Potentials and pitfalls of ChatGPT and natural-language artificial intelligence models for the understanding of laboratory medicine test results. An assessment by the European Federation of Clinical Chemistry and Laboratory Medicine (EFLM) Working Group on Artificial Intelligence (WG-AI)

Introduction

Laboratory medicine has always struggled with the fact that although it contributes to the majority of medical decisions with its test results [1], [2], [3], it is rarely able to interpret these results in the context of the patient’s clinic, especially when this information is usually not provided by the requesting physicians or is not directly consultable in the laboratory. Therefore, numeric values are usually only provided to the clinician, which is responsible for their proper interpretation. However, if this interpretation is not communicated to the patient, he or she will be left alone with a laboratory report and no clear guidance on how to interpret it. Consequently, many patients turn to the information available on the Internet, commonly referred to as “Dr. Google.” Recently, a freely available Artificial Intelligence (AI) chatbot called “Chatbot Generative pre-trained Transformer” (ChatGPT for short) was made available to the public, which simulates human-like communication [4]. This chatbot has been demonstrated to pass the United States Medical Licensing Exam (USMLE), a set of three standardized tests of expert-level knowledge, which are required for medical licensure in this country [5]. Additionally, it has been speculated that ChatGPT can be of aid for clinicians, for example by providing clinical decision support, or by offering support for differential diagnosis or preliminary treatment plans [6]. In a recent study, it was shown that the capability of ChatGPT to solve higher-order reasoning questions in pathology had a relational level of accuracy [7].

With currently half a billion users, it is more than likely that patients are already using this tool to have their laboratory results translated into layman’s terms, especially when considering that many web online search engines are ready to integrate (or have already integrated) AI-based chatbots. Since ChatGPT’s ability to interpret laboratory medical test results has not yet been tested, the European Federation of Clinical Chemistry and Laboratory Medicine (EFLM) Working Group on Artificial Intelligence (WG-AI) aimed to take a closer look at ChatGPT in this regard.

We focused on a series of fictional (but realistic) clinical cases, each with different pathological conditions that were asked to be interpreted by ChatGPT v4.0. The obtained statements of the AI tool were checked for their correctness, patient safety, helpfullness and relevance.

To our knowledge this is the first attempt to inspect the ChatGPT ability to evaluate laboratory results, simulating a real-life scenario by imitating patients in need of a medical interpretation.

Materials and methods

We have decided to consider the use case of a patient receiving his/her laboratory tests results after a routine check-up at his/her general practitioner (GP), and not a specific diagnostic question to ChatGPT from a medical doctor. In this case, and under the assumption that the patient has not yet discussed the results with his/her GP, it is likely that the patient has no other support for the interpretation of the results than the limited indications in the results themselves (e.g. written comments in the report or the observation that a certain value is in or out of the reference range). The patient might therefore seek a more informed opinion, either by asking an acquaintance or, as is often the case, to turn to internet searches [8]. We decided to focus on this use case, because we believe it could be the most frequently encountered, considering also that ChatGPT users continue to grow in an exponential fashion and the reputation for trustworthiness set to grow similarly. Other uses of ChatGPT, including those by a healthcare professional and specialist to receive interpretation support, would be contrary to the intended purpose of the system and deontologically problematic. Moreover, the aforementioned use case may have a relevant impact on both patient safety and the appropriateness of access to primary (i.e., the GP) or secondary healthcare services.

