Original Paper
Abstract
Background: Cervical cancer remains the fourth leading cause of death among women globally, with a particularly severe burden in low-resource settings. A comprehensive approach—from screening to diagnosis and treatment—is essential for effective prevention and management. Large language models (LLMs) have emerged as potential tools to support health care, though their specific role in cervical cancer management remains underexplored.
Objective: This study aims to systematically evaluate the performance and interpretability of LLMs in cervical cancer management.
Methods: Models were selected from the AlpacaEval leaderboard version 2.0 and based on the capabilities of our computer. The questions inputted into the models cover aspects of general knowledge, screening, diagnosis, and treatment, according to guidelines. The prompt was developed using the Context, Objective, Style, Tone, Audience, and Response (CO-STAR) framework. Responses were evaluated for accuracy, guideline compliance, clarity, and practicality, graded as A, B, C, and D with corresponding scores of 3, 2, 1, and 0. The effective rate was calculated as the ratio of A and B responses to the total number of designed questions. Local Interpretable Model-Agnostic Explanations (LIME) was used to explain and enhance physicians’ trust in model outputs within the medical context.
Results: Nine models were included in this study, and a set of 100 standardized questions covering general information, screening, diagnosis, and treatment was designed based on international and national guidelines. Seven models (ChatGPT-4.0 Turbo, Claude 2, Gemini Pro, Mistral-7B-v0.2, Starling-LM-7B alpha, HuatuoGPT, and BioMedLM 2.7B) provided stable responses. Among all the models included, ChatGPT-4.0 Turbo ranked first with a mean score of 2.67 (95% CI 2.54-2.80; effective rate 94.00%) with a prompt and 2.52 (95% CI 2.37-2.67; effective rate 87.00%) without a prompt, outperforming the other 8 models (P<.001). Regardless of prompts, QiZhenGPT consistently ranked among the lowest-performing models, with P<.01 in comparisons against all models except BioMedLM. Interpretability analysis showed that prompts improved alignment with human annotations for proprietary models (median intersection over union 0.43), while medical-specialized models exhibited limited improvement.
Conclusions: Proprietary LLMs, particularly ChatGPT-4.0 Turbo and Claude 2, show promise in clinical decision-making involving logical analysis. The use of prompts can enhance the accuracy of some models in cervical cancer management to varying degrees. Medical-specialized models, such as HuatuoGPT and BioMedLM, did not perform as well as expected in this study. By contrast, proprietary models, particularly those augmented with prompts, demonstrated notable accuracy and interpretability in medical tasks, such as cervical cancer management. However, this study underscores the need for further research to explore the practical application of LLMs in medical practice.
doi:10.2196/63626
Keywords
Introduction
Cervical cancer is a significant global public health challenge, ranking fourth among all female cancers and remaining the leading cause of death in many low-income countries []. In 2020, approximately 604,127 new cases and 341,831 deaths from cervical cancer were reported worldwide []. Effective cervical cancer control necessitates an integrated approach that combines screening, accurate diagnosis, and personalized treatment to reduce morbidity and mortality. Despite a substantial decline in cervical cancer incidence in the United States since the introduction of screening [], up to 25% of women remain inadequately treated [], with even higher rates observed in resource-limited and developing countries []. Moreover, precise diagnosis and appropriate treatment are essential for addressing abnormalities detected through screening, particularly to prevent disease progression and improve survival outcomes []. Hence, strengthening these efforts is essential for reducing the global burden of cervical cancer and improving patient outcomes across diverse health care contexts.
Large language models (LLMs), as cutting-edge technologies in artificial intelligence, are trained on vast data sets and enable a wide range of applications, from text polishing to complex problem-solving, thanks to their unprecedented natural language understanding capabilities. In the health care domain, LLMs hold the potential to revolutionize medical practices, including decision-making, patient management, and clinical data interpretation [,]. Notably, OpenAI’s proprietary LLMs, ChatGPT-3.5 and ChatGPT-4.0, have demonstrated high performance on the United States Medical Licensing Examination (USMLE), with ChatGPT-4.0 achieving particularly impressive results [,]. Additionally, ChatGPT has shown competence across various medical fields, including surgery [], cardiology [], and plastic surgery []. Compared with generic language models, medical-specialized models—fine-tuned on domain-specific data sets and subjected to specialized adjustments—have achieved equivalent or superior performance [].
To date, only a limited number of studies [] have applied LLMs to questions related to cervical cancer, as well as explainability analyses on either closed- or open-source LLMs to assess transparency and interpretability. The management of abnormal cervical cancer screening results, diagnosis, and treatment is a complex task that requires careful interpretation and follow-up []. When deploying LLMs in cervical cancer management, it is crucial to evaluate their performance in managing abnormalities and to identify their strengths and limitations, particularly regarding model transparency and interpretability.
In this study, we aim to compare the performance of current prevalent LLMs in cervical cancer management by evaluating their responses to a set of specifically designed questions (
Figure 1). This research may provide valuable evidence to help clinicians manage screening results more effectively and accurately, particularly in regions with limited health care infrastructure.
Methods
Model Selection
The AlpacaEval leaderboard is an automated system designed to evaluate language models based on their adherence to instructions, ranking them by comparing their responses to reference answers from top-performing models such as GPT-4. It aims to reduce biases, such as those related to output length. Unlike other leaderboards that may focus on a single type, this leaderboard includes both open- and closed-source models. The selection of potential models—whether closed-source, open-source, or medically specialized—is determined by their win rates on version 2.0 of the leaderboard, updated on March 3, 2024.
