Multidisciplinary expert evaluation of large language models on questions regarding bariatric surgery: a comparative analysis of ERNIE Bot 4.0, ChatGPT-4, Claude 3 Opus, and Gemini Pro

Abstract

Large language models (LLMs) have potential in bariatric surgery consultations, but current evaluations are limited to bariatric specialists, contradicting guidelines that call for multidisciplinary assessment. This study uses a multidisciplinary framework to evaluate LLM performance on bariatric surgery queries. Four LLMs (ERNIE Bot 4.0, ChatGPT-4, Claude 3 Opus, and Gemini Pro) were tested on 50 common bariatric surgery questions, generating 200 responses. A panel of seven experts (4 bariatric surgeons, 1 obesity physician, and 2 dietitians) assessed accuracy and comprehensiveness. Three prompt approaches were used to evaluate self-correction: basic review, web-enabled review, and evidence-based review. Qualitative analysis identified poorly rated responses. The study adheres to TRIPOD-LLM Statement reporting guideline. Rater agreement was fair, accuracy rating, Fleiss’ kappa = 0.210 (95% CI: 0.208–0.212; Z = 7.815; P < 0.001); comprehensiveness rating, Fleiss’ kappa = 0.464 (95% CI: 0.453–0.476; Z = 2.543; P < 0.011). Claude 3 Opus provided the longest answers, while ERNIE Bot 4.0 had the highest accuracy (19.46 ± 2.07). ChatGPT-4 had 90.0% “good” responses, compared to 84.0% for ERNIE Bot 4.0, 80.0% for Claude 3 Opus, and 48.0% for Gemini Pro (P ≤ 0.05). All LLMs performed sub-optimally in comprehensiveness (scores: 2.86–3.13 out of 5). However, they showed significant self-correction capabilities, especially when using internet searches and evidence-based resources. LLMs show potential for bariatric surgery education, but their direct clinical use is challenging due to variable accuracy and suboptimal comprehensiveness. Future efforts should focus on developing specialized LLMs with robust evidence and multidisciplinary input to ensure patient safety and optimal outcomes.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-46766-6.

Introduction

In the evolving landscape of medical artificial intelligence (AI)—computer systems designed to perform tasks that typically require human intelligence. Particularly LLMs, have introduced significant opportunities to optimize clinical workflows and educational initiatives. LLMs (a type of artificial intelligence system trained on massive amounts of text to understand and generate human-like language) are at the forefront of medical AI, with substantial potential to improve clinical, educational, and research productivity. However, these technologies require rigorous validation and further refinement to address existing limitations^1–4. Owing to their strong natural language processing capabilities, tools such as ChatGPT are increasingly recognized as valuable assets for supporting surgical diagnosis and treatment⁵. Nonetheless, research on using chatbots to handle bariatric surgery inquiries is nascent, requiring further investigation to fully realize their potential in this domain.

Obesity and overweight rates continue to rise worldwide. Current trends suggest that by 2030, 50% of adult men and women will be living with high BMI. In the same year, 17% of men and 22% of women will be living with obesity⁶. Bariatric surgery remains one of the most effective and reliable interventions for weight loss and improves obesity-related comorbidities such as diabetes, hypertension, and obstructive.

sleep apnea^7[,8. However, less than 1% of eligible patients undergo this procedure, a low uptake attributed to multiple factors⁹, including patient misunderstandings, insurance limitations, restricted access to resources, and provider misconceptions^10[,11. The lack of accessible information on the safety and efficacy of bariatric surgery further impedes well-informed decision-making among patients and healthcare providers⁹.

LLMs may help overcome two key barriers to broader adoption: misinformation and limited access to specialist expertise¹². When properly supervised by healthcare professionals, LLMs may be a supportive tool in countering widespread misconceptions about bariatric procedures that often circulate on social media and the internet¹³, thereby expanding education, easing workforce pressures, and reducing care inequities¹⁴. Although international guidelines require bariatric surgery to be managed by a multidisciplinary team, existing studies remain limited to evaluations from single bariatric surgeons, use non-standardized criteria, and lack blinding, which limits generalizability and risks bias^15[,16. Additionally, no published evaluations have assessed LLM performance in bariatric contexts within China. This gap is noteworthy given projections that by 2025 up to 41% of Chinese adults may have BMI ≥ 25 kg/m² and 9% BMI ≥ 30 kg/m², with an estimated 515 million adults affected by 2030¹⁷. To address these gaps, we compared leading international LLMs and China’s domestic models (including ERNIE Bot) using a multidisciplinary framework to assess their ability to answer bariatric surgery–related queries.

Methods

Ethics

All patient‑derived queries were fully anonymized prior to analysis. Anonymization included the removal of names, dates, geographic information, and any contextual details that could potentially identify an individual. The dataset was subsequently screened by the study team to exclude sensitive or inappropriate content and to ensure compliance with ethical standards. The study met criteria for IRB exemption, as no identifiable private information was used. The study adheres to TRIPOD-LLM Statement reporting guideline (Supplementary Table 8).

Study design

The study began with an initial pool of 200 questions collected from literature^10[,18, social-media platforms (Little Red Book and Douyin), and patients at our bariatric center. All prompts and model outputs were generated in English and subsequently translated into Chinese by the first author, with verification by two additional co‑authors (both bariatric physicians) before being provided to the expert graders. All LLMs were queried with temperature = 0 and default top-p settings to ensure deterministic outputs; no randomness-enhancing parameters were used. Three bariatric surgeons (Y**, C**, and C**) independently assessed all items for clinical relevance, domain coverage, clarity, and redundancy using a 4-point scale, and item level Content Validity Index (I-CVI) values were calculated. Items with I-CVI < 0.78 or judged redundant were revised or removed, resulting in 50 validated questions. The experts also confirmed balanced distribution across five domains: (1) eligibility, efficacy, and procedure options; (2) preoperative preparation; (3) recovery, risks, and complications; (4) lifestyle changes; and (5) other (Fig. 1 shows the percentage of bariatric surgery questions in each domain). From February 1–10, 2024, responses to these questions were generated using four widely used and technically advanced LLMs—ERNIE Bot 4.0 (Baidu, China), ChatGPT-4 (Open AI, USA), Claude-3 (Anthropic, USA), and Gemini Pro (Google, USA)—selected for their market prominence, geographic representation, and relevance to clinical applications¹⁸. All the models are publicly accessible via stable APIs, support multiple languages, and demonstrate potential for clinical applications^19[,20.