Thus, in order to define the set of laboratory exams to include in the study, a preliminary round table was performed across participants of the study (WG-AI members), to define laboratory parameters, fitting the patient-oriented use case. Firstly, a standard expanded set of exams was collected by allowing WG-AI members to define laboratory exams as the most requested by general practitioners (GPs). The agreement over the standard expanded set of tests was finally achieved by selecting the most common laboratory parameters (common set of parameters), which included the following set of exams: complete blood count (CBC) with differential (leucocyte subsets), gamma glutamyl transferase (GGT), glucose, total, high-density lipoprotein (HDL)-, and low-density lipoprotein (LDL)-cholesterol, creatinine, aspartate aminotransferase (AST), alanine aminotransferase (ALT) and total bilirubin (Bilirubin) levels. In addition to this series of common laboratory parameters, a second series made of additional tests were defined, including ferritin, prostate-specific antigen (PSA), thyroid-stimulating hormone (TSH), free thyroxine (FT4), alkaline phosphatase, activated partial thromboplastin time (aPTT), prothrombin time (PT) and glycated hemoglobin (HbA_1c). Tests of this second series were provided to ChatGPT only if they were of added value for the specific case.

Following, 10 plausible fictional clinical cases were defined by WG-AI members in light of the above use case, encompassing a common set of laboratory parameters, their results, the reference ranges, the age and biological sex of the patient. To make the definition of the clinical cases more flexible, a maximum of three additional parameters, chosen from the second series of tests, were allowed to be specified. These additional tests were considered important either for better defining the specific clinical context or for testing ChatGTP abilities (Supplementary Table 1).

Prompts were defined by one author (FC), an human-computer interaction (HCI) researcher specialized in human-AI interaction [9]. The term “prompt” refers to the few lines of instructions given to the chatbot, in order to elicit the specific response from it. In this study, the prompt was designed by following the state-of-the-art heuristics to mitigate the risk of hallucinations (that is incoherent and inaccurate responses) and of receiving only shallow recommendations to simply consult a doctor:

“Act as a personal assistant who is a laboratory medicine expert and can interpret lab exam results and help patients understand them. I will give you a list of test results, their unit of measure, reference intervals, and relevant information about the individual, such as age and sex. In particular, the result pattern will be the following one: “Test Name (unit of measure): Test Result (Reference range)”. Your task is to interpret these results both collectively and individually to inform the person, and raise alerts if values are out of normal ranges and advice for a referral if this is the case. Be as evidence-based as possible. If you are unable to interpret the results, or a single result, simply acknowledge that. If you recommend that the person consults a medical doctor, explain your reasoning for doing so. I report the case in what follows:”

In order to simulate the patient use case on routine laboratory results (therefore excluding cases where a specific diagnostic inquiry was performed), no other context information (e.g., in regards to pre-existing chronic conditions or pre-analytical events) was given to ChatGPT.

In addition, two cases were provided to ChatGPT with a much simpler prompt, simulating a layperson request:

“Please, help me understand these blood exams. I’m a (biologic sex), XX year old. Should I call the doctor or be worried about them?”

One of the cases was proposed to ChatGPT two times (in different use sessions) with the same prompt to check whether the associated responses changed significantly from the semantics or pragmatics point of view.

The responses produced by ChatGPT (embedding the GPT ver. 4, 03/28) were collected in different chats (i.e., use sessions) and user sessions to avoid data leakage across the interpretations of different cases. Seven members of the EFLM WG-AI group, all laboratory specialists, independently evaluated the responses with respect to relevance, correctness, helpfulness and safety, by rating response along each dimension on an six-option ordinal scale, i.e., a semantic differential ranging from 1 (very low) to 6 (very high). The cases, the responses and the scales were presented to the raters involved by means of an online multi-page questionnaire developed on the Limesurvey platform (ver. 5.5), where each case was presented in a different page. The quality dimensions of relevance, correctness, helpfulness, and safety are defined in Table 1, in the same way as they were explained to the raters before they undertook the evaluation.

