Abstract
Background:
Artificial intelligence and large language models (LLMs) have developed rapidly in recent years and involved in medical education, in addition to clinical care. However, it remains unknown how the performance of LLMs in neonatal resuscitation compares to that of healthcare professionals (HCPs). In this exploratory study, we aimed to investigate and compare the performance of LLMs with those of HCPs on 3 sources of examination questions in the neonatal resuscitation training in Canada and China.
Methods:
In this bi-center study, we evaluated the overall accuracy, accuracy across question types, and reliability of LLMs’ (ChatGPT-5 and DeepSeek-R1) responses to neonatal resuscitation questions from workshop in China, NRP® textbook (8th edition), and Kahoot quizzes in Canada. Each LLM was tested three times per question with one-week intervals in August and September, 2025. Comparisons with historical scores of HCPs in the training courses in China (2022–2024) and Canada (2022–2025) were performed.
Results:
Both LLMs performed comparably to HCPs in Chinese examinations, and showed similar accuracy in NRP® textbook questions with higher scores on multiple-choice than on short answer questions. ChatGPT achieved higher accuracy than DeepSeek and HCPs in the Kahoot quizzes. ChatGPT also had higher accuracy on scenario-based than on non-scenario-based questions in the workshop examination. High reliability of LLMs’ responses was found (Fleiss’ Kappa>0.89).
Conclusion:
Both ChatGPT and DeepSeek achieved accuracy comparable to that of HCPs and showed high consistency on selected neonatal resuscitation written examinations. The findings warrant further research to explore their potential integration in neonatal resuscitation training.
Introduction
The limited training opportunities and rapid decay of neonatal resuscitation knowledge and skills among healthcare providers (HCPs) after training represent critical barriers in neonatal resuscitation training programs (Sihota et al., 2025). Large language models (LLMs) have the potential to learn from vast medical corpora and thus may offer a solution for providing educational opportunities (Goh et al., 2024; Demirtas et al., 2025; Singhal et al., 2023).
Despite the growing integration into various medical domains, it remains unknown how the performance of LLMs in neonatal resuscitation assessment compares to that of HCPs. This knowledge gap hinders the ability to determine their safe and effective role in education and clinical practice. In addition, LLMs might exhibit linguistic strength based on their primary language of training, achieving concordance in other languages might be a challenge (Dzuali et al., 2024). In languages with limited training data, LLMs are prone to filling knowledge gaps by generating fabricated information, a phenomenon known as artificial intelligence (AI) hallucination. The persuasive but inaccurate, or biased output carries substantial risk of harm (Singhal et al., 2023).
In this exploratory study, we aimed to investigate and compare the capability of LLMs with those of participants (all HCPs) in the neonatal resuscitation training in Canada and China. We assessed the performance of ChatGPT and DeepSeek in three sources of neonatal resuscitation examinations in English and Chinese languages. We tested the hypothesis that the accuracy of both LLMs would be comparable to that of HCPs on the selected neonatal resuscitation written examinations.
Methods
Study design, testing LLMs and population
This bi-center comparative study evaluated the performance of ChatGPT and DeepSeek on neonatal resuscitation questions from neonatal resuscitation workshop in China (38 questions, Chinese), the Neonatal Resuscitation Program® (NRP®) 8th edition textbook (75 questions, English) (American Academy of Pediatrics, 2021), and Kahoot quizzes in Canada (21 questions, English). Both the workshop examinations and Kahoot quizzes were designed by the institutional educators (CX and QZ, SK and DW, respectively, and other clinicians in the local neonatal resuscitation training team) in accordance with the NRP® 8th edition (American Academy of Pediatrics, 2021) (Figure 1). All questions including images were excluded. Scenario-based questions were defined as those set within a specific clinical scenario or context. These questions assessed clinical reasoning, decision-making and application of knowledge, as opposed to non-scenario-based questions which tested isolated facts or direct recalls of knowledge (Supplementary material).
Figure 1
ChatGPT-5 (via the website, released in August 2025, OpenAI, San Francisco, CA, United States) and DeepSeek-R1 (via API, 671B parameters, R1-0528 version, June 2025 update, DeepSeek, Hangzhou, China) were used. Browsing and retrieval were disabled. Each LLM was tested three times with one-week intervals in August and September 2025. Questions were posed in their original language using simple prompt: “Please answer the following questions.” Chain-of-thought or multi-step reasoning was provided. In each of three separate testing sessions, every query was submitted once without repetition. A new conversation session was started without any prior chat history.
