Evaluating the Educational Value of Artificial Intelligence in Pharmacotherapy Education

ABSTRACT

Final‐year medical students frequently make prescribing errors, partly because they have had insufficient training in therapeutic reasoning. Students need to understand the underlying rationale behind a drug choice in order to subsequently apply that knowledge critically and independently. Artificial intelligence (AI) models are becoming increasingly popular among students for therapeutic case solving, but the educational value of these models remains unclear. Therefore, the reasoning process of ChatGPT 3.5 (a general AI model) and EvidenceHunt basic (a specialized medical AI model) was compared with that of a medical specialist/clinical pharmacologist, according to the WHO 6‐step. Although both AI models proposed reasonable treatment options, notable differences from those of the medical specialist were observed. The AI models tended to focus primarily on pharmacological recommendations for the main clinical problem, giving less consideration to broader contextual factors, such as the severity of illness. Some inaccuracies and illogical therapeutic alternatives were also identified. While AI models, such as ChatGPT and EvidenceHunt, show potential to improve education in therapeutic reasoning by suggesting plausible treatment options, their reasoning remains limited. They often justify rather than critically evaluate treatment choices, which restricts their depth, especially when it comes to considering alternative therapies and complex patient factors. For AI to be of real educational benefit, students will need to develop reflective skills so that they can critically appraise AI recommendations instead of passively accepting them. Teachers can encourage this by asking students to explain their treatment choices, thereby helping them to strengthen these skills.

1. Introduction

Final‐year medical students often feel insecure about prescribing medicines and tend to make a relatively large number of prescribing errors, some of which are potentially harmful [1, 2]. One contributing factor is that medical education tends to emphasize diagnostic skills rather than therapeutic decision‐making [3, 4]. Medical curricula often provide extensive training in history taking, physical examination, and making differential diagnoses, but provide comparatively little training in selecting and justifying appropriate treatments.

This process, known as therapeutic reasoning, refers to the cognitive steps involved in translating a diagnosis into a rational treatment plan [5, 6]. This involves balancing multiple, sometimes competing considerations, such as clinical guidelines, pharmacological properties of drugs, comorbidities, patient preferences, and contextual factors [5, 7, 8, 9, 10, 11]. Moreover, there is often no single “correct” treatment option, which makes therapeutic reasoning inherently complex, less structured, and context‐dependent [5]. The development of strong therapeutic reasoning skills is therefore an essential, yet underemphasized, component of medical education.

Teaching therapeutic reasoning is challenging. Students, lacking experience, must rely on slower, analytical type 2 reasoning to construct rational treatment plans [6, 12, 13, 14, 15]. Clinician supervisors, on the other hand, often rely on fast, unconscious, intuitive, experience‐based, type 1 thinking, which enables efficient decision‐making but is difficult to articulate explicitly, making it challenging for students to learn therapeutic reasoning from them [6, 7, 16]. Time pressures and safety concerns in clinical environments further limit opportunities for students to engage in reflective discussions about prescribing and to effectively apply type 2 thinking [17, 18]. As a result, students frequently resort to mimicking their supervisors' decisions without fully understanding the underlying rationale, which may lead to incorrect or incomplete therapy scripts [7]. As therapy scripts are built through experience, if they are based on flawed reasoning, students may unknowingly reuse them inappropriately in future cases [7].

To support the development of rational prescribing skills, cognitive forcing strategies, such as structured approaches, may help students to optimize their reasoning [19]. In response to this, many medical curricula have successfully adopted the World Health Organization (WHO) 6‐step model to provide a structured approach to the prescribing process [20, 21, 22, 23, 24, 25]. This structured framework guides learners through the therapeutic reasoning process in a stepwise, deliberate manner: (1) define the patient's problem; (2) specify the therapeutic objective; (3a) select a standard treatment; (3b) verify its suitability for the individual patient; (4) start treatment; (5) provide patient information; and (6) monitor treatment [26]. This model, in combination with case‐based education, not only facilitates a deeper understanding during teaching sessions and assessments, but also helps students build more refined therapy scripts that can later support more intuitive type 1 decision‐making grounded in solid understanding [19, 27, 28, 29].

In recent years, however, a new development has emerged in the educational landscape: the growing use of artificial intelligence (AI) models. Large language models, such as ChatGPT, are increasingly used by students to solve therapeutic case studies, prepare for exams, or rehearse Objective Structured Clinical Examination (OSCEs) using simulated scenarios [30, 31, 32, 33, 34, 35, 36]. This shift raises two pressing educational questions: to what extent are current AI systems capable of generating rational, individualized treatment plans? And do these models support meaningful learning by promoting understanding of the underlying reasoning, or do they merely encourage superficial copying?

For AI models to be of real educational value, they must do more than generate correct treatment suggestions. They must also provide transparent, step‐by‐step reasoning that helps students grasp the why behind each decision. In other words, AI should function not just as a shortcut to an answer, but as a catalyst for reflective learning. Therefore, the aim of this study is to assess whether two large language models (LLMs) can generate guideline‐consistent explanations for pharmacotherapeutic decisions using the WHO 6‐step model and thereby its educational value.

2. Methods

2.1. Study Design

This study assessed the educational value of AI by comparing the clinical decision‐making of two AI‐based systems, namely ChatGPT (a general‐purpose AI model developed by OpenAI) and EvidenceHunt (a specialized medical AI platform that delivers evidence‐based recommendations via systematic literature searches [37]) with that of a medical specialist, using two medical cases. The medical specialist's evaluation served as the human benchmark in this study. For reproducibility, the free versions of both ChatGPT (GPT 3.5) and Evidencehunt (Basic) were used. The medical specialist is an internist, clinical pharmacologist, and professor of clinical pharmacology at a large academic hospital in the Netherlands.

2.2. Task Description

Two clinical cases were entered into ChatGPT and EvidenceHunt. One case focused on hypertension and the other on a urinary tract infection, which are common issues in internal medicine. Each entity was tasked with selecting an appropriate treatment for the two cases, with emphasis on providing a thorough rationale for their recommendations based on the WHO 6‐step model [26], to standardize the comparison. The following prompt was used: “Justify, using the WHO 6‐step method, which medication should be prescribed in this case. Be thorough in your reasoning, so that a beginning prescriber can learn from it and be able to handle a similar case independently in the future”. The prompt used for this task emphasized the importance of completeness of the argumentation, balancing structure with accessibility to ensure that even users with limited experience in AI prompting could achieve meaningful results. The same instructions were also provided to the medical specialist.

2.3. Comparative Analysis

The responses generated by ChatGPT and EvidenceHunt were then systematically compared with the responses of the medical specialist across each step of the WHO 6‐step by the authors. They discussed each response to reach expert consensus. In cases where full agreement was not immediately reached, the majority opinion guided the final evaluation. A qualitative analysis of the findings was performed and the results were structured per step and compared based on adherence to the WHO model, clarity and depth of argumentation, and overall quality of therapeutic reasoning.

3. Results

The cases are summarized in Table 1 and the responses of the different entities in Table 2. The complete cases and corresponding output of the AI models and medical specialist are shown in Appendix A.

TABLE 1.

Overview of the cases.

