The Utility of ChatGPT for Assisting Patients with Study Preparation and Report Interpretation of Myocardial Viability Scintigraphy: Exploring the Future of AI-driven Patient Comprehension in Nuclear Medicine

Abstract

Purpose:

India’s high-volume healthcare system, including nuclear medicine departments, restricts the quality and frequency of patient–clinician communication. Poor understanding of preparation requirements increases cancellations, image artefacts, and repeat studies. Artificial intelligence (AI) chatbots like chat generative pre-trained transformer (ChatGPT) can be a promising tool to mitigate these challenges. We evaluated the efficiency of ChatGPT to address patients’ queries about study instructions and report findings while undergoing nuclear myocardial viability study.

Subjects and Methods:

Six myocardial-viability mock reports were created. OpenAI ChatGPT-4o responses were evaluated for the set of 14-questions regarding patient preparation and 2-questions regarding the reports. All questions and reports were entered as single prompts in separate chats. Each prompt was repeated twice using regenerate-response function. Furthermore, references used to generate responses were analyzed. The responses were then rated based on 5 key parameters: appropriateness, helpfulness, empathy, consistency, and validity of references.

Results:

Most responses were appropriate and helpful for both preparation (1.5 ± 0.76; 1.64 ± 0.63) and report prompts (1.67 ± 0.49; 2.0). However, empathy and consistency had lower scores in preparations (1.43 ± 0.76; 1.14 ± 0.66) than in report prompts (1.58 ± 0.51; 1.67 ± 0.49). Reference validity remained an issue, as only one response had a valid reference. A hallucinatory response was noted twice. The study demonstrated that none of the prompt responses could have caused harm to the patient under real-life conditions.

Conclusions:

ChatGPT helps in query resolution in myocardial viability studies. It enhances patient engagement, quality of patient preparation, and comprehension of nuclear medicine reports. However, inconsistent and less empathetic responses mandate supervised use and further refinement before incorporating it into routine practices.

Introduction

India’s high-volume healthcare system places extraordinary demands on physicians’ availability and time, often restricting the quality and frequency of patient–clinician communication.[1] It is more so evident in nuclear medicine, where a single department may offer a variety of studies. Some of them, such as myocardial studies, demand unique and stringent preparatory measures, specific contraindications, and acquisition protocols. Inadequate understanding of these details often leads to poor compliance with fasting or medication-withdrawal requirements, increased scan cancellations, image artefacts, and ultimately repeat studies that inflate radiation exposure and costs.[2]

Conversational AI chatbots have emerged as readily accessible tools for bridging such information gaps and enhancing patient compliance. Large language models (LLMs) such as ChatGPT deliver context-sensitive, natural-language responses also in regional languages. Moreover, recently, AI chatbots are being explored for triage, appointment scheduling, and patient education across several medical specialties.[3] For nuclear medicine, an on-demand chatbot could enhance prescan preparation by translating department instruction sheets for patient preparation into vernacular, interactive questions and answers (QnA) that patients can revisit at home, thereby improving adherence and reducing “noninterpretable” studies.[4] It can also foster inclusion and autonomy by empowering patients and caregivers – especially those with limited health literacy – to clarify doubts without fear of judgement in multiple Indian languages. Finally, patients and their family members can also gain insights through improved postscan comprehension by summarizing and explaining common radiotracer concepts and key findings in the final report. Although definitive interpretation and management decisions remain the clinician’s prerogative, this could greatly help to reduce the immediate anxiety of patients and caregivers until the reporting physician is available to brief them about the reports.

Even though customized chatbots can be built for individual centres and settings, ChatGPT is an easily accessible, continuously updated, multilinguistic tool with no need of local infrastructure. This makes it an attractive solution for resource limited settings – provided its responses are accurate, consistent and free of unsafe advice.[5,6]

The present study, therefore, evaluates ChatGPT as a patient-facing information tool within the nuclear medicine workflow. Focusing on myocardial viability imaging as a representative nuclear medicine scan, we propose to assess its response to common preparation-related patient queries, the clarity of its report explanations, and the extent to which it complements, rather than substitutes, physician counseling based on the reports generated. We aim to define the utility and limitations of deploying ChatGPT in India’s busy nuclear medicine departments, with a view toward scalable patient-centric care.

Subjects and Methods

We performed a single-center, descriptive study in the Department of Nuclear Medicine at our institute, evaluating the ability of ChatGPT-4o (OpenAI) to answer frequently asked questions (FAQs) about nuclear myocardial viability imaging. No patient-identifying data were used. Most common questions regarding nuclear myocardial viability preparation and report interpretation were collected from the residents and front desk staff based on their regular interactions with patients and their caretakers.

ChatGPT-4o was accessed between September 24, 2024 and September 30, 2024. Interactions were performed through the public web interface of OpenAI ChatGPT-4. To avoid contextual carryover, every question was entered in a fresh conversation window. For each response, the regenerate response button was clicked twice, producing three independent outputs per prompt. Complete transcripts were exported in plain-text format immediately after capture and time-stamped to document the exact access dates.

A sixteen-prompt catalogue of frequently asked questions (FAQs) was developed by (i) reviewing departmental instruction leaflets and (ii) prospectively logging telephone and desk enquiries from patients and caregivers. Six mock nuclear myocardial viability reports were made. Fourteen prompts addressed prescan preparation, and two-prompt-explored post-report interprétations [Table 1]. Wording was kept deliberately nontechnical to approximate genuine patient dialog. A separate prompt requesting valid references for the responses generated was also created.

