The application and challenges of ChatGPT in laboratory medicine

Capabilities of ChatGPT in addressing clinical laboratory inquiries

Despite not undergoing specialized training on medical data, ChatGPT has demonstrated initial practicality in the medical field. A study by Munoz-Zuluaga et al. involved submitting 65 questions covering multiple topics to ChatGPT to assess its capabilities and accuracy in answering laboratory medicine-related questions [36]. The results showed that ChatGPT correctly answered 50.7 % of the questions, provided incomplete or partially correct responses for 23.1 %, delivered incorrect or misleading information for 16.9 %, and produced irrelevant answers for 9.3 %. Notably, correct answers were more concentrated in questions related to foundational medical or technical knowledge (59.1 %), while errors were more common in questions pertaining to laboratory procedures or regulations (31 %). Although GPT-4 shows significant improvements in accuracy, limitations in certain domains remain evident. A study by Girton et al. evaluated ChatGPT’s ability to address 49 real patient clinical laboratory questions on social media platforms such as Reddit and Quora, and compared these responses to those provided by medical students [37]. Reviewers tended to prefer ChatGPT’s answers over those from medical professionals. This highlights ChatGPT’s progress in handling complex clinical questions, while simultaneously revealing gaps in medical professional education in this area. Approximately half of the medical students’ answers were rated as “poor,” compared to less than 10 % of ChatGPT’s responses receiving similar ratings [37]. The limited emphasis on clinical laboratory knowledge in medical education may contribute to even experienced medical professionals struggling to provide high-quality answers to real-world questions.

Another study compared responses from three chatbots (ChatGPT, Gemini, and Le Chat) with those from certified physicians in online consultations [38]. Overall, chatbots outperformed online medical professionals in interpreting laboratory results. While online doctors ranked first in 60 % of cases, Gemini received lower scores in only 39 % of cases. ChatGPT demonstrated comparable quality and accuracy to human practitioners. However, chatbots showed higher tendencies for overestimating clinical severity (22–33 % incidence) compared to physicians’ 1.0 % overestimation rate. This highlights chatbots’ challenges in processing complex contextual information and interpreting laboratory data, with the lack of unified reference standards potentially leading to inconsistent interpretations of patient test data [38].

Implementation of chatbots in medical laboratory diagnostics

Recent advancements in artificial intelligence have garnered significant research interest regarding biochemical data interpretation. Kaftan AN et al. demonstrated varying accuracy levels among AI models (Copilot, Gemini) in biochemical data analysis [39]. Copilot exhibited superior performance across all evaluation metrics, achieving statistically significant advantages in biochemical parameter interpretation compared to other models. This suggests Copilot’s enhanced data processing capabilities may stem from its sophisticated algorithmic architecture. While other models demonstrated suboptimal performance relative to Copilot, their domain-specific competencies retain practical relevance for targeted clinical applications [39].

The microbiology domain has emerged as a strategic focus for artificial intelligence (AI) integration due to its demanding requirements for diagnostic precision and rapid data processing. AI demonstrates transformative potential in enhancing clinical decision-making efficiency and optimizing infectious disease management, thereby improving patient outcomes. However, AI systems require further refinement in handling complex clinical scenarios and generating contextually appropriate responses to ensure clinical relevance [40]. Notably, ChatGPT-4.0 exhibits significant limitations when addressing Antimicrobial Susceptibility Testing (AST)-related inquiries [41], 42]. The model inappropriately recommended clindamycin for Enterococcus infections and failed to incorporate standard methodologies for polymyxin susceptibility assessment. Such erroneous guidance could misguide clinical practice, resulting in ineffective therapeutic regimens and potential patient harm.