Case A

A 67‐year‐old male, with hypertension, type 2 diabetes, hyperlipidaemia, coronary artery disease, chronic kidney disease (eGFR 50 mL/min/1.73 m2; reference value: > 90 mL/min/1,73 m2), and intermittent claudication. When examined, his blood pressure was 160/100 mmHg. He drank 10 units of alcohol per week. His medication was metformin 500 mg 2× daily, simvastatin 40 mg once daily, acetylsalicylic acid 80 mg once daily, and lisinopril 10 mg once daily.

Case B

A 72‐year‐old female, with type 2 diabetes treated with metformin, insulin, and atorvastatin. Her medical history revealed chronic kidney disease (eGFR 38 mL/min/1.73 m2; reference value: > 90 mL/min/1,73 m2); hypertension treated with lisinopril and amlodipine; history of recurrent urinary tract infections; and heart failure, treated with furosemide. She presented with lumbar region pain, fever, and urinalysis positive for nitrite and leukocyte, which could be caused by pyelonephritis.

TABLE 2.

Comparison of therapeutic reasoning between AI models and a medical specialist across the WHO 6‐step.

Step	ChatGPT	EvidenceHunt	Medical specialist
1	Identified main clinical problems; limited contextual considerations	Identified main problems; limited contextual considerations	Broad assessment including patient specific factors and comorbidities
2	Specific target for main problem; vague cardiovascular goals	Specific target for main problem; limited broader goals	Multimorbidity‐based treatment targets
3	Case 1: multiple antihypertensive classes; limited prioritization. Case 2: broad antibiotic options, some inappropriate	Case 1: focused on current ACE inhibitor and did not propose alternatives Case 2: ciprofloxacin only	Case 1: considered combinations, comorbidities and non‐pharmacological advice Case 2: guideline‐based stepwise antibiotic choice
4	Case 1: amlodipine, but no treatment for comorbidities Case 2: ceftriaxone, 7–10 days	Case 1: telmisartan based on unsupported concern for ACE inhibitor side effects Case 2: ciprofloxacin	Case 1: amlodipine, advice for cholesterol, diabetes and peripheral vascular disease Case 2: ceftriaxone 10–14 days (if admitted) or ciprofloxacin if outpatient
5	General lifestyle advice and adverse effects; outlined expected therapeutic effects	Lifestyle advice and adverse effects; emphasized medication adherence	Detailed non‐pharmacological advice in both cases; structured counseling including sequencing and adherence
6	Follow‐up relevant but non‐specific timing; emphasis on laboratory monitoring	Less frequent follow‐up; inclusion of comorbidities such as HbA1c; emphasis on laboratory monitoring	Emphasis on clinical symptoms and structured follow‐up of comorbidities

3.1. Step 1: Problem Definition

In Case A, both AI models identified the same key clinical problems as the medical specialist. ChatGPT and the medical specialist both emphasized the importance of preserving renal function, whereas EvidenceHunt did not address this. In addition, the medical specialist questioned whether blood pressure had been measured repeatedly or only once, and considered the possibility of poor adherence to therapy. He also reviewed the current treatment, including both pharmacological and non‐pharmacological interventions.

In case B, both AI models recognized the diagnosis and relevant comorbidities. However, the medical specialist also identified missing contextual information regarding the causative pathogen, such as whether the patient lived in a nursing home (increasing the likelihood of extended‐spectrum beta‐lactamase‐producing E. coli ) or which pathogens caused the earlier urinary tract infections.

3.2. Step 2: Specify the Therapeutic Objective

In case A, all entities gave a target blood pressure. The medical specialist additionally incorporated a broader cardiovascular risk management approach, including cholesterol, body mass index (BMI), and HbA1c. While ChatGPT and EvidenceHunt also aimed to minimize cardiovascular and renal risk, neither provided explicit target values. In case B, all entities agreed that the primary goal was treatment of the bacterial infection.

3.3. Step 3: Verify the Suitability of the P‐Drug

In case A, the medical specialist recommended both pharmacological and non‐pharmacological treatment options. Both AI models omitted non‐pharmacological advice. In line with the therapeutic objective, the medical specialist addressed not only hypertension, but also comorbid conditions such as claudication and diabetes. In contrast, the AI models focused exclusively on hypertension. EvidenceHunt addressed only the suitability of the current angiotensin‐converting enzyme (ACE) inhibitor and did not propose alternatives. ChatGPT discussed ACE inhibitors, calcium channel blockers, and thiazide diuretics, as did the specialist. However, ChatGPT focused primarily on safety, whereas the specialist considered both safety and efficacy—including the synergistic potential of combination therapies and their effects on comorbidities. ChatGPT also made some inaccurate statements in this case, such as suggesting that diuretics are ineffective for patients with an estimated glomerular filtration rate (eGFR) < 30 mL/min/1.73 m², which was not applicable in this case. While the AI models did not mention beta‐blockers as a potential option, the medical specialist evaluated their potential and explained why they were inappropriate in this case.

In case B, both AI models again failed to include non‐pharmacological recommendations. The medical specialist referred to clinical guidelines recommending a stepped‐care approach: starting with a preferred first‐line treatment and switching to alternatives if the initial therapy is contraindicated or ineffective. Neither AI model mentioned this approach. ChatGPT suggested four antibiotics: ciprofloxacin and ceftriaxone matched the specialist's recommendations, while piperacillin‐tazobactam and nitrofurantoin were not mentioned by the specialist. Cotrimoxazole, however, was notably absent. EvidenceHunt suggested only one oral antibiotic (ciprofloxacin) and did not provide intravenous options.

3.4. Step 4: Start Treatment

In case A, both AI models suggested treatment that was consistent with the specialist's recommendation, namely, amlodipine 5 mg. EvidenceHunt added telmisartan based on concerns about side effects associated with ACE inhibitors, particularly coughing, despite the absence of such symptoms in the case, which may reflect insufficient patient‐specific reasoning. Neither AI model suggested treatment for cholesterol, peripheral vascular disease, or diabetes.

In case B, the medical specialist identified the need to determine whether oral or intravenous treatment was appropriate, depending on the patient's clinical status. If the patient needed to be admitted to hospital, ceftriaxone 2 g once daily for 10–14 days would be preferred. If the patient was less ill and, for example, visited the general practitioner, then ciprofloxacin could be recommended. None of the AI models discussed this in detail. ChatGPT also started with ceftriaxone because of the severity of the infection, but the recommended treatment duration was short (only 7–10 days). Neither AI model addressed this distinction in detail. EvidenceHunt selected ciprofloxacin, consistent with the specialist's outpatient strategy.

3.5. Step 5: Give Information, Instructions, and Warnings

In case A, all entities acknowledged the importance of lifestyle changes and potential side effects such as oedema, although there were minor differences in the specifics. The AI models provided non‐pharmacological advice that was less detailed and specific than that of the medical specialist. The specialist also indicated the sequencing of medication initiation to aid identification of agents that caused side effects, if present, which was not done by any of the AI models.

In case B, all entities mentioned potential adverse effects. EvidenceHunt and the medical specialist emphasized the importance of medication adherence. ChatGPT additionally outlined the expected therapeutic effects. Both AI models omitted non‐pharmacological guidance.