Fig. 1.

Percentage of bariatric surgery questions in each domain.

Figure 2 illustrates the overall study design. Initially, a general prompt, “I have some questions about bariatric surgery,” was used to set the context. The 50 selected questions were input into each LLM as standalone queries, generating a total of 200 responses for analysis. After each query, the conversation was reset to minimize memory retention bias. To prevent graders from identifying the LLMs, all the responses were formatted into plain text, with chatbot-specific features concealed. The responses (Supplementary Table 1) were then randomly shuffled before being presented to the multidisciplinary expert team for grading. Grading was conducted over three separate rounds, each with a 72-hour washout period, to minimize carryover effects.

Fig. 2.

Flowchart of overall study design.

Expert consultation and scoring process

During the expert consultation phase, our questionnaire included a dedicated section for the experts to raise concerns or questions regarding the questions and answers. This section ensures transparency and allows for direct resolution of any issues. During this period, only one expert raised a concern, questioning the appropriateness of using LLMs for medication selection in clinical contexts. Our team promptly clarified that the study was a research exercise, not actual clinical practice. The expert accepted our response and raised no further concerns. Importantly, the grading process was conducted independently by each expert without discussion, ensuring objectivity and unbiased results.

Accuracy evaluation

We assembled a multidisciplinary team of seven specialists with a minimum of ten years of clinical experience (18.57 ± 7.87 years). These specialists were selected based on the core organizational structure recommended for bariatric surgery^17[,22, which includes four metabolic and bariatric surgeons (L*, W*, G*, P**), one obesity physician (S**), and two dietitians (H**, F**). The academic qualifications of the team are as follows: one member holds a bachelor’s degree, five members have master’s degrees, and two hold doctoral degrees. Professionally, the team includes three attending physicians, three associate chief physicians, and one chief physician. Their collective expertise ensured a thorough and rigorous evaluation of the LLMs’ performance in this clinical context. To maintain objectivity, the identities of the LLM chatbots were concealed from the graders. The graders’ primary responsibility was to independently assess the accuracy of each LLM response using a three-point scale: (1) “Poor” for responses with significant inaccuracies that could mislead patients and potentially cause harm. Example: A response that incorrectly states that bariatric surgery has no risks or provides advice that contradicts established medical guidelines; (2) “Borderline” for responses containing potential errors but unlikely to mislead or harm patients. Example: A response that mentions most but not all common risks associated with bariatric surgery or provides correct information but lacks detail; and (3) “Good” for error-free responses. Example: A response that accurately lists all common risks associated with bariatric surgery, provides detailed explanations, and includes recommendations for managing these risks. The total accuracy score for each LLM response was calculated by summing the scores from all seven graders (Supplementary Table 2).

We also employed a majority consensus approach to determine the final rating for each response based on the most frequent grade assigned by the seven graders. In cases where no consensus was reached (e.g., each grader assigned a different rating), we applied a conservative approach, assigning the lowest rating (“poor”) to the response.

Comprehensiveness evaluation

For chatbot responses rated “good” by majority consensus, the graders performed an additional evaluation to assess comprehensiveness. Comprehensiveness refers to the extent to which the response covers all key clinical aspects of the topic in sufficient depth, including procedure details, risks/complications, patient selection criteria, management strategies, and evidence-based guidelines, without major omissions.” Responses were rated on a five‑point scale: 1 = not comprehensive (major omissions of essential elements); 2 = slightly comprehensive (only a minority of key elements with limited detail); 3 = moderately comprehensive (about half of the key elements with reasonable detail); 4 = comprehensive (most key elements adequately covered in clinical context); 5 = very comprehensive (all or nearly all key elements covered in depth and consistent with guidelines. The mean comprehensiveness score was calculated by averaging the scores from all the graders for the responses rated “good.”

The selection of accuracy and comprehensiveness as primary metrics was informed by established medical communication frameworks and previous studies²²²³. Accuracy assessment is fundamental in medical contexts to ensure patient safety and prevent misinformation. A comprehensive evaluation reflects the need for complete patient education that facilitates informed decision-making. This two-dimensional approach aligns with standard clinical practice, where healthcare providers must deliver both accurate and thorough information to patients.

Re-evaluation of accuracy for LLMs’ self-corrected responses

For responses initially rated as ‘poor’, we explore the self-correction behavior of LLMs via three distinct prompting approaches, resulting in small and uneven sample sizes. No inferential statistics were planned, and the standardized prompts were not intended to reflect real patient behavior. Each approach was re‑evaluated for accuracy by the same expert panel after a two‑week washout, using the same criteria and consensus protocol. The three prompting approaches were as follows: (1) Basic review without web access: “That does not seem quite right. Could you kindly review?” (2) Web-enabled review: “That does not seem quite right. Could you kindly review? Please look up the web.” (3) Evidence-based review: “That does not seem quite right. Could you kindly review? Please look up the web for evidence-based information.”

Detailed qualitative analysis of poorly rated LLMs’ responses

To investigate the limitations and risks of relying solely on LLM chatbot responses for bariatric surgery information, we conducted a detailed analysis of responses rated “poor” by at least four graders. An assigned expert (Z**) carefully identified and explained inaccuracies within these responses.

Statistical analysis

Statistical analyses were performed via IBM SPSS Statistics for Windows, version 25.0 (Armonk, NY: IBM Corp.). One-way ANOVA and Tukey’s Honest Significant Difference test were used to compare character counts across responses from the four LLMs, as the data met parametric assumptions. The Kruskal-Walli’s rank sum test and Dunn’s multiple comparisons were used to compare word counts, accuracy scores, and comprehensiveness scores among the LLMs. A two-tailed Pearson’s χ² test was conducted to compare the proportions of “good,” “borderline,” and “poor” ratings across the LLMs.

