Sikaris, 2015).
Our findings indicate, while iterative refinements in LLMs have led to notable improvements in repeatability and internal consistency – particularly in GPT-4o and GPT-4 compared to GPT-3.5-Turbo – relevant challenges remain. Although this trend aligns with previous observations of substantial performance gains achieved by GPT-4 over its predecessor in medical exams (Meyer et al., 2024a; Ibrahim et al., 2023; Yang et al., 2024), it also underscores ongoing concerns regarding repeatability, as reported in other clinical contexts (Franc et al., 2024; Franc et al., 2024).
Such persistent difficulties with repeatability, especially regarding reference intervals, are concerning, potentially leading to misinterpretation of results if patients or clinicians inadvertently rely on LLM-derived reference intervals. Variability of this magnitude exceeds typical allowable error thresholds used in clinical laboratory quality assurance, highlighting a fundamental misalignment between the probabilistic nature of LLM outputs and the precision required in laboratory medicine. Indeed, previous work has shown that incomplete prompts – those lacking reference intervals – can exacerbate post-analytical errors in LLM outputs (Meyer et al., 2024b). Ensuring that reference intervals are consistently included in prompts appears essential to mitigate these inaccuracies and reduce post-analytical errors (Nguyen et al., 2024; Wang et al., 2024).
Of particular concern are analytes that lack adequate or maintained harmonization. For these tests the chatbots consistently produced high coefficients of variation, plausibly because the models have comparatively little high-quality training data for such parameters. Therefore, analytes standardized to a limited extent not only confuse patients but also challenge LLMs (Jones and Barker, 2007). These findings underscore the urgency of international efforts to establish reference materials [International Federation of Clinical Chemistry and Laboratory Medicine (IFCC), 2024a] and to harmonize reference intervals, such as those led by the International Federation of Clinical Chemistry and Laboratory Medicine (Gillery and Young, 2013) or by working groups from Canada and Australia (Bohn et al., 2023; Koerbin et al., 2018). It is a well-known fact that the clinical interpretation of reference intervals and limits asks for harmonization (Klee, 2004) or if possible, for standardization. Thus, it is evident, that global standardization and harmonization is not merely a means of reducing data complexity (Vesper et al., 2016), thereby preventing comprehension problems and interpretation errors (Jones and Barker, 2007), but is also essential for enhancing the quality of AI-assisted tools in clinical laboratory medicine.
Moreover, the observed substantial variability in units, their notations and their spelling suggests that data heterogeneity extends beyond reference intervals to the units themselves. Notably, the analytes most affected by unit inconsistencies may also correspond to those for which LLMs are least reliable. Excluding such outputs to preserve analytic rigor may therefore have even led to an underestimation of real-world variability. These errors reflect the underlying heterogeneity within the LLM’s training corpus, where language variations and non-standard data introduce noise. Such inconsistencies exacerbate interpretive challenges, particularly for laypersons, and underscore the limitations of LLMs as “stochastic parrots”—systems driven by statistical correlations rather than causal understanding (Bender et al., 2021). When combined with the variability in reference intervals, these inconsistencies exacerbate interpretive challenges. Addressing this issue requires adherence to standardized units and notations, as recommended by the European Federation of Clinical Chemistry and Laboratory Medicine (Hansen, 2019) and underline the importance of harmonization the total testing process (Plebani, 2013; Plebani, 2016; Plebani, 2018). Implementing these recommendations within LLM training frameworks could thus further foster output consistency and usability in clinical contexts.
Nevertheless, the nature of this study poses several limitations. This study focused on three versions of ChatGPT from a single provider (OpenAI). This scope reflects pragmatic resource constraints, as large-scale API access to other commercial LLMs (e.g., Gemini, Claude, Llama 3) would have required substantial funding. Nevertheless, ChatGPT represents not only one of the most widely used chatbot platforms worldwide, but its users also display a high tendency for self-diagnosis, making these findings clinically relevant. Moreover, the presented methodology is model-agnostic and can be applied to other systems in future collaborations or funded projects.