3.6. Step 6: Monitor Treatment and Follow‐Up

In case A, all entities recommended follow‐up of blood pressure, renal function, and electrolytes within 2–4 weeks. The suggested intervals varied: ChatGPT became vague after the first month (“periodically”), while EvidenceHunt recommended less frequent monitoring (every 4–6 weeks). ChatGPT did not suggest follow‐up for diabetes or cholesterol. Neither AI model included glucose, creatine kinase, or urine sediment monitoring when indicated, for instance, for side effects. ChatGPT outlined a structured follow‐up plan similar to that of the medical specialist; EvidenceHunt did not.

In case B, all entities anticipated clinical improvement within 48 h; otherwise, re‐evaluation was necessary. Whereas the specialist focused on symptom resolution, both AI models emphasized laboratory monitoring. EvidenceHunt did not address adverse effects. The specialist also reflected on which side effects should not be emphasized unnecessarily. While all mentioned antibiotic resistance based on culture results, neither AI model considered anatomical abnormalities, such as urethral obstruction or hydronephrosis, as potential causes of non‐response.

3.7. Rationale for Not Initiating the Treatment Proposed by the Other Model

For case B, because the treatment recommendations of the two AI models and medical specialist did not align, we additionally asked the models to reflect on why they did not choose the recommendation of the other AI model. Both AI models cited concerns about drug interactions and adverse effects; however, some concerns appeared inappropriate or unsupported, such as a presumed interaction with ceftriaxone and warfarin, which was not a comedication in this case. ChatGPT also cited infection severity as a reason for not choosing ciprofloxacin, whereas according to the medical specialist, information about the severity was not clearly provided in the case. Other arguments included renal function and antimicrobial resistance. EvidenceHunt highlighted an increased risk of Clostridioides difficile infection and considered the balance between safety and efficacy in its rationale for not selecting ceftriaxone.

4. Discussion

This study evaluated the potential of AI to support therapeutic decision‐making and investigated its use as supplementary educational tool in medical training. Several notable findings emerged when comparing the decision‐making processes and treatment choices of AI models with those of a medical specialist.

4.1. Therapy Choices

Overall, the problem definition, therapeutic objectives, patient instructions, and monitoring advice for the primary diagnosis were largely consistent between the AI models and the medical specialist, although some nuanced differences were noted. However, there were distinctions in the comprehensiveness and contextual awareness of the clinical reasoning. Both AI models concentrated on the primary diagnosis (e.g., hypertension), often overlooking the broader context. In contrast, the medical specialist had a more holistic approach, integrating multiple clinical factors, such as cholesterol levels, diabetes status, and patient history. ChatGPT tended to provide more detailed explanations linking symptoms to diagnosis, whereas EvidenceHunt's responses were more consistent with clinical prioritization by focusing on pressing issues. Moreover, the two models often failed to perform a comprehensive review of the patient's current medication list, which is a critical step for identifying possible contraindications and drug interactions. While the pharmacological interventions were generally in line with those of the medical specialist, non‐pharmacological interventions were consistently under‐represented in AI‐generated output.

4.2. Argumentation for Treatment Choice

While the therapeutic recommendations themselves were often appropriate, the underlying reasoning and justification lacked the depth and precision expected in clinical practice. For example, although therapeutic goals were mentioned, they were often vague and lacked details such as specific timeframes or target values—elements that are emphasized in student education and outlined in the WHO Teacher's Guide to Good Prescribing [38].

In step three, ChatGPT offered multiple pharmacological options with basic pros and cons but did not clearly weigh them. EvidenceHunt suggested fewer options, focusing on current treatment appropriateness. Both models neglected factors such as treatment costs and environmental impact. Neither model was transparent about the clinical guidelines it used, which undermined the credibility of the recommendations. Although EvidenceHunt cited some literature, it was unclear why these sources were selected and how this information was weighed.

The reasoning presented in the AI‐generated output was not always valid. For instance, the AI models used some clinically irrelevant interactions as an argument for their treatment choice. Additionally, EvidenceHunt recommended initiating telmisartan without clarifying whether to discontinue the existing ACE inhibitor therapy, thereby risking adverse effects. Although ChatGPT recommended the same drug as the medical specialist, it omitted several important justifications—such as the impact on claudication—and included irrelevant arguments, such as concerns regarding the reduced efficacy of diuretics at renal function levels below 30 mL/min. This was not applicable here, as the patient's eGFR was 50 mL/min. These examples highlighted potentially unsafe recommendations arising from incomplete or inaccurate reasoning. A similar lack of depth was observed in case B. The medical specialist acknowledged that both treatments (ceftriaxone if the patient was hospitalized; ciprofloxacin if the patient attended her general practitioner) could be valid, but highlighted the difficulty of assessing disease severity from limited case details—a nuance absent in both AI models, which instead made assumptions that led to divergent recommendations. When the AI models were asked to explain their choices in comparison to those of the specialist, they overlooked other valid clinical considerations and both mentioned, for example, side effects as an important argument—each using this argument to reject the other's preferred medication, demonstrating a lack of objective consideration of all relevant clinical factors.

4.3. Shortcomings of AI

To understand these shortcomings, it is useful to consider why AI models may struggle with therapeutic reasoning. Previous research has shown that AI performs better in generating differential diagnoses than in formulating treatment plans [39]. This may be because diagnostic reasoning is more text‐ and pattern‐based, whereas therapeutic decisions involve context‐sensitive trade‐offs between multiple valid options. A study analyzing AI‐generated illness scripts identified multiple shortcomings, especially in treatment, including outdated advice or inappropriate antibiotic use [39]. Our findings are consistent with these observations: both AI models struggled to integrate complex clinical context and provided superficial therapeutic arguments, showing that AI can assist in diagnosis, but remains limited in nuanced treatment planning. Research on large language models and ethics also showed that AI often uses wrong or misleading arguments to justify unethical actions—as observed in an earlier study, when AI models justified non‐optimal treatment decisions [40]. Despite these shortcomings, ChatGPT performed at or near the pass threshold of the United States Medical Licensing Examination, demonstrating its potential usefulness in medical education and clinical support [41]. Because of this, AI is useful for knowledge retrieval and diagnosis, but is less reliable when it comes to therapeutic reasoning and practical skills. However, it is possible that asking targeted follow‐up questions could elicit more specific output.

4.4. Strengths and Limitations

As an explorative study with two cases, our aim was not to draw generalizable conclusions but to examine the reasoning and educational value of AI in specific therapeutic contexts. A key strength of this study lies in its comparison of two distinct AI models with a medical specialist across two realistic clinical scenarios, with a specific focus on the quality of therapeutic reasoning. The study demonstrates that AI could have a role in an educational context. However, it should be noted that the evaluation was based on qualitative expert consensus rather than formal numeric scoring, and no blinding was applied, which may limit reproducibility.

Additionally, comparisons were made against a single clinical pharmacologist, which may limit generalizability. However, it should be borne in mind that, when assessing the reasoning of AI models relative to that of specialists as gold standard, the reasoning process of medical specialists can also be influenced by cognitive bias and heuristics, such as anchoring or availability bias. A potential limitation is the user‐provided input to the AI models, which can introduce bias that is reflected in the output. The way clinical information is presented—or omitted—can significantly influence a model's response. This phenomenon has been reported in the literature on prompt sensitivity and input bias in large language models, emphasizing the importance of input framing and the need for cautious interpretation of AI‐generated clinical recommendations [42, 43].