When multiple hypotheses were tested, p values were adjusted via the Bonferroni correction, with statistical significance set at p < 0.05.

Results

Inter-rater agreement analysis and response length

The inter‑rater agreement was as follows: accuracy rating, Fleiss’ kappa = 0.210 (95% CI: 0.208–0.212; Z = 7.815; P < 0.001); comprehensiveness rating, Fleiss’ kappa = 0.464 (95% CI: 0.453–0.476; Z = 2.543; P < 0.011). Response lengths varied among LLMs, with Claude-3 Opus producing the most verbose responses (105.36 ± 96.69 words; 610.7 ± 554.05 characters), followed by Gemini Pro (92.08 ± 121.49 words; 516.34 ± 709.54 characters), ChatGPT-4 (84.9 ± 73.79 words; 486.42 ± 421.17 characters), and ERNIE Bot 4.0 (75.96 ± 17.22 words; 431.88 ± 99.23 characters) (Table 1).

Table 1.

Overview of response length from LLM-Chatbots to Bariatric Surgery-Related Questions.

LLM	Response length (words)			Response length (characters)
	Mean (SD)	Minimum	Maximum	Mean (SD)	Minimum	Maximum
ERNIE Bot 4.0	75.96 ± 17.22	34	128	431.88 ± 99.23	191	713
ChatGPT 4.0	84.9 ± 73.79	34	426	486.42 ± 421.17	191	2523
Claude-3 Opus	105.36 ± 96.69	36	438	610.7 ± 554.05	247	2532
Gemini Pro	92.08 ± 121.49	21	645	516.34 ± 709.54	115	3695

Overall accuracy performance and accuracy across domains

Figure 3 shows the average accuracy scores of the LLMs’ responses to the bariatric surgery-related questions. ERNIE Bot 4.0 achieved the highest average total accuracy score of 19.46 ± 2.07. This score surpassed those of ChatGPT-4 (18.98 ± 1.67; p = 0.602), Claude-3 (18.78 ± 1.87; p = 0.238), and Gemini Pro (16.52 ± 2.79; p = 0.000). Supplementary Table 2 presents the total accuracy scores for each LLM-Chatbot’s response to each question.

Fig. 3.

LLMs performance: Total Average Scores and Comprehensiveness of Responses With ‘Good’ Accuracy.

Figure 4 shows the consensus-based accuracy ratings for the four LLMs. ChatGPT-4 received “good” ratings for 90.0% of its responses, which were significantly higher than those of ERNIE Bot 4.0 (84.0%), Claude 3 Opus (80%), and Gemini Pro (48.0%; all p ≤ 0.05). Notably, ChatGPT-4 did not receive any ‘poor’ ratings, whereas ERNIE Bot 4.0 and Gemini Pro each had 4% “poor” ratings, and Claude 3 Opus had 14%. Supplementary Table 2 provides detailed ratings for each LLM’s response.

Fig. 4.

Consensus-based accuracy ratings of four LLMs responses to bariatric surgery-related questions.

Table 2 presents a detailed sub-analysis of the consensus-based accuracy ratings across the five bariatric surgery domains. Three LLMs—ERNIE Bot 4.0, ChatGPT 4.0, and Claude 3 Opus—achieved 100% “good” ratings in the “lifestyle changes” and “others” domains. However, Gemini Pro received 57% (Χ²=10.080, p = 0.121) and 0% (Χ²=16.000, p = 0.014) of these domains, respectively. In the “preoperative preparation” domain, ChatGPT 4.0 and Claude 3 Opus achieved 100% “good” ratings, whereas ERNIE Bot 4.0 and Gemini Pro received 1 and 2 ‘borderline’ ratings, respectively (Χ²=4.513, p = 0.211). For the “eligibility, efficacy, and procedure options” domain, ChatGPT 4.0 led with 88% “good” ratings, followed by ERNIE Bot 4.0 at 82%, Claude 3 Opus at 71%, and Gemini Pro at 59% (Χ²=6.618, p = 0.358). In the ‘recovery, risks, and complications’ domain, all LLMs performed worse, with Gemini Pro achieving only a 44% “good” rating, lower than ChatGPT 4.0 (83%), ERNIE Bot 4.0 (78%), and Claude 3 Opus (72%; Χ²=10.987, p ≥ 0.05).

Table 2.

Consensus-based accuracy ratings of LLM-Chatbot responses across five bariatric surgical care domains.

Domain LLM	Eligibility, efficacy, and procedure options(n = 17)	Preoperative preparation (n = 4)	Recovery, risks, and complications (n = 18)	Lifestyle changes (n = 7)	Others (n = 4)
ERNIE Bot 4.0
Poor	0(0)	0(0)	2(11%)	0(0)	0(0)
Borderline	3(18%)	1(25%)	2(11%)	0(0)	0(0)
Good	14(82%)	3(75%)	14(78%)	7(100%)	4(100%)
Chat GPT 4.0
Poor	0(0)	0(0)	0(0)	0(0)	0(0)
Borderline	2(12%)	0(0)	3(17%)	0(0)	0(0)
Good	15(88%)	4(100%)	15(83%)	7(100%)	4(100%)
Claude-3 Opus
Poor	2(12%)	0(0)	0(0)	0(0)	0(0)
Borderline	3(18%)	0(0)	5(28%)	0(0)	0(0)
Good	12(71%)	4(100%)	13(72%)	7(100%)	4(100%)
Gemini Pro
Poor	2(12%)	0(0)	2(11%)	2(29%)	1(25%)
Borderline	5(29%)	2(50%)	8(44%)	1(14%)	3(75%)
Good	10(59%)	2(50%)	8(44%)	4(57%)	0(0)
Χ2	6.618	4.513	10.987	10.080	16.000
P value	0.358	0.211	0.089	0.121	0.014

Comprehensiveness performance

Figure 3 summarizes the comprehensiveness scores for “good” rated responses. All four LLMs performed sub-optimally, but Claude 3 Opus (3.13 ± 0.46) and Gemini Pro (3.06 ± 0.44) outperformed ChatGPT 4.0 (2.86 ± 0.41) and ERNIE Bot 4.0 (2.91 ± 0.48). A comparison of comprehensiveness scores across common questions (Table 3) revealed similar performance across all four LLMs, with no significant difference (F = 2.342, p = 0.080).