Li Y et al. conducted a systematic evaluation of two LLMs, ChatGPT and Google Gemini, in addressing HBV-related clinical inquiries [43]. ChatGPT-4.0 demonstrated superior overall performance, achieving 80.8 % accuracy on evidence-based questions compared to Google Gemini (73.1 %). Domain-specific analysis revealed ChatGPT-4.0’s enhanced diagnostic utility in HBV serological interpretation, while Google Gemini provided more comprehensive descriptions of clinical manifestations. All models exhibited critical limitations in conveying HBV prevention strategies, particularly regarding vaccine information currency. Notably, only Google Gemini referenced the current consensus recommending neonatal hepatitis B vaccination within 12 h of birth. Furthermore, despite accuracy variations, both ChatGPT and Gemini responses exceeded the recommended 8th-grade readability level (Flesch-Kincaid grade 10.2–11.4), potentially compromising patient comprehension. These findings underscore the necessity for clinician verification of LLM outputs to ensure both interpretability and clinical validity. This comparative study provides crucial insights into model-specific strengths and limitations for evidence-based clinical implementation.

Operational challenges of ChatGPT in clinical laboratory interpretation

The European Federation of Clinical Chemistry and Laboratory Medicine (EFLM) Working Group evaluation reveals significant limitations in ChatGPT’s clinical utility despite its ability to generate “broadly correct and safety-relevant” laboratory result analyses [44]. Notably, ChatGPT fails to identify subclinical disease markers within normal reference ranges – for instance, it overlooks that elevated GGT may not conclusively indicate hepatic injury, and normal leukocyte subpopulation distributions don’t guarantee immune system integrity. This parameter-isolated analysis risks missing early pathology indicators. Critical deficiencies emerge in pre-analytical factor recognition. While ChatGPT appropriately flags diabetes risk with elevated glucose and HbA1c, it neglects crucial pre-collection variables (e.g., fasting status) when interpreting discordant results (elevated glucose with normal HbA1c). Moreover, ChatGPT demonstrates inadequate synthesis of multi-analyte profiles – interpretations of liver function markers (ALT, AST, bilirubin, GGT) lack essential clinical context integration required for comprehensive diagnostic reasoning [44]. Furthermore, ChatGPT fails to thoroughly analyze the health risks associated with test results. For example, in cases of severe anemia or lipid profile abnormalities, it merely advises patients to consult a doctor but does not clearly inform them of the potential clinical severity of these conditions, which could pose significant threats to patients’ health. Lastly, ChatGPT is unable to effectively distinguish between outliers and critical clinical threshold values. This limitation could potentially lead to medical errors in critical clinical scenarios. The inability to differentiate between abnormal values and emergency-level clinical values is particularly concerning, as such distinctions are crucial for timely and appropriate medical interventions.

Diversity and variability of chatbots

In the current environment of diverse chatbot ecosystems, the variability among models and the challenges they pose in cross-study comparisons have become particularly significant. Generative AI models vary greatly in architecture, settings, and objectives, resulting in substantial differences in their performance and quality when generating content. Their architectural design and capabilities directly influence the accuracy and effectiveness of their outputs. Notably, there are significant differences in performance across different generative AI models for specific tasks. For instance, Bing has demonstrated the highest accuracy and specificity in predicting drug interactions, significantly outperforming Bard and ChatGPT-4, highlighting that certain models may have inherent advantages in specific application scenarios [45].

Furthermore, variability within the same model cannot be overlooked. The same model may produce inconsistent results under different configurations, input data, or generation strategies, especially in the public health domain, where such variability could have significant implications for decision-making [45], 46]. Differences in model performance not only affect the effectiveness of information dissemination but also directly influence user experience and satisfaction. For example, Bard has been found to provide the most easily understandable information regarding rhinoplasty, followed by ChatGPT and Bing [47]. Therefore, researchers and users should exercise careful selection when applying generative AI models to ensure their alignment with specific tasks and user needs.