Another important limitation is that the prompt used in this study was intentionally simple and mirrored real‐world use by novice AI users. However, LLM rationale quality is substantially influenced by the level of detail and structure within the prompt. Specific prompts, such as prompting for stepwise reasoning, justification of guideline choices, explanation of evidence sources, and consideration of alternatives can elicit more complete and nuanced clinical explanations. Such prompting strategies were not applied here, which means the results represent a realistic baseline of how LLMs are commonly used in practice. This also highlights that these tools can be readily used by a broad range of users, while more experienced prompting may further enhance their performance.

Another limitation is the scope of AI models included. Newer models such as Google's Med‐PaLM, which are specifically trained for medical tasks, may outperform the models assessed here. However, we did not use these models because they are currently not widely accessible. The AI models used in this study were freely accessible basic versions available at the time of data collection (ChatGPT‐3.5 and EvidenceHunt basic) and were selected to increase reproducibility and accessibility. However, it should be noted that more advanced models, such as OpenAI's GPT‐4 or GPT‐4o, may perform better in therapeutic reasoning [44]. The introduction of GPT‐5.2 thinking, with its focus on complex reasoning, indicates that prior models may not have been fully tailored for such tasks.

While the WHO 6‐step model provides a structured approach to prescribing, it may not fully capture the depth of therapeutic reasoning, such as weighing alternatives or citing evidence. This limitation should be considered when interpreting the AI's outputs. Future research could investigate prompt optimization and evaluate whether this improves the clinical relevance and reliability of AI‐generated responses. Incorporating more recent and specialized models may also yield more accurate and nuanced therapeutic reasoning.

5. Conclusion

In general, AI models can produce effective output, although they fall short in the area with the greatest educational potential: the rationale. Their rationale does not offer a systematic or balanced evaluation of options, but merely justifies the selected treatment. Though clinical context, goals, instructions, and monitoring are addressed adequately, both AI models lacked a structured, thorough analysis of alternatives. This selective focus and reliance on user input significantly limits their reliability and consistency. Without strong pharmacological knowledge, users risk overlooking critical gaps in AI‐generated reasoning.

In their current form, these freely available AI systems are not capable of genuine therapeutic reasoning and should not be considered a replacement for, or standalone educational tool in, medical training. The current inability of AI systems to engage in context‐sensitive therapeutic reasoning limits their value as an autonomous tool, yet they could provide valuable educational support when used prudently during medical training.

5.1. Implications for Teaching Practices

These findings highlight the importance of preparing medical students to work with generative artificial intelligence in a structured and critical manner, as these tools are increasingly becoming part of both medical education and clinical practice. While AI can produce seemingly effective output, this output often lacks a clear underlying explanation and may contain errors. To understand this, students should first be introduced to the fundamental principles underlying GenAI systems. LLMs are trained on large‐scale text datasets and generate probabilistic outputs rather than evidence‐based clinical judgments [45]. As a result, they may produce inaccurate or fabricated information (hallucinations) or generate responses that align too readily with user input (sycophancy) [46]. Awareness of these limitations, as well as the ethical and professional responsibilities associated with their use in clinical contexts, is essential for safe and responsible application.

Building on this foundational understanding, students should learn how to critically engage with AI‐generated output rather than accept it uncritically. This presents a key opportunity for teachers: by consistently asking students to provide a rationale for the treatment choice, they can encourage deeper reasoning and reduce uncritical reliance on AI. It may also help students recognize the limitations of AI‐generated output. Asking students to explain their therapeutic choices may improve therapeutic reasoning in two ways. First, when AI‐generated output is incorrect or incomplete, students must engage in more analytical, type 2 reasoning to identify and correct these errors. Second, understanding the rationale helps them construct robust therapy scripts, which can later support faster, intuitive type 1 reasoning. This is especially important given that type 2 reasoning requires substantial cognitive effort and is therefore not always feasible in busy clinical practice [18].

Experienced prescribers may struggle to explain their decision‐making processes to students, because type 1 reasoning is largely automatic. In such cases, AI can support learning by offering insights into the rationale behind therapeutic choices made by supervisors. Even when its reasoning is imperfect, AI may provide plausible explanations that make expert decisions more understandable to students. However, students need sufficient knowledge to critically evaluate this output, because otherwise there is a risk of over‐reliance and error propagation, as incorrect output occurs regularly [47, 48, 49, 50]. Therefore, teachers must remain vigilant to ensure AI is used appropriately and that ethical implications are carefully considered [51, 52].

To further support effective use of AI tools, training in prompt engineering should be incorporated into medical curricula. This involves teaching students how to formulate structured prompts, request stepwise reasoning, ask for justification of recommendations, and explicitly prompt for consideration of alternatives and guideline‐based evidence. Such skills may not only improve the quality of AI‐generated output but also reinforce students' own clinical reasoning processes. At the same time, students should be made aware of the diversity of AI tools currently used in healthcare and by patients, ranging from freely accessible general‐purpose systems to more specialized clinical decision support tools. Understanding differences in purpose, reliability, and regulatory context is essential for interpreting AI‐generated medical information and anticipating its influence in clinical encounters.

Beyond optimizing interaction with AI systems, it is equally important that students understand how these outputs fit within structured clinical reasoning processes, which include steps such as problem identification, information gathering, evaluation of benefits and risks, selection of the most appropriate option, and subsequent evaluation. Making this structured process explicit provides a reference point against which AI‐generated recommendations can be critically assessed and integrated into clinical reasoning.

Once this structured reasoning approach is established, students can be further trained to critically evaluate and verify AI‐generated outputs. Teachers can further strengthen students' critical thinking by integrating activities that involve fact‐checking the output of AI systems. This would help students develop the evaluative skills necessary for responsible and safe use of AI in both educational and clinical contexts. As AI becomes increasingly embedded in clinical and educational practice and widely used by students in their studies [53, 54], it is essential that they learn not only to recognize its limitations, but also to work with it effectively. Encouraging students to explain their rationale and fact‐check AI output fosters therapeutic reasoning skills AI cannot replace, while preparing them for responsible use in their future clinical work. To further support teachers in this, a method for teaching students how to use and fact‐check AI output in CPT education will be developed as part of The European Open Platform for Prescribing Education (EurOP²E) and made freely available via prescribingeducation.eu.

Author Contributions

Mariëlle G. Hartjes: writing – original draft, conceptualization. Michiel A. van Agtmael: conceptualization, writing – review and editing. Jelle Tichelaar: writing – original draft, conceptualization. Milan C. Richir: conceptualization, writing – review and editing. Michiel J. Bakkum: conceptualization, writing – review and editing.

Funding

The authors have nothing to report.

Conflicts of Interest

The authors declare no conflicts of interest.

Appendix A: Worked‐out cases A.

Evaluating the educational value of artificial intelligence in pharmacotherapy education.

Prompt: Justify, using the WHO 6‐step method, which medication should be prescribed in this case. Be thorough in your reasoning, so that a beginning prescriber can learn from it and be able to handle a similar case independently in the future.