Table 3.

Comprehensiveness assessment for common questions answered by the four LLM-Chatbots, with responses that received ‘good’ accuracy rating.

LLM	n	Mean (SD)	Median
ERNIE Bot 4.0	19	2.83(0.42)	2.71
Chat GPT 4.0	19	2.78(0.37)	2.71
Claude-3 Opus	19	3.1(0.46)	3.14
Gemini Pro	19	3(0.43)	3.00

Self-correction performance and qualitative analysis

Table 4 highlights all LLMs demonstrated notable self-correction, except for Gemini Pro’s Question 45, which retained a consensus-based “poor” rating. ERNIE Bot 4.0 and Claude 3 Opus achieved a perfect 100% improvement rate (2 of 2 responses), while Gemini Pro showed an improvement rate of 85.71% (6 of 7 responses). Notably, the accuracy of responses from ERNIE Bot 4.0 and Claude 3 Opus improved from “poor” to “good” following self-correction. Supplementary Tables 3a-c offer a detailed comparison of original and self-corrected responses for each LLM. When prompted to use Internet searches and evidence-based resources, ERNIE Bot 4.0 achieved three “Good” and one “Borderline” rating, while Claude 3 Opus maintained a perfect 100% improvement rate (4 of 4 responses). In contrast, Gemini Pro’s four responses retained “poor” ratings (4 of 14 responses).

Table 4.

Demonstration of three LLMs’ ability to self-correct when prompted and self-correction capabilities using the Internet search.

Topic	Question	Summed score(max:21)				Consensus-based rating
		Initial	Self- Corrected	Internet Search 1†	Internet Search 2‡	Initial	Self-Corrected	Internet Search 1†	Internet Search 2‡
ERNIE Bot 4.0
Recovery, risks, and complications	36.Why is acid reflux worse after bariatric surgery?	16	21	20	18	Poor a	Good	Good	Good
	39. Does bariatric surgery increase risk of hernias?	12	19	15	16	Poor a	Good	Borderline	Good
Claude-3 Opus
Eligibility, efficacy, and procedure options	7. Will bariatric surgery lower blood pressure?	16	18	17	16	Poor a	Good	Good	Good
	15. Will bariatric surgery cure sleep apnea?	16	21	18	15	Poor a	Good	Good	Good
Gemini Pro
Eligibility, efficacy, and procedure options	9. Is bariatric surgery necessary or can weight be lost with only a restrictive diet?	8	17	16	16	Poor	Borderline	Good	Good
	15. Will bariatric surgery cure sleep apnea?	16	18	17	17	Poor a	Good	Good	Good
Recovery, risks, and complications	28. How often do I need to see the doctor during the first year after bariatric surgery?	10	18	11	11	Poor	Good	Poor	Poor
	37. Does bariatric surgery cause anemia?	16	20	16	16	Poor a	Good	Poor a	Poor a
Lifestyle changes	45. How long do I take multivitamins after bariatric surgery?	10	13	17	17	Poor	Poor	Good	Good
	46. Is it recommended to start personal training months after bariatric surgery?	12	18	20	20	Poor a	Good	Good	Good
Others	48. Will the baby be healthy if the mother gets bariatric surgery?	13	20	19	19	Poor	Good	Good	Good

^a Where consensus on final accuracy rating was not reached (i.e., each grader provided a different rating), the lowest score (‘poor’) was assigned.

†Refers to responses generated by three LLMs during web searches.

‡Refers to evidence-based responses generated by three LLMs from web searches.

Only Gemini Pro’s responses were rated “poor” by at least four graders and therefore received a detailed analysis by a metabolic and bariatric surgeon (Z**) (Supplementary Table 4).

Discussion

To the best of our knowledge, this is the first study to perform a multidisciplinary expert evaluation of four leading LLMs—ERNIE Bot 4.0, ChatGPT-4, Claude-3 Opus, and Gemini Pro—in addressing frequently asked bariatric surgery-related queries. Because obesity management and postoperative recovery require coordinated, long‑term care, a multidisciplinary team (MDT) approach is recommended by systematic reviews and clinical guidelines^21[,24[,25, A MDT comprising surgeons, obesity medicine physicians, dietitians, and nurse specialists provides comprehensive healthcare. In contrast, previous studies evaluating LLMs’ responses to common weight‑loss surgery questions have generally relied on a single‑discipline perspective and have focused primarily on accuracy, reproducibility, or information comprehensiveness⁹¹³. Furthermore, these studies have largely neglected to evaluate LLMs’ self-correction abilities, especially when the internet is connected—a now standard feature in mainstream LLMs²². Additionally, prior research often lacks blinding and randomization⁹¹³, potentially introducing biases. The study employs a robust design that conceals LLM-specific characteristics in responses, randomly shuffles responses before they are reviewed, and includes a washout period between evaluation sessions²². These design elements can improve the reliability and validity of human evaluation studies, thereby enabling more rigorous assessments of LLM performance²³. Notably, this study delivers a focused China–international comparison of ERNIE Bot against three global models. Together, the strict blinding and randomization procedures, cross professional expert review, and assessment of internet enabled correction form a robust methodological and empirical foundation for standardized evaluation, although further validation in real-world clinical settings is required before any consideration of deployment.