Risks associated with updates to chatbot versions

While continuous updates to chatbots aim to enhance their capabilities and performance, they also pose challenges in terms of data consistency, data quality, and reliability. Model updates may introduce new knowledge while risking reference to outdated data, particularly in rapidly evolving fields such as medicine. For example, the data used to train ChatGPT did not include the latest blood lead reference value of 3.5 μg/dL from the CDC for children. In contrast, CopyAI was able to provide accurate blood lead reference values, indicating that differences in training data updates across chatbots may lead to inconsistent judgment results [48]. Furthermore, updates do not always result in improved performance, as accuracy in certain tasks may decline. A study conducted by Stanford University and the University of California, Berkeley, found that while GPT-4.0 can provide more comprehensive information in certain areas, its accuracy in tasks such as mathematics, coding, and visual reasoning actually decreased [49]. This phenomenon underscores the importance of continuously evaluating the performance of LLMs, particularly in fields like laboratory medicine where accuracy is critical.

Challenges of consistency in content generation by generative AI models

The consistency of chatbot output in content generation deserves close attention, particularly in specialized fields such as medicine. Research indicates that minor differences in the wording of prompts or background information can lead to significant variations in generated content, potentially affecting reliability in practical applications [11], 50], 51]. A study by Kochanek and colleagues demonstrated that the consistency of responses from GPT-4.0 when answering the same question over four days was 85–88 %, respectively, highlighting the influence of uncertainties in content generation [52]. In the medical field, where high consistency in output is critical, variability in answers could pose notable risks. Further evaluation of the reproducibility and accuracy of ChatGPT in laboratory medicine is essential to ensure its reliability as a reference in medical applications [37].

Moreover, generative AI models can be influenced by cultural and socio-cultural biases inherent in their training data, leading to inconsistencies in generated content across different cultural or linguistic contexts. For instance, Wang and colleagues found that ChatGPT performed significantly better in English than in Chinese during the Taiwanese Pharmacy Licensing Examination [53]. Alfertshofer and colleagues further observed that ChatGPT exhibited substantial differences in performance across six countries in medical licensing examinations, highlighting the impact of national and linguistic factors [54].These findings underscore the need for deeper investigation and optimization of generative AI models in multilingual and cross-cultural settings to enhance their adaptability and accuracy in diverse scenarios.

Challenges in the quality of health information dissemination

Chatbots are playing an increasingly important role in the dissemination of health information, but the clarity, conciseness, and relevance of their responses require careful attention [37]. Clarity demands that the generated content is easy to understand, avoiding the use of complex medical jargon so that the general public can grasp the key points effortlessly. Simultaneously, conciseness requires straightforward communication without unnecessary length, enhancing the appropriateness of health information and aiding the public in improving their health literacy, thereby enabling them to make informed health decisions. Additionally, the relevance of AI-generated content is crucial. Accurate and relevant information helps prevent misunderstandings, particularly in scenarios where patients cannot access face-to-face explanations from doctors, avoiding unnecessary anxiety or health risks caused by misinformation [55]. Prioritizing the relevance of information prevents information overload, ensuring the clear delivery of critical health messages. Information unrelated to health inquiries may disrupt people’s understanding of their health conditions, increasing their confusion when dealing with personal health issues.

However, the dissemination of health information via AI must balance the differing needs of patients and medical professionals. Patients typically prefer short, direct answers to quickly comprehend key information, while medical professionals tend to favor detailed and comprehensive responses. This divergence of needs creates a double-edged sword: while abundant information aids professionals in thorough analysis, it may overwhelm ordinary users, leading to comprehension difficulties and anxiety. Regardless of the presentation method, ensuring the completeness of information is paramount. The absence of necessary details may lead to misdiagnosis or inaccurate self-diagnosis, exacerbating health risks. For instance, patients with limited health literacy may struggle to interpret laboratory test results accurately, such as identifying high cholesterol levels in lipid panels, potentially leading to severe health issues like heart disease or stroke [44]. Therefore, enhancing patients’ ability to understand laboratory test results is of utmost importance.