Results

ChatGPT results were recorded and used to generate the evaluation with respect to relevance, correctness, helpfulness and safety, on a six-option ordinal scale (Supplementary Table 2). Figure 1 shows the obtained results. No significant difference, with respect to any quality dimensions, was found among the ratings of the responses generated by ChatGPT after the two alternative prompts, that is the optimized and the simpler one, for case 1 and 8 (p-values after a Mann-Whitney test were ranging from 0.7 to 0.4, with effect sizes ranging from 0.1 to 0.2). In addition, at qualitative inspection, no relevant (that is semantically substantial) difference was also found between the two responses provided by ChatGPT to the same prompt, in two distinct chats, for the same case (which nevertheless were different with regard to wording and outcome presentation). In regard to the ordinal ratings associated with the quality dimensions above, median (and the corresponding interquartile range – IQR) values of ratings were, respectively, six (from 4 to 6) for relevance, five (from 3 to 6) for correctness, four (from 2 to 5) for helpfulness, and six (from 5 to 6) for safety. All confidence intervals of the medians were above the value of 3; the proportion of positive ratings (that is above three) was significantly different from the proportion of negative ratings (that is below four) for all dimensions except helpfulness (i.e. relevance, correctness and safety), for which we observed a 56 % of positive responses (95 % confidence interval: 40–60 %). No response received a rating of one (lowest) with respect to safety, but case 6, 9 and 10 received one rating of two each, and case 8 received four ratings of two (two ratings associated with the optimized prompt and two ratings associated with the layperson prompt). Responses to case 8 also received two ratings of three for safety (Table 2).

Figure 1:

Box plot describing the results of the assessment made by the seven members of the EFLM WG-AI group, all laboratory specialists, who independently evaluated the ChatGPT responses with respect to relevance, correctness, helpfulness and safety.

Discussion

Since many years, AI has entered the medical field, reaching relevant achievements in certain domains [10]. Most radiological departments already work with image recognition software based on AI, while AI has not yet become widely adopted in laboratory medicine. One reason for this is the fact that AI models need to consider a multitude of quantitative and qualitative variables (e.g. symptoms, physical examination, medical history, previous diseases, test result, medication, pre-analytics, …) for a valid interpretation of a full laboratory test report [11, 12]. Other reasons could be the limited informatics skills and knowledge of laboratory specialists, as well as the lack of adequate infrastructural support for easy retrieval from laboratory information systems (LIS) data [13]. However, likely, AI will probably also be used in medical laboratories in the near future [14], [15], [16]. A first step in this direction is made by the recently released chatbot ChatGPT, a natural language processing (NLP) model, which offers an almost infinite number of application possibilities, including creative fun, brainstorming, writing code and presentations, conducting literature presentations and reviews, writing research manuscripts and writing coursework or exams [17].

ChatGPT uses NLP techniques to interpret and train on a massive amount of online text data, including books, articles, and other sources of information [18]. In the last months, ChatGPT has been demonstrated to be a tool used for writing scientific articles and/or abstracts, in the search for specific literature, or for summarizing data or information [18], serving more or less as an interactive encyclopedia [5]. This contribution has been paradoxically so important that some journal editors, researchers and publishers are now debating the place of such AI tools in the published literature, and whether it’s appropriate to cite the bot as an author [19].

Regardless of the fact that ChatGPT has not been trained on medical data specifically, but is intended to undertake a human-like conversation, the utility of the tool for medical purposes has already being tested [5], [6], [7]. Currently, the discussion about the benefits, potential hazards [20], possible areas of applications or needed regulations of this tool [21] is ongoing. In the meanwhile, patients who want to decipher the meaning of their “cryptic” laboratory report, are most certainly turning to ChatGPT as an interpretative aid. This is not too surprising, given the fact that laboratory reports are mostly not tailored for patients [22]. Importantly, if the use of such chatbots will include future scenarios, as those depicted above, these tools would fall under the Medical Device Directive, classified at least into Class IIa, or the medium risk profile, thus requiring conformity assessment by a notified body. Furthermore, if the intended purpose of a chatbot mentions patients among its prospective users (e.g., for prevention-related objectives), it is important that manufacturers also envision the patients’ misuses. In light of its potential, the EFLM WG-AI has taken it upon itself to test the capabilities of ChatGPT (a conversational agent still in controlled-access demo as of writing) to interpret clinical laboratory values. For this purpose, the ChatGPT was fed with 10 fictional and yet realistic cases containing some of the common laboratory parameters and their results, the reference ranges, the age and biologic sex of the patient (Supplementary Table 1). The EFLM WG-AI members, all laboratory specialists, independently evaluated the interpretations of ChatGPT with respect to safety, helpfulness, correctness, and relevance.