The variations in accuracy observed across different LLMs (accuracy range from 48% to 90.0%) raise potential safety concerns regarding direct patient education applications. While these models may provide useful insights, the disparity in their performance highlights the need for cautious interpretation and appropriate clinical oversight when translating model outputs into patient-facing recommendations²⁶. Compared with the other three LLMs, ERNIE Bot 4.0 showed the highest numerical mean accuracy score (19.46 ± 2.07), although the differences compared with ChatGPT‑4 and Claude‑3 Opus were not statistically significant. This result indicates that Chinese AI systems that ERNIE Bot 4.0 can perform competitively in English-based evaluations²⁷. Another notable finding emerged in the consensus-based accuracy ratings, where ChatGPT‑4 showed a higher proportion of ‘good’ ratings in the consensus evaluation. Although this reflects a favorable descriptive pattern, the interpretation of its underlying causes should remain cautious. The absence of poor ratings may be relevant for clinical contexts, although further studies are required to determine real-world implications, as it indicates a consistent level of reliability and minimal risk of providing misleading medical information. This high degree of accuracy and consistency is crucial for clinical applications where misinformation can have serious consequences for patient care and safety. These descriptive patterns suggest that ChatGPT‑4 may hold potential for certain clinical-support applications, pending further validation in real-world settings, particularly in areas such as patient education, preoperative counseling, and postoperative follow-up care. Our findings echo those of prior studies by Chen and Ozcivelek, who reported relatively stronger performance of ChatGPT‑4.0 over other LLM counterparts in dental and total knee arthroplasty, respectively^28[,29.

The domain-specific analysis revealed strong performance from ChatGPT-4, ERNIE Bot 4.0, and the Claude-3 Opus in the “lifestyle changes” and “others” categories (100%), while Gemini Pro’s performance was substantially lower in these areas (lifestyle changes: 57%; others: 0%, χ2 = 16.000, P < 0.05). In the “others” domain (questions 47–50, Supplementary Table 5a), Gemini Pro’s responses demonstrated significant limitations that have several practical implications. First, its oversimplified and generic medical advice, without adequate attention to individual patient factors, may lead to inappropriate post-surgical care decisions and increased complication risks. Second, the model’s inability to provide comprehensive guidance may result in unnecessary clinical visits as patients seek clarification, increasing the burden on healthcare resources. Third, the absence of multidisciplinary perspectives could disrupt the integrated care pathway between surgical teams and other healthcare providers. Fourth, its inadequate risk assessment capabilities could lead to patients underestimating post-surgical complications, potentially reducing compliance rates with recovery protocols. These descriptive limitations suggest that Gemini Pro may require additional human oversight in bariatric-surgery contexts. The remaining domains showed numerically lower ‘good’ ratings, although these differences were not statistically significant (44% to 59%, P > 0.05), so interpretations should remain descriptive. Gemini Pro showed lower descriptive ratings in several domains, though most differences were not statistically significant.

Interestingly, all LLMs showed relative weakness in the ‘recovery, risks, and complications’ domain, suggesting that this area might represent a particular challenge for current LLM technology, possibly owing to the complexity and nuanced nature of bariatric surgery post-surgical complications. Our findings differ from Aghamaliyev’ s study³⁰, which reported that GPT4.0 performed worse on preoperative questions for general abdominal surgery because of its complexity, which involves blood transfusions, nasogastric tubes, and surgical complications. In bariatric surgery, postoperative management presents unique challenges, including immediate recovery issues, long-term lifestyle modifications, and specialized complication management requiring multidisciplinary care. Unlike preoperative considerations that occur within a defined timeframe, postoperative bariatric care demands lifelong attention to nutrition, exercise, complication prevention, and lifestyle maintenance. These contrasting results suggest that LLM performance may indicate possible variation across domains, although these exploratory findings require further study. This domain-specific variation emphasizes the importance of domain-based performance assessment in LLMs, as aggregate accuracy metrics may obscure critical differences in capability across different aspects of patient care.

We conducted a deep analysis to investigate why ERNIE Bot 4.0 appeared less detailed in qualitative examples compared with ChatGPT‑4 and Claude‑3, although these observations are descriptive rather than statistically tested (Supplementary Table 5b). (1) Regarding the type of bariatric surgery (Question 5. What are the different types of bariatric surgery? ERNIE Bot 4.0 (Borderline) excels at providing concise medical topic overviews but consistently lacks the clinical depth demonstrated by ChatGPT-4.0 (Good) and Claude-3 Opus (Good), particularly in complex scenarios such as bariatric surgery consultation. This emphasis on brevity over comprehensiveness results in ERNIE Bot 4.0’s tendency to list procedures without addressing crucial clinical considerations, limiting its utility for medical decision support. ChatGPT-4.0 achieves a balance between technical accuracy and clinical relevance, incorporating specific parameters and current trends while maintaining accessibility, despite occasional verbosity. (2) Regarding the weight loss timeline (Question 10. How much weight is normally lost in the first month after sleeve gastrectomy? However, ERNIE Bot 4.0’s(Borderline) response provides an understanding of the multifactorial nature of post-surgical weight loss. Its core prediction of 10–20% total body weight loss within the first month represents a significant overestimation that could establish unrealistic patient expectations and compromise clinical outcomes^31[,32. ChatGPT-4.0 (Good) provides more reasonable predictions of 15–25 pounds (6.8–11.3 kg) of first-month weight loss, with appropriate consideration of individual variations³³. This approach better reflects typical clinical outcomes and offers a more practical framework for patient guidance. (3) Regarding psychological support (Question 21. How important is it to have family support before undergoing bariatric surgery? ) ChatGPT-4.0 and the Claude-3 Opus demonstrated superior performance, achieving “Good” ratings by providing comprehensive explanations that encompassed both physical and emotional support aspects, along with specific examples of family involvement throughout the surgical journey. In contrast, ERNIE Bot 4.0’s response, while fundamentally accurate, received a “Borderline” rating due to its brevity and lack of detailed guidance on practical support mechanisms. Notably, all three LLMs maintained clinical accuracy, though they differed significantly in their depth of explanation and practical applicability. These findings suggest that while ERNIE Bot 4.0 maintains basic clinical accuracy and provides concise medical topic overviews, its emphasis on brevity may reduce the depth of clinical detail in certain contexts. The system requires substantial enhancement of its medical knowledge base and improvement in delivering nuanced, clinically relevant responses that better align with established medical guidelines and practical patient care requirements. Despite avoiding harmful misinformation, the differences in depth of explanation and practical applicability compared with those of ChatGPT-4.0 and the Claude-3 Opus indicate areas for necessary improvement in ERNIE Bot 4.0’s capabilities.