For closed-source models, both free and paid versions are included, excluding those that are not publicly available or are in private beta. Open-source models are required to perform effectively on consumer-grade computers with standard configurations, given their potential use in resource-limited applications such as cervical cancer screening. The computer specifications for deploying these models are detailed in , with a maximum model size capacity of approximately 7 billion trainable parameters. The selection of medical-specialized models, which are limited in number on leaderboards, is informed by a study [] summarizing existing medical LLMs and their respective GitHub star counts. The performance of these medical LLMs is assessed based on the benchmark scores of their underlying models.
Criteria for Question and Prompt Designing
Questions Designing
A comprehensive question set was developed to evaluate model performance, including general questions and those specifically focused on cervical cancer screening, diagnosis, and treatment. General questions were designed by our gynecological experts to address the most common queries about cervical cancer, covering essential, foundational information frequently encountered in clinical practice. Screening-related questions were crafted with reference to the Chinese Society for Colposcopy and Cervical Pathology of the China Healthy Birth Science Association (CSCCP) Consensus on cervical cancer screening and abnormal management in China []. To ensure relevance and keep our questions up to date, we also incorporate the 2019 American Society for Colposcopy and Cervical Pathology (ASCCP) Risk-Based Management Consensus Guidelines for Abnormal Cervical Cancer Screening Tests and Cancer Precursors []. The questions comprehensively address each clinical decision outlined in the CSCCP guideline flowcharts, as detailed in . Additional screening questions were developed based on the Chinese Society of Clinical Oncology (CSCO) Guidelines for the Diagnosis and Treatment of Cervical Cancer (2023) []. The diagnosis and treatment questions were developed with reference to the Sociedad Española de Oncología Médica-Grupo Español de Investigación en Cáncer de Ovario (SEOM-GEICO) Clinical Guidelines on Cervical Cancer (2023) [], the CSCO Guidelines for the Diagnosis and Treatment of Cervical Cancer [], and The International Federation of Gynecology and Obstetrics (FIGO) 2018 Gynecologic Cancer Report – Interpretation of the Cervical Cancer Guidelines []. The design was guided by the principles outlined in 
Textbox 1.
1. Diverse complexity levels
- A combination of basic and advanced questions was included to evaluate the model’s ability to address both routine and complex clinical scenarios.
2. Strict guideline adherence
- Questions were structured to prioritize guideline-based knowledge, minimizing reliance on outdated or nonevidence-based practices.
3. Primarily closed-ended format
- Predominantly closed-ended questions were used to reduce subjective bias, with a few open-ended questions included to assess the model’s capacity for divergent medical problem-solving.
4. Definitive answers
- Each question was designed to have a clear, definitive answer.
These questions aim to evaluate the models’ understanding of clinical guidelines, their decision-making processes, and their ability to provide clear, actionable advice.
Prompt Designing
The prompt was designed using the Context, Objective, Style, Tone, Audience, and Response (CO-STAR) framework, which was ranked as the top prompt in the inaugural GPT-4 Prompt Engineering Competition. This framework was applied to guide the LLM in generating expert-level responses in gynecology, with a clear focus on defining the context, objective, style, tone, audience, and response format.
Questioning Method
Each designed question was sequentially input 3 times for each model, both with and without the designed prompt, to test consistency. The coherence of the responses was evaluated using semantic textual similarity [] by ChatGPT-3.5, the top-performing model on AlpacaEval, which was not used as a test model. If the semantics of all 3 responses are identical, they are sent back to the originating model to select the most suitable answer. In cases of discrepancies, a pairwise comparison is performed with scores ranging from 0 (not typical at all) to 100 (extremely typical) []. The 2 responses with the highest similarity scores are returned to their model, which then selects the most appropriate answer (
Figure 2).
Scoring Process and Criteria
Two gynecological experts independently and anonymously reviewed the responses to each question. If both experts agreed on a score, it was directly accepted; otherwise, they discussed it to determine the final score. Responses were evaluated based on accuracy, adherence to clinical guidelines, clarity of communication, and practicality. A scoring system, modified from a previous study [], was used to categorize responses into 4 grades: A, B, C, and D, to minimize subjective bias. Grades A and B were considered effective, and the model’s effective rate was calculated as follows:
Effective rate = (NA+NB)/(NA+NB+NC+ND) × 100%
where N represents the number of each grade. Scores are weighted at 3, 2, 1, and 0 points for statistical analysis (
Table 1).
| Grade | Description | Scores |
| A | Completely correct with comprehensive information | 3 |
| B | Mostly correct, but with missing information or minor errors | 2 |
| C | Contains major errors but with some correct content | 1 |
| D | Completely wrong or off-topic | 0 |
Model Explainability Analysis
Local Interpretable Model-Agnostic Explanations (LIME) is widely recognized for generating locally interpretable explanations of machine learning model predictions, including natural language processing models [,]. In this study, LIME was used to interpret LLM outputs by adapting methods previously successful in natural language processing. The primary LIME parameter, the number of samples, was set to 10 times the input sentence’s token count, based on preliminary experiments and prior applications of LIME to LLMs []. Each input question was analyzed to identify key terms with assigned weights, and the top 5 key terms by weight were selected. Our experts manually annotated 5 key terms per question for comparison. An intersection-over-union (IoU) analysis was performed between the LIME-selected key terms and the expert-annotated key terms to evaluate their alignment (
Figure 3).