The AI tool knew all specified parameters and presented possible causes in case of deviation from the reference value. Every parameter was evaluated individually including a final overall statement of the findings (Figure 2). It must be noted, however, that for each value, the corresponding reference value was given. In a real scenario, reference values would be probably omitted by patients’ queries, and it is therefore all the more important to strive for standardization/harmonization of laboratory parameter reference values wherever possible. Thus, the use of such types of devices for interpreting laboratory tests results, strongly support the importance of international ongoing initiatives on harmonization nomenclature, units and reference intervals [23].

Figure 2:

Output of ChatGPT obtained with this study for the specific case 4.

In some cases, the interpretation of normal results, in terms of the suspected underlying diseases, were not fully correct. For example, an increase in GGT alone was considered a sign of liver injury, normal platelet count was sometimes associated with normal coagulation, and normal distribution of leukocyte subpopulations was regarded as a correctly functioning immune system. In a case with increased glucose and HbA_1c levels, ChatGPT correctly suspected a possible diabetic condition and recommended to consult a medical doctor for further investigation (Supplementary Table 2). However, in other cases with increased glucose levels and normal HbA_1c value, the recommendation remained similar, not taking pre-analytical issues into account, such as the possibility of a non-fasting sample (Supplementary Table 2).

Another major finding was that the chatbot was unable of synoptically interpreting all coherent laboratory test results. Some parameters (e.g. ALT and AST) were mentioned in conjunction to each other, while others (e.g. Bilirubin or GGT) were treated separately thereof. In addition, we found that even though the system mentions results to be slightly or severely in-or decreased, its interpretation did not seem to be influenced by this fact as for example a severe anemia (Hb [g/L]: 77 (128–168); Hct [%]: 24.3 (38.4–50.4)) was weighed equally as the deviated lipid profile of this patient, recommending that seeing a doctor would be a “good idea”. Therefore, we could state that chatbot was unable to discriminate between abnormal values (defined as a test result above or below the upper or lower limits of the reference interval) and critical values, being potentially harmful situations to be immediately communicated to physicians in order to prevent potential patients’ harm [24]. Similarly, ChatGTP never takes care of the fact that for certain laboratory parameters, pre-analytical conditions should be taken into account for interpretation of results (e.g. fasting for glucose, hepatic enzymes, etc. …) [25], [26], [27].

Overall, ChatGPT is very cautious in its statement, even when provided with tailored prompts. In each of the cases, a visit to the physician was recommended and no recommendation for follow-up diagnosis or therapy was ever suggested. In order to test for the appropriateness of these recommendations for doctors’ visits, the laboratory experts gave individual case-related interpretations, which were then compared with those of ChatGPT (Supplementary Table 1, Figure 1). Marked differences were observed across the cases, thereby illustrating the suboptimal recommendations of ChatGPT.

To test for continuity of the chatbots answers, we evaluate the ChatGPT response using twice the same prompt for the case for which its output had been found more helpful. In the two responses generated, we found a high similarity in terms of semantics, overall interpretation and recommendations: this suggests an adequate consistency among responses. As the prompt that was used to feed ChatGPT with the fictional cases was designed according to current prompt engineering best practices, we additionally tested the chatbots output using a naive prompt, intended to simulate the request by a layperson. The response text changed slightly, but the overall interpretation was almost equivalent, suggesting adequate consistency among stochastically-generated responses.

Apart from all the drawbacks described, ChatGPT, being a conversational tool rather than a medical adviser or decision aid, demonstrated an impressive capability to detect and interpret altered values, even if did it only on a parameter-by-parameter basis, as a probable consequence of not having been trained, nor optimised, on laboratory medicine reports.