Case 1

—Hypertension

Case presentation

Patient Overview

Mr. Jan de Vries is a 67‐year‐old male with a height of 1.78 m, a weight of 92 kg (BMI: 29.0 kg/m², classified as overweight), a blood pressure of 160/100 mmHg, and a pulse of 85 bpm.

Medical History:

• Hypertension (5 years)

• Type 2 Diabetes Mellitus (8 years)

• Hyperlipidemia (LDL‐cholesterol: 3.5 mmol/L)

• Coronary Artery Disease (PCI 3 years ago)

• Chronic Kidney Disease (CKD stage 3a, eGFR: 50 mL/min/1.73 m²)

• Intermittent Claudication (max walking distance: 200 m)

• Moderate alcohol use (10 drinks/week)

• Non‐smoker (quit 10 years ago)

Current Medication:

• Metformin 500 mg, 2× daily

• Simvastatin 40 mg, 1× daily

• Acetylsalicylic acid 80 mg, 1× daily

• Lisinopril 10 mg, 1× daily

Lab Values:

• HbA1c: 7.4% (57 mmol/mol) (reference: < 5.7% / <39 mmol/mol)

• Serum creatinine: 130 μmol/L (reference: 60–110 μmol/L)

• Potassium: 4.8 mmol/L (reference: 3.5–5.0 mmol/L)

• Sodium: 138 mmol/L (reference: 135–145 mmol/L)

• LDL‐cholesterol: 3.5 mmol/L (reference: < 3.0 mmol/L)

• HDL‐cholesterol: 1.0 mmol/L (reference < 1.0 mmol/L)

• Triglycerides: 2.0 mmol/L (reference: < 1.7 mmol/L fasting)

Lifestyle:

• Limited physical activity due to claudication.

• Mediterranean diet, not strictly low‐sodium.

Working diagnosis: hypertension

Step 1: define the patient's problem

ChatGPT:

Mr. de Vries is a 67‐year‐old man with the following health issues:

• Hypertension: Blood pressure of 160/100 mmHg, which is elevated despite treatment with lisinopril.

• Type 2 Diabetes Mellitus: Diabetes is a risk factor for cardiovascular complications and requires stringent blood pressure control.

• Chronic Kidney Disease (CKD Stage 3a): His eGFR is 50 mL/min/1.73 m², indicating that kidney protection is crucial.

• Coronary Artery Disease: He has undergone PCI, indicating a significant cardiovascular risk.

• Hyperlipidemia and Intermittent Claudication: These conditions further increase his cardiovascular risk.

EvidenceHunt: Mr. de Vries has uncontrolled hypertension (160/100 mmHg) with multiple comorbidities including Type 2 diabetes, hyperlipidemia, coronary artery disease, chronic kidney disease, and intermittent claudication.

Medical specialist:

Work Diagnosis (Severity, Cause, Possible Consequences) and evaluation of Existing Therapy (Effectiveness, Side Effects, Therapy Adherence): Essential hypertension (RR 160/100) with vascular disease, diabetes mellitus, kidney disease, and overweight. The latter limits his physical activity. Despite 10 mg ACE inhibition, his blood pressure remains too high. I wonder whether this is a one‐time measurement or if repeatedly elevated blood pressure has been observed (using a Dynamap or 24‐h measurement) to rule out white coat hypertension. I also question whether therapy adherence is an issue and if the patient is actually taking his lisinopril. ACE inhibition is useful, also for proteinuria in diabetes mellitus, and is the first choice for claudication.

Step 2: specify the therapeutic objective

ChatGPT:

The treatment goals for Mr. de Vries are:

• Blood Pressure Reduction: To reduce cardiovascular risk and slow the progression of kidney disease.

• Kidney Protection: To preserve kidney function through adequate blood pressure control and avoiding nephrotoxic agents.

• Minimization of Side Effects: Considering his comorbidities and existing medication regimen.

Given his diabetes and kidney disease, the target blood pressure should be below 130/80 mmHg.

EvidenceHunt: The primary goal is to reduce Mr. de Vries' blood pressure to < 130/80 mmHg to minimize cardiovascular and renal risks, considering his diabetes and CKD.

Medical specialist: Goal (curative, symptomatic, preventive, palliative):

• Target BP < 135/75

• Target LDL < 1.8

• Target BMI 25

• Target HbA1c < 53 in patients < 70 years

Step 3: verify the suitability of the P‐drug

ChatGPT: Since Mr. de Vries is already using lisinopril, an ACE inhibitor that has both antihypertensive and kidney‐protective effects, we must consider whether to increase the dose or add another medication.

Possible options include:

1. Increasing the Lisinopril Dose: Lisinopril can be increased up to a maximum of 40 mg per day. A higher dose may help achieve better blood pressure control while maintaining kidney protection. However, his potassium level is slightly elevated (4.8 mmol/L), which poses a risk of hyperkalemia if the dose is further increased.

2. Adding a Calcium Channel Blocker (e.g., Amlodipine): This is a logical step if monotherapy with an ACE inhibitor is insufficient. Amlodipine does not cause hyperkalemia, has a strong antihypertensive effect, and does not adversely affect kidney function.

3. Adding a Thiazide Diuretic (e.g., Chlorthalidone): This can further reduce blood pressure and has a complementary effect with ACE inhibitors. However, thiazide diuretics can increase blood glucose levels and are less effective in patients with reduced kidney function (eGFR < 30 mL/min).

Choice: Given Mr. de Vries' current medication and comorbidities, the best option is to add amlodipine. This medication does not negatively impact kidney function, does not cause hyperkalemia, and has a strong antihypertensive effect. Adding amlodipine can help lower blood pressure without the risks associated with increasing lisinopril or starting a diuretic.

EvidenceHunt:

Current Antihypertensive: Lisinopril (ACE inhibitor).

Additional Considerations:

• Effectiveness in BP Reduction: ACE inhibitors are effective but may need combination therapy for optimal control.

• Renal Protection: ACE inhibitors are beneficial for CKD patients.

• Cardiovascular Protection: ACE inhibitors reduce cardiovascular events.

• Potential Interactions: Monitor for hyperkalemia, especially with CKD.

• Side Effects: Cough, hyperkalemia, renal function deterioration.

Medical specialist: Non‐pharmacological treatment:

• Supervised walking training by a specialized exercise or physiotherapist for ≥ 6 months.

• Lifestyle coaching program for weight loss to achieve BMI 25.

• Foot care/pedicure if wounds are present, caution for diabetic foot.

• Formal recommendation: alcohol intake to 0 units/week, but I am often less strict—10 units/week is acceptable.

Pharmacotherapeutic treatment (based on evidence‐based medicine, guidelines and standards: selecting drug classes, evaluating mechanisms, choosing a registered drug, determining NNT, NNU, Time until benefit and pharmacokinetics):

• Hypertension:A second agent is needed. Within the ABCD framework, a calcium antagonist is preferred due to its vasodilatory effect, which may benefit claudication symptoms.