In terms of comprehensiveness, both Ernie Bot 4.0 and ChatGPT-4 fail to meet the acceptable threshold of three points, underperforming relative to Claude 3 Opus and Gemini Pro. Our detailed analysis of selected questions reveals notable performance differences across LLMs (supplementary table 5c). (1) Concerning clinical knowledge depth (Question 4: Medication management post-surgery), Claude-3 Opus demonstrated greater comprehensiveness (3.86/5) by providing the most comprehensive guidance by systematically addressing six medication categories with specific examples, mechanistic explanations, and a focus on individualized clinical management. Gemini Pro (3.00/5) demonstrated moderate comprehensiveness, covering four medication categories and key monitoring points, but lacked the depth and specificity evident in the Claude-3 Opus’ response. In contrast, ChatGPT-4 (2.71/5) and Ernie Bot 4.0 (2.43/5) provided general recommendations without addressing specific medication classes or underlying pharmacological principles. (2) Regarding the clinical decision-making structure (Question 6: surgical revision for GERD), Claude-3 Opus (4.14/5) provided a structured evaluation approach including systematic assessment criteria, surgical indications, and multidisciplinary considerations. ChatGPT-4 (3.14/5) offered relevant but fragmented information without a clear decision-making framework. (3) Regarding vitamin deficiency and surgical complications (Question 26: Do bariatric surgery patients have serious health problems caused by vitamin deficiency? ) The ChatGPT-4.0 score was 2.43/5, indicating that it covered the post-surgery care but still lacked detailed exercise and dietary guidelines. ERNIE Bot 4.0 received a score of 2.71/5, reflecting its basic lifestyle recommendations but lacking specifics, particularly in terms of portion control, psychological support, and concrete exercise or dietary guidelines. Claude-3 Opus and Gemini Pro both achieved scores of 3.43/5. Claude-3 Opus provided specific, actionable recommendations, quantified exercise guidelines, and a focus on both physical and psychological aspects, including dietary specifics, portion control, and support systems. Gemini Pro addressed broader lifestyle factors such as sleep and stress but lacked quantifiable guidelines and a more structured presentation. Based on the analysis, the Claude‑3 Opus provided more structured or detailed responses in selected qualitative examples through comprehensive evaluation frameworks, clear surgical indications, and integrated multidisciplinary considerations. Additionally, it provides exceptional postoperative care guidance with specific, actionable recommendations spanning exercise, dietary specifications, and psychological support frameworks.

Our detailed qualitative analysis revealed several critical instances where Gemini provided inaccurate, oversimplified or unindividualized information that was graded as “poor” responses by experts regarding bariatric surgery management (Supplementary Table 4). These findings highlight potential concerns about the quality and precision of medical information dissemination, although these observations come from a subset of poor-rated responses. Gemini pro’s response to the question regarding the necessity of bariatric surgery versus weight loss through diet al.one is overly simplistic (Question 9, Eligibility, efficacy, and procedure options domain). It fails to address the complexities of long-term weight management and the specific eligibility criteria for bariatric surgery. This oversimplification could mislead patients regarding the necessity and effectiveness of bariatric surgery, highlighting the need for more nuanced and detailed responses. The response does not adequately address the importance of postoperative lifestyle changes and the specific eligibility criteria for bariatric surgery, which are critical for informed decision-making. Regarding post-operative care protocols (Question 28, “Eligibility, efficacy, and procedure options” domain), the imprecise follow-up schedule recommendation of “every 2–4 weeks” could compromise patient safety. In China, bariatric and metabolic centers implement personalized follow-up schedules tailored to patients’ specific conditions. After bariatric surgery, the follow-up schedule typically includes an initial visit at 3–4 weeks post-operation, followed by subsequent visits at 1–3-month intervals throughout the first year. While some follow-ups can be conducted through online consultations with nutritionists, it is important to note that different bariatric and metabolic centers customize their follow-up schedules based on individual patient needs rather than adhering to a one-size-fits-all timeline. The lack of personalized guidance could result in delayed identification of complications, potentially leading to severe outcomes. The response about nutritional supplementation (Question 45, “Lifestyle changes” domain) was generalized. The blanket recommendation for lifelong supplementation fails to recognize procedure-specific requirements. While patients are generally advised to take multivitamins after bariatric surgery owing to the risk of nutrient deficiencies, the duration and necessity of supplementation can vary depending on the type of surgery. According to the guidelines³⁵, it is recommended that patients continue supplementation with multivitamins after bariatric surgery and undergo lifelong monitoring of nutritional status. Malabsorptive procedures, such as Roux-en-Y gastric bypass (RYGB), one anastomosis gastric bypass (OAGB), and biliopancreatic diversion with a duodenal switch (BPD/DS), generally require lifelong multivitamin supplementation, whereas restrictive procedures, such as sleeve gastrectomy (SG) and laparoscopic adjustable gastric banding (LAGB), require supplementation for at least one year.