IoU(x1, x2) = (|x1∩x2|)/(| x1∪x2|)
Ethical Considerations
This study did not involve human participants, identifiable patient data, or protected health information. The data utilized in this study comprised publicly available sources, including leaderboards, clinical guidelines, and secondary analyses of model-generated outputs. Therefore, an ethical review was not required under Zhongnan Hospital of Wuhan University’s secondary research policies. The study complied with the Declaration of Helsinki and institutional guidelines for secondary data use.
Statistical Methods
Analyses were conducted using R version 4.3.1 (R Foundation) and RStudio 2023.12.1+402 (R Foundation). Differences across models were assessed using the chi-square test for categorical variables. For paired comparisons, data were first tested for normality. If normally distributed, a paired t test was applied, with results reported as mean and SD; otherwise, a paired Wilcoxon rank sum test was used, with outcomes presented as median and IQR. Effective rates were reported as mean values with 95% CIs. A P value of less than .05 was considered indicative of a significant difference.
Results
Model Selection
After screening for win rates and conducting tests on our computers, our study included 9 models. The proprietary models are ChatGPT-4.0 Turbo, Claude 2, and Gemini Pro, which are accessible through their official websites. The open-source LLMs include Mistral-7B-v0.2, Starling-LM-7B Alpha, and Microsoft Phi-2. The medical-specialized models are the Chinese models HuatuoGPT and QiZhenGPT, along with the English model BioMedLM 2.7B. The expected performance ranking of the selected models is as follows: ChatGPT-4.0 Turbo > Gemini Pro > Claude 2 > Mistral-7B-v0.2 > Starling-LM-7B Alpha > ChatGLM 6B (QiZhenGPT) > Phi-2 > Baichuan2-7B-Chat (HuatuoGPT). BioMedLM 2.7B is excluded from this ranking because it is not listed on the AlpacaEval Leaderboard. The characteristics of the included models are presented in Table 2.
| Model and access reference | AlpacaEval win rate | Description |
| ChatGPT-4.0 Turbo [] | 50% | Developed by OpenAI, ChatGPT-4.0 Turbo is an LLMa that is currently the most powerful in terms of performance. |
| Claude 2 [] | 17.19% | Developed by Anthropic, it is built on the GPT-3 architecture. This model features a context window of 100,000 ultra-long tokens, enabling it to handle longer context inputs efficiently. |
| Mistral-7B-v0.2 [] | 14.72% | Mistral-7B-v0.2 is the strongest open-source model on the list that can be deployed on consumer computers. Furthermore, the popularity of this model is high, as it received more than 700,000 downloads in January 2024. |
| Starling-LM-7B alpha [] | 14.25% | A fine-tuned model that outperforms all models to date on MT-Bench except for OpenAI’s GPT-4 and GPT-4 Turbo. |
| Gemini Pro [] | 18.18% | Developed by Google DeepMind, the more advanced Gemini Ultra is not yet available to the public, so we used the Pro version. |
| HuatuoGPT 2-7B [] | 1.99% (base model) | Developed by the Shenzhen Institute of Big Data and The Chinese University of Hong Kong, this Chinese medical LLM is fine-tuned based on Baichuan2-7B. Uses deploying method. The online demo is available at []. |
| QiZhenGPT [] | 3.01% (base model) | Released by Zhejiang University, the project includes 3 versions, each fine-tuned from the base models of ChatGLM-6B, Chinese-LLaMA-7B, and CaMA-13B. |
| Phi-2 [] | 2.34% | Released by Microsoft, this small language model has a data size of only 2.7 billion. Easy to deploy, even on consumer-grade computers, where it exhibits exceptionally fast response times. |
| BioMedLM 2.7B [] | N/Ab | Previously known as PubMedGPT 2.7B, this model was developed through pretraining. |
aLLM: large language model.
bN/A: not applicable.
Questions and Prompts for LLMs
The question set consisted of 100 questions designed to encompass a broad range of clinical scenarios commonly encountered in cervical cancer management. The first 22 questions focused on general knowledge, emphasizing foundational aspects frequently encountered in clinical gynecology. The next 40 questions addressed cervical cancer screening, aligning with the latest consensus guidelines and decision-making protocols. Subsequently, 6 and 32 questions covered diagnosis and treatment, respectively, offering a comprehensive evaluation of the models’ ability to interpret diagnostic criteria and recommend evidence-based treatment options. By including both routine and complex queries, the question set serves as a robust benchmark for assessing model performance, accuracy, and adherence to evidence-based medical practices. The complete list of questions is provided in Table 3.
| Category | Questions |
| Questions related to general knowledge |
|
| Questions related to screening |
|
| Questions related to diagnosis |
|
| Questions related to treatments |
|
aHPV: human papillomavirus.
bASC-US: atypical squamous cells of undetermined significance.
cASC-H: atypical squamous cells, cannot exclude high-grade squamous intraepithelial lesion
dLSIL: low-grade squamous intraepithelial lesion.
eHSIL: high-grade squamous intraepithelial lesion.
fAGC: atypical glandular cell.
gTZ3: Type 3 transformation zone.
hCIN: cervical intraepithelial neoplasia.
iFIGO: International Federation of Gynecology and Obstetrics.