A further major limitation of this study is the use of a rigid, uniform prompt for a single hypothetical patient profile. While this approach enhanced internal consistency and allowed us to isolate model-intrinsic variability, it inevitably limited the assessment of the models’ contextual reasoning. More complex and clinically variable inputs – for instance, including age-related comorbidities such as renal insufficiency or hepatic dysfunction – could elicit different or even inconsistent adjustments of reference intervals. Thus, our design may have underestimated the variability that could occur in real-world, patient-specific queries, where models dynamically adapt to contextual cues. Future studies should systematically incorporate such contextual variability to better understand how patient- and prompt-related factors influence the reliability of LLM outputs.
Another important limitation concerns the unit-consistency threshold: by excluding analytes for which fewer than 80% of outputs used a consistent unit notation, we may have introduced a selection bias in regard to analytes. This criterion was implemented to ensure processability and avoid artificial inflation of variability metrics. However, it likely led to the preferential exclusion of analytes with intrinsically high unit variability. Future studies could mitigate this bias by applying unit-harmonization algorithms or stratifying analyses based on unit-consistency levels.
Despite these limitations, this study yields four important implications for the clinical use of AI-based chatbots in laboratory medicine.
Firstly, due to the inherent variability and potential for misinterpretation, such chatbots should be exclusively utilized by healthcare professionals who are trained to formulate high-quality prompts and able to critically assess the outputs. It is imperative to discourage patients from using chatbots to interpret their own laboratory results.
Secondly, the newer LLM versions tested, specifically GPT-4 and GPT-4o, demonstrate significantly lower variability compared to GPT-3.5-Turbo, making them preferable for future healthcare applications.
Thirdly, to further mitigate variability, adopting standardized prompt designs that consistently incorporate reference intervals is essential, as the absence of these safeguards will lead to persistent post-analytical errors.
Fourthly, concerted global initiatives for standardization or harmonization and the identification of reference measurement procedures and reference materials are important [International Federation of Clinical Chemistry and Laboratory Medicine (IFCC), 2024b]. Especially for LLMs it is of worth to have semantic and syntactic standards for laboratory reports available (Bietenbeck and Streichert, 2021). This is of course not limited to reference intervals and units but needs to be extended to Logical Observation Identifiers Names and Codes or laboratory reporting formats. Once standardization and harmonization are achieved, integrating these laboratory datasets into LLM training could enhance reproducibility, particularly for analytes that are currently poorly defined.
Finally, the misalignment between the precision required in laboratory medicine and the probabilistic nature of LLM outputs, combined with the high tendency of non-medical users to self-diagnose and omit reference intervals, creates a potentially hazardous situation. From a regulatory perspective, these findings underscore the importance of aligning LLM use in clinical contexts with emerging frameworks such as the EU Artificial Intelligence Act and FDA guidance on Software as a Medical Device. These frameworks emphasize transparency, risk management, and traceability—principles that are particularly relevant when outputs may influence clinical decision-making (Baumgartner and Baumgartner, 2023).
In summary, while advances in AI models hold promise for clinical laboratory medicine, relevant challenges remain in ensuring reliable and reproducible interpretations of laboratory data.
At present, ChatGPT’s high variability in regard to reference intervals leaves “curbside consultation” a mere aspiration for the future (Lee et al., 2023). Even after minimizing linguistic ambiguity (Yang et al., 2023) challenges such as hallucinations and output variability persist. Therefore, two practical implications remain. First, chatbots accessible to medical laypersons should be trained (or rule-augmented) to detect when a laboratory query lacks an explicit reference interval and prompt the user to supply it, rather than returning a possibly misleading value. Second, every gain in real-world reference interval harmonization and standardization will indirectly stabilize LLM outputs, because the models’ training corpora will contain fewer conflicting examples. Thus, while ongoing model refinement and domain-specific training are essential, they are not sufficient on their own. True progress depends on the global standardization and harmonization of reference intervals and units, as well as continued investigation into the capabilities and limitations of language models. Achieving these objectives will bridge the gap between current experimental applications and the reliable, real-world use of AI-driven chatbots in clinical decision support.