The scores of 286 HCPs in the neonatal resuscitation workshop conducted by the University of Hong Kong-Shenzhen Hospital, China (2022–2024), and Kahoot scores of 251 HCPs in NRP® training in Royal Alexandra Hospital, University of Alberta, Canada (2022–2025) were reviewed. The Kahoot HCP cohort completed the quizzes as part of the NRP® training, after the didactic lecture and before the simulation training.
Assessment of LLM responses
LLM response was scored in binary as correct or incorrect. For short answer questions, any response that was incorrect or incomplete was considered incorrect. Two neonatologists (YC, CX) independently evaluated all LLM responses using pre-defined scoring criteria. YC was blinded to response source, while CX was involved in LLM response collection and was aware of the source. Discrepancies were resolved through final cross-check.
Outcome measures
The primary outcome was overall accuracy (expressed as percentage of number of correct answers/total number of answers). The secondary outcomes included reliability of LLMs’ responses, accuracy for scenario-based, multiple-choice and short answer questions.
Statements of ethics
This study was reviewed and approved by the Institutional Review Board of the University of Hong Kong-Shenzhen Hospital (#hkuszh2025063) and the Institutional Review Board of the University of Alberta (#Pro00156941). Written informed consent was obtained from participants in the neonatal resuscitation workshops held at the University of Hong Kong-Shenzhen Hospital to participate in the study. For the anonymous Kahoot scores, the Institutional Review Board of the University of Alberta granted an exemption from requiring written informed consent for the use of data.
Statistical analysis
Differences in accuracy between groups (ChatGPT, DeepSeek, and HCPs) were assessed using Chi-square test or Fisher’s exact test. In addition to analyzing the accuracy of LLMs in all questions in repeated sessions (results in Supplementary material), to avoid clustering bias from three correlated runs per question, we aggregated the three binary outcomes (correct/incorrect) using majority voting. A question was scored as correct if at least two of three runs were correct. This yielded one binary outcome per model per question. The inter-rater agreement on three responses generated by LLMs was evaluated using Fleiss’ Kappa. SPSS Statistics (v.27) was used for data analyses. A p value less than 0.05 was considered statistically significant.
Results
Agreement between two assessors was high, with differences observed in 6 out of 402 answers. Discrepancies were resolved after cross-check, and a final binary outcome was agreed for each response.
Performance comparison between LLMs and HCPs
In the neonatal resuscitation workshop examinations, ChatGPT and DeepSeek showed similar accuracy rates (90 and 92%, p = 0.692) (Table 1). HCPs significantly improved the scores after training (92% vs. 82% of pre-training, p < 0.005) but both pre-training and post-training scores were comparable to those of LLMs. In the Kahoot quizzes, ChatGPT had the highest accuracy rate (95%), followed by DeepSeek (81%) and HCPs (78%) but the differences were not significant (Table 1 and Supplementary material).
Table 1
| Assessments | Overall accuracy % (n/N)# | p-value | ||
|---|---|---|---|---|
| vs. DeepSeek-R1 | vs. HCPs pre-training 81.9% (8,343/10184) N = 268 | vs. HCPs post-training 91.7% (4,702/5130) N = 135 | ||
| Neonatal resuscitation workshop examinations (Chinese) | ChatGPT-5 89.5% (34/38) | 0.692 | 0.227 | 0.628 |
| DeepSeek-R1 92.1% (35/38) | — | 0.103 | 0.921 | |
| Kahoot quizzes (English) | vs. DeepSeek-R1 | vs. HCPs 78.3% (4,128/5271) N = 251 | ||
| ChatGPT-5 95.2% (20/21) | 0.343 | 0.064 | ||
| DeepSeek-R1 81.0% (17/21) | — | 1.000 | ||
Comparison of performance between large language models and healthcare providers (HCPs).
#: n/N represents the percentage of total correct answers over the total number of questions.
Performance across question types
ChatGPT achieved higher scores in scenario-based than in non-scenario-based questions in neonatal resuscitation workshop (100% vs. 83%), but the difference was not statistically significant (Table 2 and Supplementary material). DeepSeek and HCPs had similar accuracy rates. ChatGPT and DeepSeek showed similar accuracy in NRP® 8th edition textbook questions. Both LLMs scored higher on multiple-choice than on short answer questions (Table 2).