Using the CO-STAR framework, the prompt was designed to guide the model in providing clinically relevant and detailed responses, meeting the standards necessary for accurate interpretation in cervical cancer management. The specific details of the prompt are presented in Table 4.
| Prompt element | Content |
| # Context # | Now you are a gynecologist with over 20 years of experience in medicine and you are answering questions about the medical specialty of cervical cancer treatment, diagnosis, and screening. |
| # Objective # | Please answer the following questions correctly and in strict accordance with the latest guidelines for the screening, treatment, and diagnosis of cervical cancer. |
| # Style # | The information should be clear, concise, and medically accurate, using terminology appropriate for both health care professionals and patients. |
| # Tone # | The tone should be formal and professional, recognizing the sensitive nature of cancer-related discussions. |
| # Audience # | The primary audience includes health care professionals, researchers, and patients seeking information about cervical cancer management. |
| # Response # | Generate detailed responses to specific queries regarding cervical cancer. Assess the accuracy and relevance of the information provided. |
aCO-STAR: Context, Objective, Style, Tone, Audience, and Response.
Model Stability
Among the 9 models evaluated, 7 demonstrated good reproducibility with stable responses. However, the repeatability of Phi-2 and QiZhenGPT was unsatisfactory, as posing the same question 3 times often resulted in varying answers. For Phi-2, 61 out of 100 responses with the prompt and 68 responses without the prompt exhibited semantic differences across repetitions. Similarly, for QiZhenGPT, 60 responses with the prompt and 55 without the prompt varied. In both cases, pairwise comparisons were necessary to determine the final output (see ).
Model Efficacy
The evaluation results for each model, with and without the prompt, are presented in 
Figure 4. The top 3 performers were all proprietary models. ChatGPT-4.0 Turbo achieved the highest effective rate, at 94% (mean score 2.67, 95% CI 2.54-2.80) with the prompt and 87% (mean score 2.52, 95% CI 2.37-2.67) without it, highlighting the positive impact of the prompt on its performance. Claude 2 maintained an effective rate of 85% both with and without the prompt, with similar mean scores of 2.35 (95% CI 2.16-2.54) and 2.39 (95% CI 2.22-2.56), respectively. Gemini Pro showed moderate improvement, with its effective rate increasing from 66% (mean score 2.00, 95% CI 1.80-2.20) without the prompt to 77% (mean score 2.25, 95% CI 2.06-2.44) with the prompt.
By contrast, the 3 medically specialized models exhibited lower effective rates. HuatuoGPT achieved an effective rate of 53% (mean score 2.00, 95% CI 1.80-2.20) with the prompt, which unexpectedly increased to 57% (mean score 1.76, 95% CI 1.54-1.98) without it. BioMedLM showed minimal improvement, with an effective rate of 39% (mean score 1.13, 95% CI 0.90-1.36) with the prompt and 38% (mean score 1.76, 95% CI 1.54-1.98) without it. QiZhenGPT had the lowest performance, with an effective rate of 33% (mean score 1.13, 95% CI 0.91-1.35) with the prompt and 32% (mean score 1.19, 95% CI 0.97-1.41) without it, showing limited impact from the prompt on enhancing its responses. The STS testing results are provided in . Detailed responses and original scoring are provided in .
The chi-square test revealed significant differences across models (P=.001). As the data for each model did not follow a normal distribution (P<.01), the Wilcoxon rank sum test was applied. With the prompt, ChatGPT-4.0 Turbo and Claude 2 exhibited highly significant differences (P<.001) compared with most other models, indicating substantial performance enhancement when the prompt was used. This pattern remained consistent in comparisons with lower-performing models, such as HuatuoGPT, BioMedLM, and QiZhenGPT. Without the prompt, significant differences were still observed, particularly between high-performing models such as ChatGPT-4.0 Turbo (P<.001) and Claude 2 (P<.001) and lower-performing models. However, the absence of the prompt reduced significance in certain comparisons, such as between Mistral-7B and Gemini Pro (P=.30) or BioMedLM and QiZhenGPT (P=.64). When comparing performance with and without the prompt, ChatGPT-4.0 Turbo and Gemini Pro demonstrated statistically significant improvements with the prompt (P<.001), whereas Claude 2 showed no significant difference (P=.07). By contrast, models such as BioMedLM (P=.77), Phi-2 (P=.53), and QiZhenGPT (P=.01) exhibited minimal or insignificant changes (
Figure 5).
Model Explainability
Given the nonnormal distribution of IoU values for each model, the Wilcoxon rank sum test was used to assess differences. As shown in 
Figure 6, the inclusion of prompts significantly improved the alignment between model-generated explanations and human annotations, with all models exhibiting statistically significant differences between prompted and unprompted conditions (P<.001). Specifically, Claude 2, Gemini Pro, Starling-LM-7B Alpha, ChatGPT-4.0 Turbo, and Mistral-7B-v0.2 demonstrated a consistent median IoU of 0.43 with prompts. Among these, ChatGPT-4.0 Turbo had the widest IoU range (IQR 0.56). Without prompts, the median IoU for these models dropped to 0.25, with narrower IQRs ranging from 0.32 to 0.43, indicating reduced interpretability consistency. Among the medically specialized models, QiZhenGPT showed the most substantial improvement with prompts, achieving a median IoU of 0.43 (IQR 0.42), aligning it with the performance of proprietary models under similar conditions. By contrast, BioMedLM 2.7B and HuatuoGPT exhibited lower interpretability, with median IoUs of 0.29 and 0.25, respectively, and smaller IQRs in nonprompted conditions (median IoU of 0.11 and IQR of 0.25 for both).