The raw data supporting the conclusions of this article will be made available by the corresponding author on reasonable request.
AM: Funding acquisition, Conceptualization, Formal analysis, Investigation, Data curation, Visualization, Writing – original draft, Writing – review & editing. ES: Writing – review & editing. TS: Supervision, Conceptualization, Writing – review & editing, Validation, Resources.
The author(s) declare that financial support was received for the research and/or publication of this article. This project was supported by ‘OpenAI’s Researcher Access Program and API’ and the Publication cost was funded by the German Research Foundation. Moreover, AM received Speaker Support at the Congress of the German Society for Clinical Chemistry and Laboratory Medicine 2024 as well as the ‘Digital Laboratory’ research award from the same society for her previous work on AI-based chatbots. Open AI was not involved in the study design, collection, analysis, interpretation of data, the writing of this article, or the decision to submit it for publication.
We are grateful to the Stiftung für Pathobiochemie und Molekulare Diagnostik – Referenzinstitut für Bioanalytik for their support in classifying analytes as either poorly standardized or well standardized.
The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
The author(s) declare that Gen AI was used in the creation of this manuscript. DeepL and ChatGPT for proofreading this article and their linguistic assistance as well as ChatGPT for facilitating the statistical programming, with AI-based output being critically reviewed by the authors.
Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.
The Supplementary material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/frai.2025.1681979/full#supplementary-material
Baumgartner, C., and Baumgartner, D. (2023). A regulatory challenge for natural language processing (NLP)-based tools such as ChatGPT to be legally used for healthcare decisions. Where are we now? Clin. Transl. Med. 13:e1362. doi: 10.1002/ctm2.1362,
PubMed Abstract | Crossref Full Text | Google Scholar
Bender, E. M., Gebru, T., McMillan-Major, A., and Shmitchell, S. (2021). Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency. Virtual Event, Canada: Association for Computing Machinery, 610–623.
Google Scholar
Bietenbeck, A., and Streichert, T. (2021). Preparing Laboratories for Interconnected Health Care. Diagnostics (Basel). 11:1487. doi: 10.3390/diagnostics11081487,
PubMed Abstract | Crossref Full Text | Google Scholar
Bohn, M. K., Bailey, D., Balion, C., Cembrowski, G., Collier, C., De Guire, V., et al. (2023). Reference interval harmonization: harnessing the power of big data analytics to derive common reference intervals across populations and testing platforms. Clin. Chem. 69, 991–1008. doi: 10.1093/clinchem/hvad099,
PubMed Abstract | Crossref Full Text | Google Scholar
Cadamuro, J., Cabitza, F., Debeljak, Z., Bruyne, S. D., Frans, G., Perez, S. M., et al. (2023). Potentials and pitfalls of ChatGPT and natural-language artificial intelligence models for the understanding of laboratory medicine test results. An assessment by the European Federation of Clinical Chemistry and Laboratory Medicine (EFLM) working group on artificial intelligence (WG-AI). Clin. Chem. Lab. Med. 61, 1158–1166. doi: 10.1515/cclm-2023-0355,
PubMed Abstract | Crossref Full Text | Google Scholar
Chan, C-h, Leeper, TJ, Becker, J, and Schoch, D (2023). rio: A Swiss-army knife for data file I/O.