  • ○
    Beta‐blockers are not preferred due to ISA worsening claudication.
  • ○
    Diuretics are avoided in diabetes due to increased risk of hyperglycemia.
  • ○
    A second agent is more effective than increasing lisinopril to 20 mg once daily (permitted with eGFR 50 mL/min).
  • ○
    First step: Amlodipine 5 mg, caution for edema.
• Claudication Guideline:

  • ○
    ACE inhibitor and statin: replace simvastatin with rosuvastatin 20 mg (initial pharmacokinetic dose, stronger effect for LDL target, and fewer interactions than simvastatin).
  • ○
    Add dipyridamole alongside acetylsalicylic acid.
• Diabetes Guideline:

  • ○
    SGLT2 inhibitor, e.g., dapagliflozin, given vascular disease and diabetes, at the lowest dose considering HbA1c of 57.

Step 4: start treatment.

ChatGPT:

Proposed Treatment:

• Starting Dose: Amlodipine 5 mg, once daily.

• Maximum Dose: Amlodipine can be increased to 10 mg per day if needed, depending on the response and any side effects.

EvidenceHunt:

Recommended Additions:

• Angiotensin Receptor Blocker (ARB): Telmisartan or Losartan.

• Rationale: ARBs provide similar benefits to ACE inhibitors but with a lower risk of cough. They are effective in reducing proteinuria and protecting renal function [1] [2].

• Calcium Channel Blocker (CCB): Amlodipine.

• Rationale: Effective in reducing BP and safe in patients with intermittent claudication. CCBs do not adversely affect peripheral vascular disease [3] [1].

Proposed Regimen:

• Telmisartan: Start at 40 mg daily, titrate up to 80 mg as needed.

• Amlodipine: Start at 5 mg daily, titrate up to 10 mg as needed.

Medical specialist: Suitability of medications for the patient (contraindications, interactions, side effects, warning, special populations):

• Amlodipine 5 mg.

• Rosuvastatin 20 mg (Stop Simvastatin).

• I would not give Dipyridamole for peripheral vascular disease despite the guideline because the patient is already using a platelet aggregration inhibitor (Acetylsalicylic acid).

• Dapagliflozin 1 × 5 mg, start low in older age and comorbidity.

Step 5: give information, instructions and warnings

ChatGPT: Patient Education for Mr. de Vries:

• Explain that amlodipine will be added to better control his blood pressure and reduce his risk of heart and kidney complications.

• Inform him about possible side effects of amlodipine, such as ankle swelling, headache, and dizziness. Advise him to contact his healthcare provider if he experiences severe side effects.

• Discuss the importance of adherence to medication and continuing lifestyle recommendations, such as salt restriction and maintaining a healthy diet.

EvidenceHunt:

• Telmisartan: Take once daily. Monitor BP, renal function, and potassium levels regularly.

• Amlodipine: Take once daily. Monitor for peripheral edema and adjust dose if necessary.

• General Advice: Maintain a low‐sodium diet, moderate alcohol intake, and regular physical activity as tolerated.

Medical specialist: Non‐pharmacological:

• Supervised walking training by a specialized exercise or physiotherapist for ≥ 6 months.

• Program with a lifestyle coach to lose weight to BMI 25.

• Foot care/pedicure if wounds are present, beware of diabetic foot.

• Formally, alcohol to 0 units/week, but I am often not that strict. 10 units/week is okay.

Pharmacological:

Inform the patient (desired effect, possible side effects, intake instructions):

I would not start all three new medications at the same time because, in case of side effects, it would be unclear which medication caused them. Also, explain this to the patient.

I would explain to the patient:

• Amlodipine: Contact a doctor in case of edema.

• Rosuvastatin: Report serious muscle pain or weakness.

• Dapagliflozin: Be alert for signs of a urinary tract infection.

First amlodipine, after 2 weeks rosuvastatin, and then after 2 more weeks dapagliflozin.

Step 6: monitor and follow‐up

ChatGPT:

Monitoring:

• Check blood pressure after 2–4 weeks to assess if target levels are achieved.

• Monitor kidney function (serum creatinine and eGFR) and electrolytes (especially potassium) after 1 month, and then periodically.

• Watch for side effects such as peripheral edema and adjust the amlodipine dose if necessary.

Follow‐up Plan:

• If blood pressure remains uncontrolled at the maximum dose of amlodipine, consider adding another antihypertensive, such as a thiazide diuretic (with caution due to his diabetes).

EvidenceHunt:

• BP Monitoring: Check BP bi‐weekly until target is achieved, then monthly.

• Renal Function: Monitor serum creatinine and potassium levels every 4–6 weeks initially, then every 3 months.

• Diabetes Control: Regular HbA1c checks every 3 months.

• Lipid Profile: Monitor LDL, HDL, and triglycerides every 6 months.

Medical specialist:

What to monitor (effect, side effects, therapy adherence):

Ask the patient after 2 weeks how they are doing and whether they experience the above‐mentioned side effects. Additionally, check blood pressure, creatinine, electrolytes, glucose, and in case of complaints: muscle CK for muscle pain and urine sediment for dysuria. HbA1c after 3 months.

Follow‐up plan (control schedule):

Check‐ups at 2, 4, and 6 weeks. If everything is fine, the next check‐up will be in 3 months. If blood pressure is still not optimal (< 135/85), increase amlodipine to 10 mg.

Case 2

—Urinary tract infection

Case presentation

Patient Profile

Mrs. Maria Janssen is a 72‐year‐old female with a height of 1.65 m, a weight of 78 kg, and a BMI of 28.6 kg/m².

Medical History:

• Type 2 diabetes mellitus (15 years), treated with metformin and insulin.

• Chronic kidney disease (CKD stage 3b) with an eGFR of 38 mL/min/1.73 m².

• Hypertension (10 years), treated with lisinopril and amlodipine.

• History of recurrent urinary tract infections (last episode 6 months ago).

• Recent hospitalization for heart failure exacerbation.

Current Medications:

• Metformin 1000 mg, 2× per day

• Insulin glargine 20 units, 1× per day

• Lisinopril 20 mg, 1× per day

• Amlodipine 5 mg, 1× per day

• Furosemide 40 mg, 1× per day

• Atorvastatin 20 mg, 1× per day

Clinical Presentation:

• Symptoms: Dysuria, frequent urination, fever (38.5°C), lower abdominal and back pain.

• Physical Examination: Pain on palpation in the right lumbar region, suggesting pyelonephritis.

• Laboratory Results:

• Urine analysis: Positive for nitrites, leukocyte esterase, and blood; microscopy shows bacteria and leukocytes.

• Blood tests: WBC 14000/mm³ (reference: 4,000–11 000), CRP 80 mg/L (reference: < 5), serum creatinine 150 μmol/L (reference: 60–110), potassium 5.0 mmol/L (reference: 3.5–5.0), glucose 12.5 mmol/L (reference non‐fasting: < 7.8).

Working diagnosis: Pyelonephritis

Step 1: define the patient's problem:

ChatGPT: Diagnosis:

Mrs. Janssen presents with clinical and laboratory findings indicative of a complicated urinary tract infection (UTI) with possible pyelonephritis. The lumbar region pain, fever, and urinalysis positive for nitrite and leukocyte esterase support this diagnosis.

Comorbidities:

• Type 2 Diabetes Mellitus: Increases the risk of infections and can complicate the course of the UTI.