Discussion
Principal Findings
This study systematically evaluated 9 LLMs for their performance, stability, and interpretability in cervical cancer management. The results revealed that proprietary models, such as ChatGPT-4.0 Turbo, Claude 2, and Gemini Pro, achieved superior response accuracy and interpretability, particularly with prompt guidance. By contrast, medically specialized models such as HuatuoGPT, QiZhenGPT, and BioMedLM demonstrated comparatively lower effectiveness, with limited improvement from prompt use. Notably, while proprietary models exhibited consistent reproducibility, certain open-source and specialized models, such as Phi-2 and QiZhenGPT, showed variable responses upon repeated questioning. Furthermore, the use of prompts significantly enhanced interpretability in models such as Claude 2, Gemini Pro, and Starling-LM-7B Alpha, highlighting the potential of structured input to improve alignment with clinical expectations.
Comparison to Prior Work
In terms of average score ranking, proprietary models such as ChatGPT-4.0 Turbo, Claude 2, and Gemini Pro outperformed open-source models. This result aligns with traditional views on the superiority of proprietary systems []. However, without the prompt, Mistral-7B outperformed Gemini Pro. Among the open-source models, Mistral-7B-v0.2 and Starling-LM-7B Alpha outperformed HuatuoGPT and BioMedLM 2.7B. However, the repeatability of answers from Microsoft Phi-2 was poor, making it unsuitable for medical applications, while ChatGPT-4.0 Turbo and Claude 2 provided accurate and consistent responses. Our results indicated that the performance of the 3 medical models was average, challenging the prevailing belief that medical-specific models are superior for medical queries []. Previous studies [,] have shown that larger models, characterized by increased parameter counts, tend to perform better. Additionally, as the model scale increases, its generalization ability improves []. This may explain the relative underperformance of medical models compared with proprietary models, given the substantial disparity in parameter magnitude between them.
Recent advancements in algorithms have been shown to improve the performance of LLMs in the medical field [,], with research [] indicating significant accuracy improvements using specific prompts. The integration of prompts has had a notable impact on the performance of several LLMs, emphasizing the value of structured input in guiding model responses within clinical contexts. Proprietary models, such as ChatGPT-4.0 Turbo and Gemini Pro, showed marked improvements in effective rate and response accuracy when guided by the CO-STAR prompt framework, suggesting that structured prompts help enhance focus on relevant clinical information and reduce ambiguity []. Conversely, models with specialized but limited training, such as BioMedLM, exhibited minimal sensitivity to prompts, likely due to architectural limitations in processing complex prompt structures []. Interestingly, HuatuoGPT experienced a decline in performance with the addition of prompts. This unexpected outcome suggests that the structured prompt for HuatuoGPT may have interfered with its response generation by introducing constraints that conflicted with its training data or underlying language patterns, potentially limiting its ability to accurately interpret open-ended clinical scenarios []. Additionally, smaller models often become confused when handling longer prompts []. The variation in prompt effectiveness across models underscores that, while structured prompts generally improve response precision, their impact is influenced by the model design and data scope.
The IoU serves as a robust indicator of alignment between model-generated explanations and human annotations, providing insights into the interpretability of LLMs in clinical contexts []. A higher IoU reflects greater consistency with human-provided explanations, suggesting enhanced model transparency and reliability in decision-making support. Our results demonstrate that a higher IoU corresponds to better alignment between model-generated explanations and human annotations, indicating improved interpretability. Proprietary models, particularly ChatGPT-4.0 Turbo and Claude-2, performed well in aligning with human explanations when prompts were used, highlighting their potential for generating clinically relevant interpretations. Interestingly, the rankings for model explainability based on IoU scores do not directly correlate with those based on effective rates. This discrepancy likely arises because improvements in model performance do not necessarily enhance explainability []. According to previous studies [], as models become more accurate, their alignment with human-annotated explanations does not necessarily improve. This misalignment suggests that the factors driving a model’s effectiveness in task accuracy differ from those contributing to explainability. Higher-performing models may rely on complex, implicit patterns that are not fully captured by metrics such as IoU, which primarily assess agreement with human logic rather than the model’s actual reasoning process []. However, IoU alone may not fully capture explanation quality, as it can overlook aspects such as coherence and clinical relevance. Therefore, incorporating qualitative assessments alongside IoU could provide a more comprehensive measure of model explainability in clinical contexts.
Ethical Issues
LLMs have performed well in the cervical cancer question-and-answer task, but ethical considerations, such as transparency, data privacy, and algorithmic bias, remain []. Tools such as LIME enhance transparency and simplify the explanation of AI decisions, with further progress expected []. Deployments adhere to strict data laws to ensure ongoing improvements in privacy, and technological advancements are anticipated to further safeguard patient privacy []. Bias issues are managed through explainable AI and methods such as training with multiple multiinstitutional or population data sets, as well as using generative adversarial networks to obtain more representative data []. While practical challenges remain in technology integration and staff training, LLMs are more easily adopted due to their application programming interfaces and their ability to act as personalized learning assistants, reducing the reliance on extensive medical staff training [,].
Limitations
Our study also has limitations: (1) Because of the limited capabilities of our computers, we were unable to test all existing LLMs. It is possible that there are models with better performance than ChatGPT-4.0 Turbo in handling abnormal cervical screening results. (2) Our study did not include augmented algorithms or corpora that have been shown to improve LLM performance in other studies, as not all patients or physicians are familiar with these tools. The lack of these enhancements may limit the ability of LLMs to demonstrate their full potential in answering medical questions. This absence could have restricted the models from showcasing their full capabilities in medical query resolution, potentially affecting the generalizability of our results in more advanced settings. (3) The study conducted assessments under controlled, structured questions, which may not fully reflect the model’s performance in dynamic, real-world clinical settings. This controlled environment may limit our ability to assess the adaptability of LLMs in unpredictable or complex patient interactions.