Google Scholar
Coskun, A., and Lippi, G. (2024). The impact of physiological variations on personalized reference intervals and decision limits: an in-depth analysis. Clin. Chem. Lab. Med. 62, 2140–2147. doi: 10.1515/cclm-2024-0009,
PubMed Abstract | Crossref Full Text | Google Scholar
Deutsche Gesellschaft für Klinische Chemie und Laboratoriumsmedizin (DGKL). Zulässige Messunsicherheit—Excel-Datei Zulässige Messunsicherheit 2021. Available online at: https://dgkl.de/wp-content/uploads/2024/07/2021_01_06_Zula__ssige_Messunsicherheit_DGKL.3.10.20_-1.xlsx
Google Scholar
Franc, J. M., Cheng, L., Hart, A., Hata, R., and Hertelendy, A. (2024). Repeatability, reproducibility, and diagnostic accuracy of a commercial large language model (ChatGPT) to perform emergency department triage using the Canadian triage and acuity scale. CJEM 26, 40–46. doi: 10.1007/s43678-023-00616-w,
PubMed Abstract | Crossref Full Text | Google Scholar
Franc, J. M., Hertelendy, A. J., Cheng, L., Hata, R., and Verde, M. (2024). Accuracy of a commercial large language model (ChatGPT) to perform disaster triage of simulated patients using the simple triage and rapid treatment (START) protocol: gage repeatability and reproducibility study. J. Med. Internet Res. 26:e55648. doi: 10.2196/55648,
PubMed Abstract | Crossref Full Text | Google Scholar
Frieder, S., Pinchetti, L., Griffiths, R.-R., Salvatori, T., Lukasiewicz, T., Petersen, P., et al. (2024). Mathematical capabilities of chatgpt. Adv. Neural Inf. Proces. Syst. 36, 27699–27744. doi: 10.48550/arXiv.2301.13867
Crossref Full Text | Google Scholar
Garnier, S, Ross, N, Rudis, R, Camargo, PA, Sciaini, M, and Scherer, C. (2024). Viridis(lite)—Colorblind-friendly color maps for R. Vienna, Austria: R Foundation for Statistical Computing.
Google Scholar
Gillery, P., and Young, I. S. (2013). Progress towards standardization: an IFCC scientific division perspective. Clin. Chem. Lab. Med. 51, 915–918. doi: 10.1515/cclm-2013-0081,
PubMed Abstract | Crossref Full Text | Google Scholar
Girton, M. R., Greene, D. N., Messerlian, G., Keren, D. F., and Yu, M. (2024). ChatGPT vs medical professional: Analyzing responses to laboratory medicine questions on social media. Clin. Chem. 70, 1122–1139. doi: 10.1093/clinchem/hvae093,
PubMed Abstract | Crossref Full Text | Google Scholar
Guidi, G. C., Lippi, G., Solero, G. P., Poli, G., and Plebani, M. (2006). Managing transferability of laboratory data. Clin. Chim. Acta 374, 57–62. doi: 10.1016/j.cca.2006.06.009,
PubMed Abstract | Crossref Full Text | Google Scholar
Haeckel, R., Wosniok, W., Arzideh, F., Zierk, J., Gurr, E., and Streichert, T. (2017). Critical comments to a recent EFLM recommendation for the review of reference intervals. Clin. Chem. Lab. Med. 55, 341–347. doi: 10.1515/cclm-2016-1112,
PubMed Abstract | Crossref Full Text | Google Scholar
Hansen, Y. B. L. (2019). Recommendations on measurement units—why and how. EJIFCC 30, 250–275.
Google Scholar
Hu, K. ChatGPT sets record for fastest-growing user base—analyst note: Reuters; (2023). Available online at: https://www.reuters.com/technology/chatgpt-sets-record-fastest-growing-user-base-analyst-note-2023-02-01/
Google Scholar
Ibrahim, R. B., Chokkalla, A. K., Levett, K., Gustafson, D., Olayinka, L., Kumar, S., et al. (2023). ChatGPT-exploring its role in clinical chemistry. Ann. Clin. Lab. Sci. 53, 835–839,
PubMed Abstract | Google Scholar
International Consortium for Harmonization of Clinical Laboratory Results. Measurands Harmonization. (2025). Available online at: https://www.harmonization.net
Google Scholar
International Federation of Clinical Chemistry and Laboratory Medicine (IFCC). Reference Materials (2024a). Available online at: https://ifcc.org/ifcc-scientific-division/reference-materials/.