Table 2
| Model or Data source | Overall accuracy, % (n/N)a | Scenario-based questions, % (n/N)a | Non scenario-based Questions, % (n/N)a | p-value |
|---|---|---|---|---|
| ChatGPT-5 | ||||
| Neonatal resuscitation workshop examination | 89.5% (34/38) | 100% (14/14) | 83.3% (20/24) | 0.276 |
| NRP® 8th textbook | 93.3% (70/75) | 90.5% (19/21) | 94.4% (51/54) | 0.645 |
| Kahoot quizzesa,b | 95.2% (20/21) | 100% (1/1) | 95.0% (19/20) | b |
| DeepSeek-R1 | ||||
| Neonatal resuscitation workshop examination | 92.1% (35/38) | 92.9% (13/14) | 91.7% (22/24) | 1.000 |
| NRP® 8th textbook | 90.7% (68/75) | 90.5% (19/21) | 90.7% (49/54) | 1.000 |
| Kahoot quizzesa,b | 81.0% (17/21) | 100% (1/1) | 80% (16/20) | b |
| Healthcare providers (HCPs) | ||||
| Neonatal resuscitation workshop examination (135 HCPs) | 91.7% (4,702/5130) | 92.1% (1739/1890) | 91.5% (2,963/3240) | 0.484 |
| Kahoot quizzesb (251 HCPs) | 78.3% (4,128/5271) | 85.3% (214/251) | 78.0% (3,914/5020) | b |
| NRP® 8th Textbook | Overall accuracy, % (n/N)a | Multiple-choice questions, % (n/N)a | Short answer questions, % (n/N)a | p-value |
|---|---|---|---|---|
| ChatGPT-5 | 93.3% (70/75) | 97.0% (64/66) | 66.7% (6/9)* | 0.011 |
| DeepSeek-R1 | 90.7% (68/75) | 95.5% (63/66) | 55.6% (5/9)* | 0.003 |
Comparison of performance between large language models and healthcare providers by question format, model, and data source.
a: n/N represents the percentage of total correct answers over the total number of questions.
b: No comparison was performed because there was only one scenario-based question in the Kahoot quiz.
Accuracy and reliability analysis
The reliability of LLMs’ responses for all examinations was high (Fleiss’ Kappa values all ≥0.90, p < 0.001) (Table 3).
Table 3
| Data source | Model | Test 1 (%) | Test 2 (%) | Test 3 (%) | Fleiss’ Kappa (95% CI) | p-value |
|---|---|---|---|---|---|---|
| NRP® 8th Textbook | ChatGPT | 95.3 | 96.0 | 96.7 | 0.985 (0.978–0.991) | <0.001 |
| DeepSeek | 93.3 | 95.3 | 92.7 | 0.941 (0.914–0.962) | <0.001 | |
| Neonatal resuscitation workshop examinations | ChatGPT | 86.8 | 89.5 | 92.1 | 0.988 (0.980–0.993) | <0.001 |
| DeepSeek | 94.7 | 92.1 | 89.5 | 0.898 (0.833–0.942) | <0.001 | |
| Kahoot quizzes | ChatGPT | 90.5 | 95.2 | 95.2 | 0.900 (0.807–0.955) | <0.001 |
| DeepSeek | 71.4 | 85.7 | 76.2 | 0.953 (0.905–0.979) | <0.001 |
Accuracy and reliability analysis of large language models across repeated tests.
Language concordance and AI hallucination
Simple calculation mistakes and cross-lingual inconsistency were noted in some answers by both LLMs. The performance of two LLMs varied across epinephrine and flush volume and was influenced by language. When prompted in English, ChatGPT tended to provide the correct answer, whereas DeepSeek answered more correctly in Chinese. For the question on epinephrine flush volume in English Kahoot quizzes, ChatGPT consistently answered correctly, whereas DeepSeek consistently answered incorrectly in three runs. In contrast, for the same question presented in Chinese workshop examinations, DeepSeek answered correctly in all three runs, while ChatGPT provided incorrect answers in two of the three runs (Supplementary material).
Discussion
This study suggests potential strength of LLMs in selected neonatal resuscitation written examinations with performance comparable to HCPs by rigorous, bilingual comparisons. Our results indicate that LLMs perform well in overall accuracy and scenario-based problem-solving in neonatal resuscitation, with higher scores on multiple-choice than on short answer questions.