Conclusions
This study highlights the pivotal role of LLMs, particularly proprietary ones such as ChatGPT-4.0 Turbo, in enhancing clinical decision-making in cervical cancer screening. ChatGPT-4.0 Turbo outperforms both open-source and medical-specialized models in interpreting clinical guidelines and handling medical queries. Such findings are essential for improving the accuracy and efficiency of medical screenings and diagnoses, ultimately enhancing health care delivery and patient care. Further research is needed to assess the effectiveness of LLMs in medical applications, potentially leading to the development of models more tailored for medical practice and advancing overall health care.
Acknowledgments
This work was supported by the Science and Technology Innovation Cultivation Funding of Zhongnan Hospital of Wuhan University (grant CXPY2022049).
Data Availability
The 100 questions developed for model evaluation and all analyzed data in this study are included in the published manuscript and its multimedia appendices.
Conflicts of Interest
None declared.
Multimedia Appendix 2
Cervical cancer screening and abnormal result management process by the CSCCP (Chinese Society for Colposcopy and Cervical Pathology of the China Healthy Birth Science Association).
DOCX File , 467 KBMultimedia Appendix 3
Semantic textual similarity testing for Phi-2 and QiZhenGPT model.
DOCX File , 154 KBReferences
- Sung H, Ferlay J, Siegel RL, Laversanne M, Soerjomataram I, Jemal A, et al. Global Cancer Statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin. May 2021;71(3):209-249. [FREE Full text] [CrossRef] [Medline]
- Cohen CM, Wentzensen N, Castle PE, Schiffman M, Zuna R, Arend RC, et al. Racial and ethnic disparities in cervical cancer incidence, survival, and mortality by histologic subtype. J Clin Oncol. Feb 10, 2023;41(5):1059-1068. [FREE Full text] [CrossRef] [Medline]
- Siegel RL, Miller KD, Wagle NS, Jemal A. Cancer statistics, 2023. CA Cancer J Clin. Jan 2023;73(1):17-48. [FREE Full text] [CrossRef] [Medline]
- Burmeister CA, Khan SF, Schäfer G, Mbatani N, Adams T, Moodley J, et al. Cervical cancer therapies: current challenges and future perspectives. Tumour Virus Res. Jun 2022;13:200238. [FREE Full text] [CrossRef] [Medline]
- Chavez MR, Butler TS, Rekawek P, Heo H, Kinzler WL. Chat Generative Pre-trained Transformer: why we should embrace this technology. Am J Obstet Gynecol. Jun 2023;228(6):706-711. [CrossRef] [Medline]
- Kim JK, Chua M, Rickard M, Lorenzo A. ChatGPT and large language model (LLM) chatbots: the current state of acceptability and a proposal for guidelines on utilization in academic medicine. J Pediatr Urol. Jun 02, 2023:598-604. [CrossRef] [Medline]
- Kung TH, Cheatham M, Medenilla A, Sillos C, De Leon L, Elepaño C, et al. Performance of ChatGPT on USMLE: potential for AI-assisted medical education using large language models. PLOS Digit Health. Feb 2023;2(2):e0000198. [FREE Full text] [CrossRef] [Medline]
- Knoedler L, Alfertshofer M, Knoedler S, Hoch CC, Funk PF, Cotofana S, et al. Pure wisdom or potemkin villages? A comparison of ChatGPT 3.5 and ChatGPT 4 on USMLE step 3 style questions: quantitative analysis. JMIR Med Educ. Jan 05, 2024;10:e51148. [FREE Full text] [CrossRef] [Medline]
- Beaulieu-Jones BR, Berrigan MT, Shah S, Marwaha JS, Lai S, Brat GA. Evaluating capabilities of large language models: performance of GPT-4 on surgical knowledge assessments. Surgery. Apr 2024;175(4):936-942. [CrossRef] [Medline]
- Skalidis I, Cagnina A, Luangphiphat W, Mahendiran T, Muller O, Abbe E, et al. ChatGPT takes on the European Exam in Core Cardiology: an artificial intelligence success story? Eur Heart J Digit Health. May 2023;4(3):279-281. [FREE Full text] [CrossRef] [Medline]
- Xie Y, Seth I, Hunter-Smith DJ, Rozen WM, Ross R, Lee M. Aesthetic surgery advice and counseling from artificial intelligence: a rhinoplasty consultation with ChatGPT. Aesthetic Plast Surg. Oct 2023;47(5):1985-1993. [FREE Full text] [CrossRef] [Medline]
- Singhal K, Azizi S, Tu T, Mahdavi SS, Wei J, Chung HW, et al. Large language models encode clinical knowledge. Nature. Aug 2023;620(7972):172-180. [FREE Full text] [CrossRef] [Medline]
- Hermann CE, Patel JM, Boyd L, Growdon WB, Aviki E, Stasenko M. Let's chat about cervical cancer: assessing the accuracy of ChatGPT responses to cervical cancer questions. Gynecol Oncol. Dec 2023;179:164-168. [CrossRef] [Medline]
- Stonehocker J. Cervical cancer screening in pregnancy. Obstet Gynecol Clin North Am. Jun 2013;40(2):269-282. [CrossRef] [Medline]
- Zhou H, Liu F, Gu B, Zou X, Huang J, Wu J, et al. A survey of large language models in medicine: progress, application, and challenge. arXiv. :e1. Preprint posted online on July 22, 2023. [FREE Full text] [CrossRef]
- Zhao Y, Wei L. Expert consensus interpretation on Chinese cervical cancer screening and abnormal management issues by CSCCP. Journal of Practical Obstetrics and Gynecology. Feb 15, 2018;34(2):101-104. [FREE Full text]
- Perkins RB, Guido RS, Castle PE, Chelmow D, Einstein MH, Garcia F, et al. 2019 ASCCP Risk-Based Management Consensus Guidelines Committee. 2019 ASCCP risk-based management consensus guidelines for abnormal cervical cancer screening tests and cancer precursors. J Low Genit Tract Dis. Apr 2020;24(2):102-131. [FREE Full text] [CrossRef] [Medline]
- Wu L, Li L, Huang M, Li G, Lou G, Wu X, et al. Chinese Society of Clinical Oncology (CSCO) Guidelines for the Diagnosis and Treatment of Cervical Cancer (2023). Beijing, China. People's Medical Publishing House; Aug 01, 2023.