Despite these performance limitations, all LLMs show substantial self-correction abilities when prompted. Ernie Bot 4.0 and Claude 3 Opus achieved a perfect improvement rate of 100% (2 out of 2 responses), whereas Gemini Pro demonstrated an 85.71% improvement rate (6 out of 7 responses). To enhance LLM capabilities, the predominant strategies have focused on two fronts: vertical scaling, which involves expanding their size, data volume, and computational resources; and horizontal customization, which encompasses postfiltering, fine-tuning, and the incorporation of human feedback³⁵. While human feedback is valuable, it carries risks, as it depends heavily on the veracity, expertise, and potential biases of the users. These factors can introduce variability and reduce the reliability of the model’s outputs³⁶. Moreover, the absence of an automated disclaimer in LLM chatbots poses a significant concern, especially when some responses remain suboptimal despite self-correction. As shown in Supplementary Tables 3a–c, this suggests that, even with self-correction mechanisms, LLM chatbots may still propagate erroneous medical information without disclosing the uncertainty of their outputs¹⁴. A false response from GPT-4, for example, can be referred to as “hallucination,” which is particularly dangerous in medical contexts. The term “hallucination” refers to an LLM generating plausible and often confident statements that are factually incorrect in the sense of not being grounded in the data³⁷. These errors, which are often subtle and presented convincingly, may mislead users into believing incorrect information³⁹. Case reports have highlighted instances where ChatGPT has caused significant, potentially life-threatening harm to patients³⁹. To mitigate these risks, future research should focus on improving the transparency and accountability of LLM chatbots. Possible solutions include developing more sophisticated self-correction algorithms to better identify and flag uncertain responses, as well as implementing an automated disclaimer system to alert users about the limitations of the information provided. These measures enhance the reliability of LLM chatbots and encourage a more cautious approach to their use, particularly in sensitive fields such as healthcare⁴.

When tasked with utilizing Internet search and evidence-based resources (Supplementary Tables 6a-c), the Claude 3 Opus achieved a perfect 100% improvement rate (4 of 4 responses), whereas ERNIE Bot 4.0 produced three “Good” and one “Borderline” rating. Notably, Gemini Pro’s four responses retained a “Poor” rating (4 of 14 responses). To understand these disparities, we compared the responses of Ernie Bot 4.0 and Gemini Pro, which incorporate Internet search responses, as shown in supplementary Table 7. For example, ERNIE Bot 4.0’s response demonstrated better scientific accuracy, comprehensiveness, personalization, and clinical relevance. (Question 36. Why is acid reflux worse after bariatric surgery? ) showed comprehensive evaluation of post-bariatric acid reflux mechanisms and remarkable scientific accuracy in explaining anatomical and physiological changes and their clinical implications. The analysis effectively covers six major reflux mechanisms while establishing clear links between surgical modifications and potential complications. Patient safety and personalized care are highlighted through detailed discussions of medical consultation and management approaches, with the recognition of individual symptom variations. The clinical relevance is enhanced by incorporating practical strategies and long-term considerations, presenting a balanced integration of technical information and patient guidance that follows current medical guidelines. On the other hand, Gemini Pro’s response was sub-optimal. (Question 28. How often do I need to see the doctor during the first year after bariatric surgery? ) In China, different bariatric and metabolic centers implement personalized follow-up schedules tailored to patients’ specific conditions. While a three-month follow-up interval may be appropriate for most patients, those with more severe comorbidities—such as serious diabetes, obstructive sleep apnea hypopnea syndrome (OSAHS) with respiratory failure, or uncontrolled hypertension—require individualized follow-up schedules. This oversight demonstrates insufficient consideration of special patient populations in this response, potentially compromising patient care.

Although connectivity features are expected to enhance LM performance⁴¹, this study presents a more nuanced view. The results indicate that mere access to internet resources does not guarantee improved performance^41[,42. However, our findings suggest that while internet-connected LLMs outperform their non-connected counterparts, they still fall short of the self-correction abilities demonstrated by LLMs guided by prompt-based self-correction showed more consistent improvement than internet-based correction in this study sample (Table 4)¹⁸. Our research findings align with previous test results in the field of ophthalmology regarding myopia, indicating that the effectiveness of internet-based self-correction in improving accuracy depends on multiple factors: model architecture, domain specificity, query complexity, and the quality of online resources²². These insights suggest that future development should focus on three key directions. The optimization of internet search integration mechanisms will enhance information filtering and synthesis capabilities, while the development of specialized medical knowledge frameworks will incorporate structured clinical guidelines and protocols. The enhancement of personalization capabilities will enable the processing of diverse patient profiles for tailored recommendations. Furthermore, systematic analysis of error patterns between connected and non-connected models could provide valuable insights for developing improved hybrid architectures that maximize the advantages of both approaches.

In evidence-based medicine (EBM), the retrieval and ranking of relevant evidence based on epistemological strength is essential to inform clinical practice⁴¹. Our study revealed that the resources cited by the evaluated LLMs predominantly came from open medical websites lacking explicit EBM methodologies, raising concerns about their understanding of EBM principles (Supplementary Tables 6a-c), highlighting limitations in integrating high-quality evidence. This pattern raises the possibility that users may incorrectly assume they’re accessing rigorously peer-reviewed information when consulting variable-quality web resources, potentially affecting medical decision-making and patient safety.

Despite the medical community’s push for better EBM integration into LLMs, several challenges hinder their ability to acquire EBM-compliant evidence⁴¹. These include the closed nature of medical databases, specific data training and update cycles, the lack of domain-specific logical reasoning in EBM, and unverified training data. Moreover, some companies show little interest in applying EBM principles to the medical field, which exacerbates these issues. To address these challenges, we propose medical databases should be made accessible to allow real-time internet updates, ensuring that LLMs have access to the latest and most relevant evidence while implementing HIPAA-compliant security measures including data anonymization and strict access controls⁴³. Various adaptation methods can tailor pre-trained LLMs to specific medical domains, including fine-tuning with task-specific data, instruction fine-tuning, reinforcement learning from human feedback, and prompt engineering⁴⁴. Research shows that fine-tuned models can match or outperform larger zero-shot models in medical evidence summarization⁴⁵. Development of vertical language models targeting specific medical issues offers another promising approach. Due to surgery’s complexity and high-risk nature, LLM applications in this field remain nascent, though insights from other medical disciplines may inform future surgical applications⁴⁶.