Overall, the answers provided by ChatGPT were rated by the EFLM WG-AI members as mostly correct, relevant and safe but mostly not too helpful. For instance, statements like “The Hct value is low, which suggests that the proportion of blood volume composed of red blood cells is lower than normal” or “Glucose is low, which might indicate hypoglycemia.” surely are correct, relevant and safe, but not really helpful. In other instances ChatGPT was more precise, suggesting microcytic anemia in case of low Hb- and MCV-values, and again in others was misleading like when stating “MCV, MCH, MCHC, Hb, and Hct are within the normal reference ranges, which indicates normal red blood cell function and structure”, disregarding the possibility of Hemoglobinopathies. Table 3 reports a summary of the findings of EFLM WG-AI members, listed as pros and cons.

The capabilities and usefulness of ChatGPT depend on what one compares it to as well as on user expectations. Compared to a pure Google search, ChatGPT has clear advantages (although a much higher carbon footprint, if informal estimates of nearly nine times the energy consumption are correct) because it can see and partially understand some connections between the prompted test results. Also, searching laboratory test results with search engines could be time-consuming. Indeed, search engines results, being lists of hyperlinks redirecting to other web pages, necessitate of continuing web surfing until definitive information are gathered, sometimes leading patients to inadequately written or reviewed health information [28]. Furthermore, it is possible to have conversations with the AI and ask follow-up questions, which are generally considered a good practice to obtain optimal results. On the other hand, if one compares ChatGPT with physicians or laboratory medicine specialists, its responses have major disadvantages (see Table 3 “Cons”). However, this is not surprising, since ChatGPT is not trained for this purpose, nor intended for this use.

Regarding the discussion on whether or not AI will replace laboratory specialists, we want to state the following: the risk that AI tools will make cognitive human work superfluous is partly justified, but in our view, it applies only to repetitive tasks. Current evidence suggests that creative problem solving remains a deeply human capability and that the combination of AI and human experts’ skills either adds up to a greater effect or, in the case of laboratory medicine, accelerates personalized medicine, as laboratory specialists can deal more intensively with individual complex patients [29], [30], [31].

This study presents some limitations. Our investigation merely represents a first insight into the possibilities of ChatGPT as a reader of laboratory test results, aimed at simply being the basis for writing this paper. For a statistical evaluation, a much larger number of cases, and responses (on the basis of both different and equal prompts) would be necessary. However, the cases considered were purposely chosen to be representative of a large range of cases, especially among routine use cases. Secondly, prompts and cases were prepared in English language only. As a major pitfall in using this chatbot, it should be considered that the richness of ChatGPT’s response and the intelligibility of its writing in some languages, including French and Arabic, are notably inferior to those in English [32]. Finally, perceived helpfulness would presumably be higher among less-experienced health care employees: an assumption that would need to be confirmed with an appropriate follow-up study.

Conclusions/outlook

ChatGPT can currently be considered a tool capable of detecting anomalies in laboratory parameters and interpreting deviations from the reference value on a test-to-test basis. The system is quite superficial and provides generic answers on complex cases where multiple dependencies among the results should be instead considered and patterns recognized to reach an accurate interpretation. Moreover, the chatbot has been extremely reluctant to make definitive statements about the overall findings or recommendations about a further course of action in general, with the exception of recommending to consult a physician or book a visit.

The chatbot has impressively demonstrated that AI is capable of analyzing medical data, even if it has not been specifically trained or fine-tuned to do so. It remains an open question whether conversational agents that embed large language models trained on medical data, such as BioGPT or Med-PaLM, or that are fine-tuned on laboratory reports, may not soon achieve a utility comparable to consulting a physician, thus constituting an additional and reliable filter for more appropriate use of health care services. Given the pace at which the development of these systems is proceeding, and the expectations of its ever-growing user base, we are convinced that it is not so much a question of “if this will happen”, but rather “when”. We therefore strongly recommend that all laboratory specialists familiarize themselves with this (and similar) tool(s) and acquire an informed attitude toward this potentially disruptive phase of change in medical diagnostics.