- Manso L, Ramchandani-Vaswani A, Romero I, Sánchez-Lorenzo L, Bermejo-Pérez MJ, Estévez-García P, et al. SEOM-GEICO clinical guidelines on cervical cancer (2023). Clin Transl Oncol. Nov 2024;26(11):2771-2782. [CrossRef] [Medline]
- Zhou H, Wang D, Luo M, Lin Z. FIGO 2018 gynecologic cancer report-interpretation of the cervical cancer guidelines. Zhongguo Shiyong Fuke Yu Chanke Zazhi. Feb 19, 2019;35(1):95-103. [CrossRef]
- Jiang T, Huang S, Luan Z, Wang D, Zhuang F. Scaling sentence embeddings with large language models. arXiv. :1. Preprint posted online on July 31, 2023. [FREE Full text] [CrossRef]
- Le Mens G, Kovács B, Hannan MT, Pros G. Uncovering the semantics of concepts using GPT-4. Proc Natl Acad Sci U S A. Dec 05, 2023;120(49):e2309350120. [FREE Full text] [CrossRef] [Medline]
- Lozano A, Fleming SL, Chiang C, Shah N. Clinfo.ai: an open-source retrieval-augmented large language model system for answering medical questions using scientific literature. Pac Symp Biocomput. 2024;29:8-23. [FREE Full text] [Medline]
- Barr KN, Blomberg T, Liu J, Saraiva LA, Papapetrou P. Evaluating local interpretable model-agnostic explanations on clinical machine learning classification models. New York, NY. IEEE; 2020. Presented at: 2020 IEEE 33rd International Symposium on Computer-Based Medical Systems (CBMS); July 28-30, 2020; Rochester, MN, USA. [CrossRef]
- Madsen A, Reddy S, Chandar S. Post-hoc interpretability for neural NLP: a survey. ACM Comput Surv. Dec 23, 2022;55(8):1-42. [CrossRef]
- Huang S, Mamidanna S, Jangam S, Zhou Y, Gilpin L. Can large language models explain themselves? A study of LLM-generated self-explanations. arXiv. :1. Preprint posted online on October 17, 2023. [FREE Full text] [CrossRef]
- OpenAI. URL: https://chat.openai.com/ [accessed 2025-01-10]
- Claude. URL: https://claude.ai/ [accessed 2025-01-10]
- Hugging Face. URL: https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2 [accessed 2025-01-10]
- Hugging Face. URL: https://huggingface.co/berkeley-nest/Starling-LM-7B-alpha [accessed 2025-01-10]
- Google Gemini. URL: https://gemini.google.com/ [accessed 2025-01-10]
- GitHub. URL: https://github.com/FreedomIntelligence/HuatuoGPT?tab=readme-ov-file [accessed 2025-01-10]
- HuatuoGPT. URL: https://www.huatuogpt.cn/#/ [accessed 2025-01-10]
- GitHub. URL: https://github.com/CMKRG/QiZhenGPT?tab=readme-ov-file [accessed 2025-01-10]
- Hugging Face. URL: https://huggingface.co/microsoft/phi-2 [accessed 2025-01-10]
- Hugging Face. URL: https://huggingface.co/stanford-crfm/BioMedLM [accessed 2025-01-10]
- Chen H, Jiao F, Li X, Qin C, Ravaut M, Zhao R, et al. ChatGPT's one-year anniversary: are open-source large language models catching up? arXiv. :1. Preprint posted online on January 15, 2024. [FREE Full text] [CrossRef]
- Tsoutsanis P, Tsoutsanis A. Evaluation of large language model performance on the Multi-Specialty Recruitment Assessment (MSRA) exam. Comput Biol Med. Jan 2024;168:107794. [CrossRef] [Medline]
- Yang H, Li M, Zhou H, Xiao Y, Fang Q, Zhang R. One LLM is not enough: harnessing the power of ensemble learning for medical question answering. medRxiv. :1. Preprint posted online December 24, 2023. [FREE Full text] [CrossRef] [Medline]
- Kaplan J, McCandlish S, Henighan T, Brown T, Chess B, Child R, et al. Scaling laws for neural language models. arXiv. :1. Preprint posted online on January 23, 2020. [FREE Full text]
- Liévin V, Hother CE, Motzfeldt AG, Winther O. Can large language models reason about medical questions? Patterns (N Y). Mar 08, 2024;5(3):100943. [FREE Full text] [CrossRef] [Medline]
- Sivarajkumar S, Kelley M, Samolyk-Mazzanti A, Visweswaran S, Wang Y. An empirical evaluation of prompting strategies for large language models in zero-shot clinical natural language processing: algorithm development and validation study. JMIR Med Inform. Apr 08, 2024;12:e55318. [FREE Full text] [CrossRef] [Medline]
- Chen B, Zhang Z, Langrené N, Zhu S. Unleashing the potential of prompt engineering in large language models: a comprehensive review. arXiv. :1. Preprint posted online on September 5, 2024. [FREE Full text] [CrossRef]