Google Scholar
International Federation of Clinical Chemistry and Laboratory Medicine (IFCC). Traceability in Laboratory Medicine (C-TLM) (2024b). Available online at: https://ifcc.org/ifcc-scientific-division/sd-committees/c-tlm/
Google Scholar
Jones, G., and Barker, A. (2007). Standardisation of reference intervals: an Australasian view. Clin. Biochem. Rev. 28, 169–173,
PubMed Abstract | Google Scholar
Klee, G. G. (2004). Clinical interpretation of reference intervals and reference limits. A plea for assay harmonization. Clin. Chem. Lab. Med. 42, 752–757. doi: 10.1515/CCLM.2004.127,
PubMed Abstract | Crossref Full Text | Google Scholar
Koerbin, G., Sikaris, K., Jones, G. R. D., Flatman, R., and Tate, J. R. (2018). An update report on the harmonization of adult reference intervals in Australasia. Clin. Chem. Lab. Med. 57, 38–41. doi: 10.1515/cclm-2017-0920,
PubMed Abstract | Crossref Full Text | Google Scholar
Lee, P., Bubeck, S., and Petro, J. (2023). Benefits, limits, and risks of GPT-4 as an AI chatbot for medicine. N. Engl. J. Med. 388, 1233–1239. doi: 10.1056/NEJMsr2214184,
PubMed Abstract | Crossref Full Text | Google Scholar
Liu, M., Okuhara, T., Chang, X., Shirabe, R., Nishiie, Y., Okada, H., et al. (2024). Performance of ChatGPT across different versions in medical licensing examinations worldwide: systematic review and Meta-analysis. J. Med. Internet Res. 26:e60807. doi: 10.2196/60807,
PubMed Abstract | Crossref Full Text | Google Scholar
Meyer, A., Riese, J., and Streichert, T. (2024a). Comparison of the performance of GPT-3.5 and GPT-4 with that of medical students on the written German medical licensing examination: observational study. JMIR Med. Educ. 10:e50965. doi: 10.2196/50965,
PubMed Abstract | Crossref Full Text | Google Scholar
Meyer, A., Soleman, A., Riese, J., and Streichert, T. (2024b). Comparison of ChatGPT, Gemini, and Le chat with physician interpretations of medical laboratory questions from an online health forum. Clin. Chem. Lab. Med. 62, 2425–2434. doi: 10.1515/cclm-2024-0246,
PubMed Abstract | Crossref Full Text | Google Scholar
Nguyen, D., MacKenzie, A., and Kim, Y. H. (2024). Encouragement vs. liability: how prompt engineering influences ChatGPT-4’s radiology exam performance. Clin. Imaging 115:110276. doi: 10.1016/j.clinimag.2024.110276,
PubMed Abstract | Crossref Full Text | Google Scholar
O’Connor, J. Steps to Use ChatGPT-4 for Blood Test Translation—A Guide to Interpreting Blood Test Results with ChatGPT-4: Medium; (2023). Available online at: https://generativeai.pub/steps-to-use-chatgpt-4-for-blood-work-translation-da99f266cbe3
Google Scholar
Plebani, M. (2013). Harmonization in laboratory medicine: the complete picture. Clin. Chem. Lab. Med. 51, 741–751. doi: 10.1515/cclm-2013-0075,
PubMed Abstract | Crossref Full Text | Google Scholar
Plebani, M. (2016). Harmonization in laboratory medicine: requests, samples, measurements and reports. Crit. Rev. Clin. Lab. Sci. 53, 184–196. doi: 10.3109/10408363.2015.1116851,
PubMed Abstract | Crossref Full Text | Google Scholar
Plebani, M. (2018). Harmonization in laboratory medicine: more than clinical chemistry? Clin. Chem. Lab. Med. 56, 1579–1586. doi: 10.1515/cclm-2017-0865,
PubMed Abstract | Crossref Full Text | Google Scholar
Plebani, M., and Lippi, G. (2023). Standardization and harmonization in laboratory medicine: not only for clinical chemistry measurands. Clin. Chem. Lab. Med. 61, 185–187. doi: 10.1515/cclm-2022-1122,
PubMed Abstract | Crossref Full Text | Google Scholar
Shahsavar, Y., and Choudhury, A. (2023). User intentions to use ChatGPT for self-diagnosis and health-related purposes: cross-sectional survey study. JMIR Hum. Factors 10:e47564. doi: 10.2196/47564,
PubMed Abstract | Crossref Full Text | Google Scholar
Sikaris, K. (2015). Performance criteria of the post-analytical phase. Clin. Chem. Lab. Med. 53, 949–958. doi: 10.1515/cclm-2015-0016,
PubMed Abstract | Crossref Full Text | Google Scholar
Sjoberg, DD, Larmarange, J, Curry, M, Lavery, J, Whiting, K, and Zabor, EC (2022. gtsummary: presentation-ready data summary and analytic result tables.