LLMs use deep learning to understand and generate output. They excel at tasks involving language comprehension, contextual reasoning, and synthesizing information based on artificial neural networks trained on vast datasets (Tangsrivimol et al., 2025). Scenario-based questions are fundamentally reasoning tasks. They are designed to evaluate the ability to synthesize theoretical knowledge and apply it to a specific clinical context. This requires integrating multiple variables, recognizing clues, and eliminating interfering factors to arrive at a valid conclusion. This form of contextualized reasoning aligns precisely with the strength of LLMs. However, LLMs might have been exposed to neonatal resuscitation concepts, as this knowledge is found in publicly available materials including NRP® textbook and journals. Their high accuracy rates might also reflect training data contamination, leading to memorization rather than true reasoning.
LLMs’ responses are stochastic. In evaluations of LLMs on life support exams, overall accuracy varied based on the model version and testing methodology. An earlier version of ChatGPT (version:9) failed to pass the American Heart Association (AHA) Basic Life Support (BLS) and Advanced Cardiovascular Life Support (ACLS) exams in a single-attempt setting, scoring 64–68% on BLS and 68.4–76.3% on ACLS exam (Fijačko et al., 2023). When three responses per question were generated and evaluated, ChatGPT achieved an overall accuracy of 84% on BLS exam and 78.95% on ACLS exam (Zhu et al., 2023). ChatGPT-4o achieved a 90.7% overall accuracy and exceeding all the passing score on all attempts on the Pediatric Advanced Life Support (PALS) exam (Kokulu et al., 2024), while in a recent study by Demirtas et al., it attained 75.3% on the NRP® exam over three attempts (Demirtas et al., 2025). To capture a broader spectrum of potential variation and more accurately assess real-world stability, we adopted repeated measuring over long intervals. This approach is designed to account for a wider range of fluctuation sources and thus provides a more reliable measure of real-world performance. We found that ChatGPT-5 achieved higher accuracy than models in previous studies, which might be attributable to additional training and iterative advancement of the model.
Beyond resuscitation, a growing body of research has also investigated LLMs in pediatrics and medical education (Mansoor et al., 2025; Huang et al., 2026; Suresh and Misra, 2024). Mansoor et al. reported that DeepSeek-R1 achieved an accuracy exceeding the typical first-time pass rates of the American Board of Pediatrics (Mansoor et al., 2025). However, LLM’s performance on multiple-choice questions should not be equated with superior clinical competence, as they have not been tested on history taking, physical examination, or real-world management planning (Mansoor et al., 2025). We collected data from HCPs retrospectively from previous training under time constraints and real-time feedback, whereas LLMs were tested in a untimed environment. Therefore, this was not a direct head-to-head comparison but an exploratory analysis comparing the performance of LLMs against historical data from HCPs. Direct comparisons are confounded by differences in testing environment, time pressure, and presence of feedback, possible distractions, and motivational context. High accuracy of LLMs on static written questions does not equate to clinical competence (Huang et al., 2026). Nevertheless, LLMs can pass standardized examinations and may serve as effective tools for question generation, personalized study plans, and immediate feedback (Suresh and Misra, 2024). Further research should explore how LLMs can optimally enhance HCP’s learning efficiency.
LLMs exhibit relatively poor performance on simple arithmetic tasks such as single-digit addition (Goodell et al., 2025). A study reported that ChatGPT was an unreliable clinical calculator, with only a 66% accuracy (Goodell et al., 2025). Errors in simple medication dosage calculations were also identified in our study. This inadequacy poses a substantial obstacle to the implementation of LLMs, as their outputs require careful oversight in critical tasks such as determining the dosage and concentration of neonatal resuscitation medications.
Across one Chinese and two English tests, both ChatGPT and DeepSeek showed good overall accuracy. Interestingly, for the same question with identical content in Chinese and English, we observed that each LLM performed better in its dominant training language. This pattern aligns with the concept of language concordance, where LLMs tend to perform better when prompted in their dominant training language. Reasoning errors may arise when the prompt language differs from the model’s internal processing language (Strasser et al., 2026). In practice, educators need to consider model choice and prompt language to optimize its output.
Whether the advantages of LLMs could be translated into clinical practice and medical education also depend on the way they are used. While ChatGPT-4 had superior independent diagnostic reasoning capability compared to physicians, its use by physicians in facilitating diagnosis had not significantly improved diagnostic performance (Goh et al., 2024). However, ChatGPT-generated structured feedback significantly enhanced medical students’ clinical decision-making skills (Brügge et al., 2024). This indicates that simply using LLMs as information tools might fail to leverage their reasoning potential. A recent pediatric simulation study reported that LLM could generate real-time debrief scripts from simulation transcripts, improving discussion organization and learner engagement (Hong et al., 2025). Such applications suggest that LLMs may be useful in neonatal resuscitation training, where structured feedback and scenario-based learning are essential. However, further research is needed to assess LLMs’ performance in real-world settings before recommending their integration into neonatal resuscitation education.