- Lu A, Zhang H, Zhang Y, Wang X, Yang D. Bounding the capabilities of large language models in open text generation with prompt constraints. arXiv. :1. Preprint posted online on February 17, 2023. [FREE Full text] [CrossRef]
- Lester B, Al-Rfou R, Constant N. The power of scale for parameter-efficient prompt tuning. arXiv. :1. Preprint posted online on September 2, 2021. [FREE Full text] [CrossRef]
- Alvarez-Melis D, Kaur H, Wallach H, Vaughan J. From human explanation to model interpretability: a framework based on weight of evidence. arXiv. :1. Preprint posted online on September 20, 2021. [FREE Full text] [CrossRef]
- Lazebnik T, Bunimovich-Mendrazitsky S, Rosenfeld A. An algorithm to optimize explainability using feature ensembles. Appl Intell. Feb 01, 2024;54(2):2248-2260. [CrossRef]
- Heyen H, Widdicombe A, Siegel N, Perez-Ortiz M, Treleaven P. The effect of model size on LLM post-hoc explainability via LIME. arXiv. :1. Preprint posted online on May 8, 2024. [FREE Full text]
- Chakraborty S, Tomsett R, Raghavendra R, Harborne D, Alzantot M, Cerutti F, et al. Interpretability of deep learning models: a survey of results. 2018. Presented at: 2017 IEEE SmartWorld, Ubiquitous Intelligence & Computing, Advanced & Trusted Computed, Scalable Computing & Communications, Cloud & Big Data Computing, Internet of People and Smart City Innovation (SmartWorld/SCALCOM/UIC/ATC/CBDCom/IOP/SCI); August 4-8, 2017; San Francisco, CA, USA. [CrossRef]
- Zhui L, Fenghe L, Xuehu W, Qining F, Wei R. Ethical considerations and fundamental principles of large language models in medical education: viewpoint. J Med Internet Res. Aug 01, 2024;26:e60083. [FREE Full text] [CrossRef] [Medline]
- Dong X, Wang S, Lin D, Rajbahadur G, Zhou B, Liu S, et al. PromptExp: multi-granularity prompt explanation of large language models. arXiv. :1. Preprint posted online on October 30, 2024. [FREE Full text]
- Zhan X, Seymour W, Such J. Beyond individual concerns: multi-user privacy in large language models. 2024. Presented at: CUI '24: Proceedings of the 6th ACM Conference on Conversational User Interfaces; July 8-10, 2024; Luxembourg City, Luxembourg. [CrossRef]
- Ong JCL, Seng BJJ, Law JZF, Low LL, Kwa ALH, Giacomini KM, et al. Artificial intelligence, ChatGPT, and other large language models for social determinants of health: current state and future directions. Cell Rep Med. Jan 16, 2024;5(1):101356. [FREE Full text] [CrossRef] [Medline]
- Hart SN, Hoffman NG, Gershkovich P, Christenson C, McClintock DS, Miller LJ, et al. Organizational preparedness for the use of large language models in pathology informatics. J Pathol Inform. 2023;14:100338. [FREE Full text] [CrossRef] [Medline]
- Su Z, Tang G, Huang R, Qiao Y, Zhang Z, Dai X. Based on medicine, the now and future of large language models. Cell Mol Bioeng. Aug 16, 2024;17(4):263-277. [CrossRef] [Medline]
Abbreviations
| ASCCP: American Society for Colposcopy and Cervical Pathology |
| CO-STAR: Context, Objective, Style, Tone, Audience, and Response |
| CSCCP: Chinese Society for Colposcopy and Cervical Pathology of the China Healthy Birth Science Association |
| CSCO: Chinese Society of Clinical Oncology |
| FIGO: International Federation of Gynecology and Obstetrics |
| GEICO: Grupo Español de Investigación en Cáncer de Ovario |
| IoU: intersection over union |
| LIME: Local Interpretable Model-agnostic Explanations |
| LLM: large language model |
| SEOM: Sociedad Española de Oncología Médica |
| USMLE: United States Medical Licensing Examination |
Edited by A Coristine; submitted 25.06.24; peer-reviewed by K Keshtkar, S Mao, A Kalluchi; comments to author 27.09.24; revised version received 01.11.24; accepted 11.12.24; published 05.02.25.
Copyright©Warisijiang Kuerbanjiang, Shengzhe Peng, Yiershatijiang Jiamaliding, Yuexiong Yi. Originally published in the Journal of Medical Internet Research (https://www.jmir.org), 05.02.2025.
This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in the Journal of Medical Internet Research (ISSN 1438-8871), is properly cited. The complete bibliographic information, a link to the original publication on https://www.jmir.org/, as well as this copyright and license information must be included.