Risks of chatbots providing “hallucination” information

The application of chatbots in the dissemination of medical information is becoming increasingly widespread, but it has also raised significant concerns regarding the risk of misinformation. While generative AI tools like GPT offer certain advantages in providing information, their inability to explain decision-making processes hinders their ability to effectively identify and correct biases or errors within the model. In the medical field, AI-generated misinformation can lead to severe consequences, including incorrect self-diagnoses, delayed medical care, the spread of potentially harmful diseases, and a loss of public trust in healthcare professionals and institutions. For instance, GPT has provided inaccurate descriptions of FDA-approved high-sensitivity troponin point-of-care testing devices. Such “hallucination” responses are not isolated incidents, highlighting the limitations of generative AI models in ensuring the accuracy of information [56], 57]. Therefore, ensuring the accuracy, reliability, and trustworthiness of AI-generated medical information is of paramount importance. Developers must prioritize this objective during the design and training of AI models, as existing studies demonstrate that the generation of inaccurate information is a widespread phenomenon in such tools.

Legal and ethical issues in chatbot training data

Legal and ethical issues surrounding chatbot training data are critical concerns in the development of generative artificial intelligence (AI) models [58], [59], [60], [61]. These issues not only pertain to the origin of the data but also involve its usage and acceptability, directly impacting the transparency and credibility of the model. Data transparency is of particular importance, ensuring that users and researchers can understand the methods of data collection and utilization, thereby validating the reliability and scientific integrity of the model. Additionally, ethical concerns must not be overlooked, especially when using copyrighted materials and clinical data, as it is imperative to ensure legal and compliant data acquisition and usage while safeguarding the privacy rights of data providers.

Generative AI models like ChatGPT are trained on internet-based data, inevitably inheriting biases present in the training data and reflecting them in their outputs [62]. Studies have shown that LLMs exhibit biases related to gender, race, and other social factors, which not only compromise the fairness of model outputs but also exacerbate social inequalities. These biases are closely linked to issues such as data gaps, insufficient sample sizes, and inherent biases in foundational datasets [63]. If the training data is older, the potential biases and errors may be amplified in subsequent model outputs [64]. This phenomenon is particularly evident in existing records, especially in key areas such as randomized design, further highlighting the potential biases in query selection processes.

EU AI act and FDA clinical-decision-support (CDS) guidance

The European Union’s Artificial Intelligence Act (AI Act), effective from July 12, 2024, establishes the world’s first comprehensive regulatory framework for AI, prioritizing ethical development, fundamental rights protection, and human-centric innovation [65]. Adopting a risk-based approach, the Act categorizes AI systems into four tiers: unacceptable risk (banned), high-risk (e.g., medical diagnostics, subject to strict safety, transparency, and accountability requirements), general-purpose AI with systemic risk (transparency obligations), and low/no-risk (minimal regulation) [66]. In healthcare, stringent standards for high-risk medical AI emphasize rigorous testing, bias mitigation, and post-market monitoring, though gaps persist in addressing patient rights and accountability for low-risk applications. Challenges include uneven implementation across member states, ambiguous liability frameworks, and evolving limitations in detecting AI-generated content [67]. The Act also addresses generative AI’s copyright concerns, mandating transparency in training data and safeguarding intellectual property, while exempting pre-market research [66]. Extraterritorial impacts akin to the GDPR is expected, reshaping global AI practices and necessitating AI ethics education. Balancing regulation with innovation, the AI Act aims to foster trustworthy AI but requires clearer guidelines, stakeholder collaboration, and tailored healthcare standards to ensure equitable access and patient safety.

Currently, no LLM is authorized by the FDA as a CDS device. Research has evaluated whether LLMs can adhere to regulations in the context of clinical emergencies. The findings revealed that while LLMs provided appropriate preventive care recommendations, 100 % of responses from GPT-4 and 52 % from Llama-3 were noncompliant in time-critical scenarios, yielding device-like support by suggesting specific diagnoses and treatments [68]. Despite prompts based on FDA guidance, compliance was not consistently achieved. These results highlight the urgent need for new regulatory frameworks to address the unique challenges posed by generative AI in healthcare, as current guidelines are insufficient to prevent unauthorized device-like outputs from LLMs.