Limitations
Our study has several limitations. First, training examinations including questions in the neonatal resuscitation workshops and NRP® provider courses could be stressful to some HCPs because the scores are important in the certification process. Second, Kahoot is a gamified, interactive platform but it has a countdown for each question. The pressure of time might have affected the ability of some HCPs to retrieve known information or forced them to make quick choices. Therefore, the performance of HCPs might be underestimated in these situations. Further, the professional background of HCPs was heterogeneous, including NICU doctors, NICU nurses, pediatricians, respiratory therapists, midwives, and obstetricians, with a wide range of clinical experience. We were unable to further stratify and compare the performance of subgroups with LLMs.
Conclusion
In this exploratory study, both ChatGPT and DeepSeek achieved accuracy comparable to that of HCPs and showed high consistency on selected neonatal resuscitation written examinations. The findings warrant further research to explore their potential integration in neonatal resuscitation training.
Statements
Data availability statement
The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.
Ethics statement
The studies involving humans were approved by Institutional Review Board of the University of Hong Kong Shenzhen Hospital (#2025063) and the Institutional Review Board of the University of Alberta (#Pro00156941). The studies were conducted in accordance with the local legislation and institutional requirements. The participants provided their written informed consent to participate in this study.
Author contributions
CX: Methodology, Data curation, Project administration, Formal analysis, Conceptualization, Validation, Writing – original draft, Software, Writing – review & editing, Resources, Funding acquisition, Investigation. YC: Project administration, Data curation, Methodology, Investigation, Writing – review & editing. SS: Software, Writing – review & editing, Investigation, Project administration, Methodology, Data curation. DW: Methodology, Conceptualization, Investigation, Supervision, Resources, Writing – review & editing, Project administration. QZ: Writing – review & editing, Investigation, Supervision, Funding acquisition, Resources, Project administration, Conceptualization, Methodology. GS: Project administration, Resources, Visualization, Supervision, Methodology, Conceptualization, Writing – review & editing. P-YC: Methodology, Supervision, Conceptualization, Writing – review & editing, Investigation, Visualization, Project administration, Validation.
Funding
The author(s) declared that financial support was received for this work and/or its publication. This study was supported by Medical Education Collaborative High-Quality Innovation and Development Research Project (YJXT20250507), Hospital Management Committee of Shenzhen University, 2025 and Guangdong Provincial Department of Education Research Project on Educational Reform (2023-157).
Acknowledgments
We sincerely thank other members of the TRAINinG (Drs. Genevieve Fung, Jiang-Qin Liu, Mabel Wong and Ming Zhou, and nurses Ms. Yin Xue and Do Zhang) NRP® instructors and participants in Royal Alexandra Hospital, Edmonton, Canada, and instructors and providers of neonatal resuscitation workshop in the University of Hong Kong-Shenzhen Hospital, Shenzhen, China.
Conflict of interest
The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
The author P-YC declare that they were an editorial board member of Frontiers, at the time of submission. This had no impact on the peer review process and the final decision.
Generative AI statement
The author(s) declared that Generative AI was not used in the creation of this manuscript.
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.
Correction note
A correction has been made to this article. Details can be found at: 10.3389/frai.2026.1934012.
Publisher’s note
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.
Supplementary material
The Supplementary material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/frai.2026.1838877/full#supplementary-material
References
1
American Academy of Pediatrics (2021). Textbook of Neonatal Resuscitation. Dallas: American Academy of Pediatrics.