Google Scholar
Team RC (2022). R: A language and environment for statistical computing. Vienna, Austria: R Foundation for Statistical Computing.
Google Scholar
Vesper, H. W., Myers, G. L., and Miller, W. G. (2016). Current practices and challenges in the standardization and harmonization of clinical laboratory tests1223. Am. J. Clin. Nutr. 104, 907S–912S. doi: 10.3945/ajcn.115.110387
Crossref Full Text | Google Scholar
Wang, L., Bi, W., Zhao, S., Ma, Y., Lv, L., Meng, C., et al. (2024). Investigating the impact of prompt engineering on the performance of large language models for standardizing obstetric diagnosis text: comparative study. JMIR Form Res. 8:e53216. doi: 10.2196/53216,
PubMed Abstract | Crossref Full Text | Google Scholar
Wickham, H., Averick, M., Bryan, J., Chang, W., McGowan, L. D. A., François, R., et al. (2019). Welcome to the {tidyverse}. J. Open Source Softw. 4:1686. doi: 10.21105/joss.01686
Crossref Full Text | Google Scholar
Yang, W. H., Chan, Y. H., Huang, C. P., and Chen, T. J. (2024). Comparative analysis of GPT-3.5 and GPT-4.0 in Taiwan’s medical technologist certification: a study in artificial intelligence advancements. J. Chin. Med. Assoc. 87, 525–530. doi: 10.1097/JCMA.0000000000001092,
PubMed Abstract | Crossref Full Text | Google Scholar
Yang, H. S., Wang, F., Greenblatt, M. B., Huang, S. X., and Zhang, Y. (2023). AI chatbots in clinical laboratory medicine: foundations and trends. Clin. Chem. 69, 1238–1246. doi: 10.1093/clinchem/hvad106,
PubMed Abstract | Crossref Full Text | Google Scholar
Keywords: chatbot, ChatGPT, reference interval, repeatability, consistency, large language model
Citation: Meyer A, Schömig E and Streichert T (2025) ChatGPT and reference intervals: a comparative analysis of repeatability in GPT-3.5 Turbo, GPT-4, and GPT-4o. Front. Artif. Intell. 8:1681979. doi: 10.3389/frai.2025.1681979
Received: 14 August 2025; Revised: 06 November 2025; Accepted: 12 November 2025;
Published: 12 December 2025.
Edited by:
Reviewed by:
Copyright © 2025 Meyer, Schömig and Streichert. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
*Correspondence: Annika Meyer, QW5uaWthLm1leWVyMUB1ay1rb2Vsbi5kZQ==
Disclaimer: All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article or claim that may be made by its manufacturer is not guaranteed or endorsed by the publisher.
Share on
Share on
Frontiers' quality
Our research integrity and auditing teams lead the rigorous process that protects the quality of the scientific record
Supplementary Material

source