• Chronic Kidney Disease (CKD Stage 3b): Her eGFR is 38 mL/min/1.73 m², which influences the choice of antibiotics due to reduced renal clearance.

• Hypertension: Well‐controlled but may be exacerbated by the infection and medication.

EvidenceHunt: Mrs. Janssen presents with a complicated urinary tract infection (UTI), likely pyelonephritis, exacerbated by her comorbid conditions: type 2 diabetes, CKD stage 3b, and hypertension.

Medical specialist: Work diagnosis (severity, cause, possible consequences): Evaluation of existing therapy (effectiveness, side effects, therapy adherence):

This patient has a pyelonephritis, the choice of antibiotic depends on the assessment of the possible resistance of the causative pathogen. This depends on where the patient comes from (a nursing home with a high prevalence of ESBL+ E. coli in urine cultures, or a history of recurrent treatments for UTIs, leading to increased resistance).

Step 2: specify the therapeutic objective

ChatGPT: Treatment Objectives:

• Eradication of the Infection: To prevent sepsis, renal failure, or further deterioration of kidney function.

• Protection of Kidney Function: Use antibiotics that are effective but not nephrotoxic, with dosage adjustments based on renal function.

• Prevention of Resistance: Avoid broad‐spectrum antibiotics unless strictly necessary, and opt for targeted therapy if possible.

EvidenceHunt:

The primary goals are to:

‐ Eradicate the bacterial infection.

‐ Prevent complications such as sepsis.

‐ Manage her comorbid conditions without causing adverse effects.

Medical specialist: Sterilization of the urinary tract.

Step 3: verify the suitability of the P‐drug

ChatGPT: Possible Antibiotics:

1. Ciprofloxacin (Fluoroquinolone):

   • ○
     Advantages: Good penetration into the kidneys, effective against common uropathogens, available in both oral and intravenous forms.
   • ○
     Disadvantages: Requires dose adjustment due to renal function, risk of tendinitis/tendinopathy, and possible interaction with metformin (risk of lactic acidosis).
2. Ceftriaxone (Cephalosporin):

   • ○
     Advantages: Broad‐spectrum cephalosporin, effective against many uropathogens including ESBL‐producing organisms, not reliant on renal function for elimination.
   • ○
     Disadvantages: Requires parenteral administration (intravenous or intramuscular), less effective against Pseudomonas aeruginosa, risk of C. difficile infection.
3. Piperacillin‐Tazobactam:

   • ○
     Advantages: Broad‐spectrum, effective against both gram‐positive and gram‐negative bacteria, including Pseudomonas.
   • ○
     Disadvantages: Requires intravenous administration, potential nephrotoxicity, dose adjustment needed in renal insufficiency.
4. Nitrofurantoin:

   • ○
     Advantages: Effective in uncomplicated UTIs, low resistance, minimal systemic absorption.
   • ○
     Disadvantages: Not suitable for eGFR < 60 mL/min, and ineffective for complicated UTIs or pyelonephritis.

Antibiotic Choice: Given Mrs. Janssen's renal function, suspected pathogens, and the clinical severity of the infection, Ceftriaxone is an appropriate choice. This antibiotic offers broad coverage, does not require dose adjustment in renal insufficiency, and has a relatively favorable safety profile in older patients with comorbidities.

EvidenceHunt:

Given Mrs. Janssen's complicated UTI and her comorbidities, the choice of antibiotic must consider:

‐ Efficacy against common uropathogens.

‐ Safety profile in patients with CKD and diabetes.

‐ Minimal interaction with her current medications.

Recommended Antibiotic.

Ciprofloxacin is a suitable choice due to its broad‐spectrum activity and efficacy in treating complicated UTIs, including pyelonephritis. It is also effective in patients with diabetes and CKD, provided renal function is monitored.

‐ Dosage: Adjusted for renal function, typically 250–500 mg every 12 h.

‐ Duration: 7–14 days, depending on clinical response.

Medical specialist:

Non‐pharmacological: Drink well, 2 L/day, and do not hold urine. Urinate after sex.

Pharmacological: According to the guidelines (SWABid):

First choice: Ceftriaxone IV 2000 mg once daily for 10 to 14 days.

If proven susceptibility is confirmed, cotrimoxazole 960 mg twice daily orally is preferred for oral treatment, unless contraindicated.

Second choice: Ciprofloxacin 500 mg orally twice daily for 10 to 14 days.

Use as an (empirical) oral option or as empirical therapy when ceftriaxone cannot be administered—see also remarks.

If proven susceptibility is confirmed, cotrimoxazole 960 mg twice daily orally is preferred for oral treatment unless contraindicated. In that case, the treatment duration is 10–14 days for all women and 14 days for men.

Step 4: start treatment

ChatGPT: Proposed Treatment:

• Ceftriaxone 1–2 g, once daily intravenously.

  • ○
    Route of Administration: Intravenous, given the severity of the infection and the need for effective blood and tissue concentrations.
  • ○
    Duration: 7–10 days, depending on clinical response and sensitivity results.

EvidenceHunt:

• Initiate Ciprofloxacin: 250 mg every 12 h, adjusted for renal function.

• Monitor: Renal function, blood glucose levels, and signs of clinical improvement.

Medical Specialist: Check Whether the Medication Is Suitable for the Patient (Contraindications, Interactions, Side Effects, Warnings, Special Populations) and Choose Form, Dose, and Dosing Frequency

From the clinical presentation, it is difficult to determine if the patient is severely ill/septic and therefore needs to be admitted. In that case, the choice—according to swabid—would be ceftriaxone 1 × 2 g. If there is a high suspicion of ESBL+, meropenem would be the choice. If the patient is less ill and, for example, presents at the general practitioner's office, ciprofloxacin would be a good choice.

Step 5: give information, instructions and warnings

ChatGPT: Patient Education for Mrs. Janssen:

• Explain that Ceftriaxone is being used to treat the infection quickly and effectively, and that intravenous administration is necessary due to the seriousness of the situation.

• Inform her about potential side effects such as diarrhea (risk of C. difficile infection), allergic reactions, and pain at the injection site.

• Advise her to monitor her blood sugar levels regularly, as infections and medications like cephalosporins can cause hyperglycemia.

EvidenceHunt:

• Inform Mrs. Janssen: About the importance of completing the full course of antibiotics.

• Warn: About potential side effects, such as gastrointestinal disturbances and tendonitis.

• Instruct: To maintain hydration and monitor blood glucose levels closely.

Medical specialist:

Drink well, 2 L/day, and do not hold urine. Urinate after sex. Prescribe Ciprofloxacin 2 × 500 mg for 1 week. Inform the patient about the desired effect (treating infection), possible side effects (nausea, diarrhea, tendonitis, or tendon rupture), and instructions to take the medication as prescribed.

Step 6: monitor and follow‐up

ChatGPT: Monitoring:

• Clinical Parameters: Daily monitoring of temperature, pain, and general condition to evaluate the effectiveness of the treatment.

• Laboratory Tests: Repeat blood tests (WBC, CRP, serum creatinine) after 48 h to assess the effect of the therapy.

• Urine Culture: Order and follow up on sensitivity results to confirm the use of Ceftriaxone or switch to a more targeted antibiotic if needed.