2
BrüggeE.RicchizziS.ArenbeckM.KellerM. N.SchurL.StummerW.et al. (2024). Large language models improve clinical decision making of medical students through patient simulation and structured feedback: a randomized controlled trial. BMC Med. Educ.24:1391. doi: 10.1186/s12909-024-06399-7,
3
DemirtasM. S.KokuluK.TuncG. (2025). Evaluating artificial intelligence competency in education: performance evaluation of ChatGPT in the neonatal resuscitation program exam. Resuscitation209:110563. doi: 10.1016/j.resuscitation.2025.110563,
4
DzualiF.SeigerK.NovoaR.AleshinM.TengJ.LesterJ.et al. (2024). ChatGPT may improve access to language-concordant Care for Patients with non-English Language Preferences. JMIR Med Educ.10:e51435. doi: 10.2196/51435,
5
FijačkoN.GosakL.ŠtiglicG.PicardC. T.JohnD. M. (2023). Can ChatGPT pass the life support exams without entering the American heart association course?Resuscitation185:109732. doi: 10.1016/j.resuscitation.2023.109732,
6
GohE.GalloR.HomJ.StrongE.WengY.KermanH.et al. (2024). Large language model influence on diagnostic reasoning: a randomized clinical trial. JAMA Netw. Open7:e2440969. doi: 10.1001/jamanetworkopen.2024.40969,
7
GoodellA. J.ChuS. N.RouholimanD.ChuL. F. (2025). Large language model agents can use tools to perform clinical calculations. NPJ Digit Med.8:163. doi: 10.1038/s41746-025-01475-8,
8
HongE.KazmirS.DylikB.AuerbachM.RosatiM.AthanasopoulouS.et al. (2025). Exploring the use of a large language model in simulation debriefing: an observational simulation-based pilot study. Simul. Healthc.20, 366–371. doi: 10.1097/SIH.0000000000000861
9
HuangT.TseG.PagelerN. M.BannettY. (2026). Large language models using clinical text in Pediatrics: a scoping review. JAMA Netw. Open9:e262443. doi: 10.1001/jamanetworkopen.2026.2443,
10
KokuluK.DemirtaşM. S.SertE. T.MutluH. (2024). ChatGPT and pediatric advanced life support: a performance evaluation. Resuscitation205:110451. doi: 10.1016/j.resuscitation.2024.110451,
11
MansoorM.IbrahimA.HamideA. (2025). Performance of DeepSeek and GPT models on Pediatric board preparation questions: comparative evaluation. JMIR AI.4:e76056. doi: 10.2196/76056,
12
SihotaD.Lee HimR.DominguezG.HarrisonL.VaivadaT.BhuttaZ. A. (2025). Effectiveness of neonatal resuscitation training programs, implementation, and scale-up in low and middle-income countries. Neonatology122, 1–32. doi: 10.1159/000542539,
13
SinghalK.AziziS.TuT.MahdaviS. S.WeiJ.ChungH. W.et al. (2023). Large language models encode clinical knowledge. Nature620, 172–180. doi: 10.1038/s41586-023-06291-2,
14
StrasserL. M.AnschuetzW.DennstädtF.HastingsJ. (2026). Performance evaluation of large language models in multilingual medical multiple-choice questions: mixed methods study. JMIR Med Educ.12:e81399. doi: 10.2196/81399,
15
SureshS.MisraS. M. (2024). Large language models in Pediatric education: current uses and future potential. Pediatrics154:e064683. doi: 10.1542/peds.2023-064683,
16
TangsrivimolJ. A.DarzidehkalaniE.VirkH. U. H.WangZ.EggerJ.WangM.et al. (2025). Benefits, limits, and risks of ChatGPT in medicine. Front Artif Intell.8:1518049. doi: 10.3389/frai.2025.1518049,
17
ZhuL.MouW.YangT.ChenR. (2023). ChatGPT can pass the AHA exams: open-ended questions outperform multiple-choice format. Resuscitation188:109783. doi: 10.1016/j.resuscitation.2023.109783,
Summary
Keywords
large language model, neonatal resuscitation, healthcare professional (HCP), examination questions, language concordance
Citation
Xu C, Chen Y, Skelding S, Wang D, Zhang Q, Schmölzer GM and Cheung P-Y (2026) Performance of large language models in neonatal resuscitation assessments versus healthcare providers: an exploratory study. Front. Artif. Intell. 9:1838877. doi: 10.3389/frai.2026.1838877
Received
25 March 2026
Revised
31 May 2026
Accepted
31 May 2026
Published
25 June 2026
Corrected
22 July 2026
Volume
9 - 2026
Edited by
Vinod Kumar Chauhan, University of Strathclyde, United Kingdom
Reviewed by
Ahmet Aykut, SBU. Van Education and Research Hospital, Türkiye
Shai Itzhaki, Schneider Children's Medical Center, Israel
Updates
Copyright
© 2026 Xu, Chen, Skelding, Wang, Zhang, Schmölzer and Cheung.
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: Po-Yin Cheung, poyin@ualberta.ca
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.