In this study, we observed heterogeneity in performance metrics among the four LLMs evaluated for bariatric surgery care applications. The accuracy rates demonstrated considerable variation (48.0%~90.0%), whereas the comprehensiveness scores showed a narrower range (2.86 ± 0.41 ~ 3.13 ± 0.46). Furthermore, these findings suggest that current LLMs are not yet suitable for independent counseling. They may have potential future roles as supportive tools, which would require additional validation^13[,23. Our analysis identified several potential application areas where LLMs might support bariatric care under supervision. In the preoperative phase, these systems can be integrated into clinical workflows through⁴⁷ (1) automated initial screening questionnaires with real-time analysis; (2) interactive educational modules explaining surgical procedures with branching dialog paths based on patient comprehension; and (3) scheduled reminder systems for preoperative preparation steps. For example, LLMs can handle routine inquiries about surgical procedures and lifestyle modifications while flagging complex cases for human expert review. The postoperative care pathway presents another viable application domain, where LLMs support structured follow-up protocols through³(1) systematic screening of post-surgical recovery markers via daily symptom monitoring; (2) personalized guidance on nutrition and physical activity adherence with automated progress tracking; (3) preliminary psychological support assessment through structured chatbot interactions; and (4) early detection of complications through automated pattern analysis with healthcare providers alerts for timely intervention.

Limitation

The evaluation of LLMs in healthcare has gained significant attention in medical research, yet this field remains nascent and is characterized by considerable variability in study designs, evaluation metrics, and assessor expertise⁴⁴. This lack of standardization limits the comparability and generalizability of findings across models, highlighting the need for more rigorous evaluation methodologies. A recent human evaluation framework for LLMs in healthcare²³, derived from a literature review, presents valuable insights but has not yet been empirically tested. Our study adheres to several of the framework’s principles, such as evaluating essential evaluation dimensions (accuracy and comprehensiveness), conducting blinded assessments, involving a multidisciplinary expert panel, and applying data analysis to compare outcomes, ensuring rigorous scientific standards.

While our methodological approach has strengths, several limitations warrant careful consideration. First, the relatively low inter-rater agreement (Fleiss’ kappa = 0.210 and 0.464) reflects both the complexity of multidisciplinary surgical judgment and the absence of pre-study calibration. To address this, future studies should incorporate structured calibration sessions and consider multiple rating rounds to improve reliability. Second, our conservative adjudication rule—assigning the lowest rating when consensus was not reached—prioritized clinical safety but may have inflated “poor” ratings and compressed borderline cases. Future evaluations could adopt alternative adjudication strategies, such as majority voting or median scoring, to better capture borderline performance while maintaining safety standards. Third, although our query set integrated sources from literature, Chinese social-media platforms, and patient inquiries, the generalizability of our findings is limited by the FAQ‑style, single‑turn query format used in this study. While suitable for evaluating educational or informational chatbot functions, this design does not extend to more complex real‑world applications such as longitudinal counseling, personalized risk assessment, shared decision‑making, or postoperative management. Future work should clearly delineate the intended use case of such models, incorporate multi‑turn and context‑rich scenarios, and diverse question sources to enhance cross‑cultural validity and support more generalizable clinical deployment. Fourth, Responses were generated in English and then translated into Chinese for expert evaluation. This translation process could introduce nuances or biases that affect the assessment of response quality and accuracy. Future studies should consider conducting assessments directly in the original language of the model outputs or validating the translation process to improve transparency and reliability. Finally, while LLMs show promise in supporting medical decision-making, their susceptibility to hallucination, underlying data and algorithmic biases, and limited transparency pose important safety risks. Continued monitoring, model refinement, and evaluation of newer or domain-specific LLMs—especially those trained on bariatric datasets—will be essential to mitigate these risks and prevent misinformation, inappropriate decisions, or inequities in clinical care.

Conclusions

In conclusion, the evaluation of LLMs remains in its early stages, with existing studies employing a variety of methods, metrics, and processes. Our study, leveraging a multidisciplinary expert panel, blinding, and randomization, offers a more scientific and comprehensive assessment of four mainstream models’ performance in addressing common questions from bariatric surgery patients. The framework effectively captures significant variations in response quality, depth, and organization, enhancing the robustness of our assessment methodology. The evaluated LLMs showed variability in accuracy and fell short of ideal comprehensiveness scores. Notably, their performance was suboptimal in complex domain (recovery, risks, and complications). Additionally, the models’ ability to improve through online searches and evidence-based resources was less effective than their self-correction capabilities. Future research should focus on the development and validation of LLMs specifically tailored for patient health education in bariatric surgery. This endeavor requires the collection of real-world consultation needs from weight loss patients, utilizing reliable, specialized evidence as training data. Furthermore, continuous involvement of interdisciplinary experts throughout the development, evaluation, and implementation processes is essential to enhance the safety and efficacy of LLMs in patient education applications.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

J.S.C. conceived and designed the study. J.S.C., J.H.C., and T.T.Y. performed the experiments and collected data. L.F.H., L.L.C., and H.O.Q. analyzed and interpreted the data. X.J.Z. contributed materials tools. J.S.C. wrote the main manuscript text. P.Y.H., W.H.Y. and X.L.H. supervised the project and revised the manuscript critically for important intellectual content. All authors reviewed and approved the final manuscript.

Funding

Pengyu Huang received financial support from health and young backbone project of Fujian province (NO.2025QNGGA010) and the Joint Funds for the innovation of science and Technology Fujian province (Grant number:2023Y9384). Jianshu Cai received financial support from the Zhejiang medical and health project (Grant number:2023KY781). Jionghuang Chen received financial support from Zhejiang Provincial Natural Science Foundation of China (Grant No. LQ24H160021). Lifang He received financial support from Hunan Province Undergraduate Innovation and Entrepreneurship Training Program (grant no. S202510545100). Liuliu Chen received financial support from Science and Technology and Health Commission of Guizhou Province (Grant number: GZWKJ2021-491).

Data availability

The data generated during this study are not publicly available due to privacy concerns. However, the data are available from the corresponding author upon reasonable request.

Declarations

Competing interests

The authors declare no competing interests.