• Renal Function: Regular monitoring of eGFR and serum creatinine to prevent further impairment of kidney function.

Complications:

• Monitor for signs of C. difficile infection, especially with prolonged use of cephalosporins.

• Expect possible improvement in blood pressure as the infection is controlled, which may require adjustments to antihypertensive medications.

EvidenceHunt:

‐ Follow‐Up: Reassess in 48–72 h to evaluate clinical response and adjust treatment if necessary.

‐ Monitor: Serum creatinine, potassium levels, and signs of infection resolution.

‐ Adjust Treatment: Based on culture results and sensitivity patterns.

Supporting Evidence

‐ Ciprofloxacin Efficacy: A study on elderly patients with complicated UTIs showed that age is not a direct factor in the effectiveness of antimicrobial agents, emphasizing the importance of managing underlying conditions [1].

‐ Diabetes and UTI: Women with diabetes have a higher risk of recurrent UTIs, and effective antibiotic treatment is crucial to prevent complications [2].

‐ CKD Considerations: Patients with CKD and UTIs require careful monitoring of renal function and appropriate dose adjustments to avoid nephrotoxicity [3].

Medical specialist:

If becoming ill, fever not reduced after 48 h, difficulty urinating, chills, contact healthcare provider (beware of urosepsis due to ciprofloxacin‐resistant bacteria or anatomical issues such as bladder emptying disorders/urinary retention). Very rare side effects like tendon rupture or aneurysm rupture, I would not inform the patient about.

Reasons for not choosing alternative treatment

ChatGPT:

Ciprofloxacin was not selected in this case due to several important considerations specific to Mrs. Janssen's medical condition. The reasons are outlined below:

1. Renal Function and Dosage Adjustment

Mrs. Janssen has reduced kidney function with an eGFR of 38 mL/min/1.73 m² (CKD Stage 3b). Ciprofloxacin requires dosage adjustment at this level of renal function to prevent drug accumulation and potential toxicity. Although adjustment is possible, it complicates dosing and increases the risk of side effects, which is risky for a patient with multiple comorbidities.

2. Side Effects and Risks

Ciprofloxacin, a fluoroquinolone, is associated with several serious side effects, including:

• Tendinitis and Tendon Ruptures: Particularly in older patients, those with renal insufficiency, and in those concurrently using corticosteroids. This is a significant concern given Mrs. Janssen's age and kidney function.

• QT Interval Prolongation: This can lead to serious cardiac arrhythmias, posing a risk for an elderly patient who is also on other medications that may affect the QT interval.

• Central Nervous System Side Effects: Such as confusion, dizziness, and in rare cases, even psychosis. Since Mrs. Janssen has already exhibited some confusion, ciprofloxacin could potentially exacerbate this condition.

3. Interactions with Existing Medications

Ciprofloxacin can affect the action of other medications, for example:

• Metformin: Ciprofloxacin can increase the risk of lactic acidosis, especially in patients with impaired kidney function.

• Insulin: There is a risk of altered blood sugar levels, which could be more difficult to manage in a diabetic patient like Mrs. Janssen, who is already on insulin.

4. Resistance

There is increasing resistance of uropathogens to fluoroquinolones like ciprofloxacin, particularly in patients with recurrent urinary tract infections (UTIs). Although resistance profiles can vary locally, it is important to consider the risk of ineffectiveness, especially in older patients with a history of recurrent UTIs.

5. Severity of the Infection

Mrs. Janssen's clinical presentation suggests a complicated UTI with possible pyelonephritis, a severe form of infection that requires prompt and effective treatment. Ceftriaxone, a broad‐spectrum cephalosporin, is often a more suitable choice for initial treatment in severe cases due to its strong bactericidal activity, broad coverage, and safety in patients with reduced renal function without the need for dosage adjustment.

Conclusion

While ciprofloxacin can be an effective choice for uncomplicated UTIs, it is less suitable in this specific case due to the risks of side effects, drug interactions, and the required dosage adjustments in renal insufficiency. Ceftriaxone offers a safer, more effective option for treating Mrs. Janssen's complicated UTI, given her medical history and the severity of her infection.

EvidenceHunt:

Overview: Ceftriaxone, a third‐generation cephalosporin, is commonly used for treating complicated urinary tract infections (UTIs). However, its use may not be ideal in patients with multiple comorbidities such as type 2 diabetes, chronic kidney disease (CKD) stage 3b, and hypertension. Below are the key reasons based on the latest medical literature.

Increased Risk of Clostridioides difficile Infection (CDI)

Higher Risk with Ceftriaxone: Ceftriaxone has been associated with a higher risk of healthcare facility‐onset Clostridioides difficile infection (HOCDI) compared to first‐generation cephalosporins like cefazolin. A retrospective study found that HOCDI developed in 0.40% of admissions with ceftriaxone exposure compared to 0.15% with cefazolin (adjusted odds ratio, 2.44; 95% CI, 1.25–4.76; p < 0.001) [1]. Given the patient's multiple comorbidities, the risk of CDI is a significant concern.

Drug Interactions and Increased Bleeding Risk

Interaction with Warfarin: Ceftriaxone can potentiate the activity of warfarin, leading to an increased risk of bleeding. A retrospective chart review found that ceftriaxone significantly increased the international normalized ratio (INR) more than other antibiotics, with a peak INR value of 3.56 compared to 2.66 for first‐generation cephalosporins (p = 0.004) [2]. This is particularly relevant for patients with CKD and hypertension, who may already be on anticoagulant therapy.

Efficacy and Safety in Complicated UTI

Alternative Options: In a combined analysis of two randomized, double‐blind, multicenter trials, ertapenem was found to be equally effective as ceftriaxone for treating complicated UTIs in adults. The success rates were 89.5% for ertapenem and 91.1% for ceftriaxone, indicating equivalent outcomes [3]. Given the patient's complex medical history, an alternative like ertapenem may be considered to minimize potential adverse effects.

Specific Considerations for CKD and Diabetes

Management of Comorbidities: Effective management of comorbid conditions like diabetes and hypertension is crucial in patients with CKD to reduce healthcare resource utilization and costs. A study found that good disease management was associated with up to 40% lower likelihood of inpatient hospitalization and up to 30% lower emergency department visits [4]. Using antibiotics with fewer side effects and interactions can help in better managing these comorbidities.

Summary

• Higher Risk of CDI: Ceftriaxone increases the risk of HOCDI significantly compared to first‐generation cephalosporins like cefazolin [1].

• Increased Bleeding Risk: Ceftriaxone interacts with warfarin, leading to a higher risk of bleeding, which is particularly concerning for patients with CKD and hypertension [2].

• Alternative Treatments: Ertapenem has been shown to be equally effective as ceftriaxone for complicated UTIs, offering a safer alternative [3].

• Comorbidity Management: Effective management of diabetes and hypertension in CKD patients is crucial, and using antibiotics with fewer side effects can aid in this [4].

In conclusion, ceftriaxone was likely not chosen due to its higher risk of CDI, potential for increased bleeding when used with warfarin, and the availability of equally effective but safer alternatives like ertapenem.

Data Availability Statement

The data that supports the findings of this study are available in the supporting information in this aricle.