Table 1.

Preparation and report-related prompts

Preparation prompts

1. What is a nuclear myocardial viability test?

2. Why is a nuclear myocardial viability test needed?

3. How much time does a nuclear myocardial viability test take?

4. Is nuclear myocardial viability test harmful for me?

5. How should I prepare myself for a nuclear myocardial viability test?

6. Are there any medications that I need to stop before the nuclear myocardial viability test?

7. Should I follow any dietary precautions before undergoing nuclear myocardial viability test?

8. Can I continue antihypertensives before undergoing a nuclear myocardial viability test?

9. I had a heart attack 2 months ago. Can I still take the nuclear myocardial viability test?

10. Being a diabetic can I take nuclear myocardial viability test?

11. Can I continue antidiabetics/insulin before undergoing a nuclear myocardial viability test?

12. I am not willing to undergo a nuclear myocardial viability test. Does it lead to any problem in the future?

13. Should I take any precautionary measures after the nuclear myocardial viability study?

14. Are there any contraindications to nuclear myocardial viability study?

Report prompts

1. Please explain my report in brief

2. Based on this report what my doctor/cardiologist will plan for me?

Validity of references

1. What are the references that ChatGPT followed while generating these answers?

An experienced nuclear medicine physician and one postgraduate trainee appraised every ChatGPT response using a five-domain rubric by consensus [Table 2]. Domains comprised appropriateness, helpfulness, empathy, consistency (across the three regenerations), and validity of references, each scored on a three-point ordinal scale in which 0 = not appropriate/helpful/empathetic/consistent/no valid reference, 1 = indeterminate, and 2 = appropriate/helpful/empathetic/consistent/valid reference present.

Table 2.

Scoring rubric for domains

Score (0–2) with criterion	Details
Appropriateness/accuracy
Appropriate	Accurate information meeting department standards
Partly appropriate	Incorrect/inconsistent information
Inappropriate	Potentially harmful information
Helpfulness
Helpful	Comprehensive and answers patient’s queries completely
Partly helpful	Specific but lacks crucial information
Unhelpful	Not specific to the question
Empathetic
Yes	Human-like empathy while generating answers
Uncommentable	Cannot be assessed
No	Neutral
Consistency/precision
Consistent	Only changes in layout, word arrangement, and narration style
Partly consistent	Differences in core content
Inconsistent	Incompatible responses in multiple regenerations
Validity of references
Valid	Appropriate, accessible, and identifiable source
Incorrect citation/outdated citation	Generic references/outdated citations/incorrect bibliographic data but relevant to the prompt
Invalid	Hallucinations (invalid references that cannot be found online)

Data were first coded into Microsoft Excel and then exported to SPSS Statistics v27 (Statistical Package for Social Sciences; Manufacturer: IBM) after data cleaning. All computations were performed in SPSS Statistics v27. Descriptive statistics were generated. Mean ± standard deviation for each domain was calculated separately for preparation prompts (n = 14) and report-interpretation prompts (n = 2). No inferential or inter-rater reliability tests were applied as the ratings were consensual. For visual comparison, the domain means were plotted on a five-axis radar chart (preparation vs. report) using Microsoft Excel, with the radial scale fixed at 0–2.

Raw transcripts, rating spreadsheets, and statistical outputs were archived. An anonymized compendium of reports was used to generate responses.

Results

We analyzed sixteen patient-focused prompts – 14 concerning prescan preparation and 2 concerning report interpretation. Each preparation prompt was submitted to ChatGPT-4.0 and regenerated twice (three responses per prompt). In addition, a prompt asking for valid references was asked in the end, yielding 56 preparation-related responses. Each of the six template reports were submitted in a separate window. Subsequently, two related prompts were submitted for each report and regenerated twice. Similarly, a prompt requesting reference validity was submitted in the end. This yielded a total of 42 responses. Every prompt was rated in five domains (appropriateness, helpfulness, consistency, empathy, and validity of reference) on a 0-to-2 scale.

Table 1 summarizes ChatGPT-4o’s quantitative performance across the five evaluation domains. For the 14 prescan preparation prompts (56 rated answers), the chatbot achieved mean scores between 1.14 and 1.64. ‘Appropriateness’ averaged 1.50 ± 0.76, whereas “Helpfulness” was slightly higher at 1.64 ± 0.63. “Empathic language” was less consistent (mean 1.43 ± 0.76), suggesting that preparation replies often lacked clear reassurance or patient-centred phrasing. “Consistency” across the three regenerations was 1.14 ± 0.67. Reference validity remained the principal weakness; none of the 42 preparation answers contained a fully verifiable citation.

For the report-interpretation prompts (42 rated answers), clinical content was similarly reliable. “Appropriateness” was 1.67 ± 0.49, and “Helpfulness” reached a perfect mean of 2.00 ± 0.00, reflecting the model’s ability to translate perfusion–metabolism findings into plain language for common understanding. In contrast, both “empathy” (1.58 ± 0.51) and, particularly, “consistency” (1.67 ± 0.49, only 33% optimal) increased, indicating lesser wording variability and infrequent omissions when the same query was regenerated. Reference validity remained poor with one fully verifiable citation, echoing the preparation findings.

“Hallucinatory” (generated responses are different in context to the prompt submitted) responses were noted twice in the responses.

Discussion

The nuclear medicine departments in India, particularly in reputed institutions, have high patient volumes, multilingual and diverse populations, and brief consultation windows, which constrain thorough physician–patient interactions. AI chatbots provide on-demand, language-flexible explanations without additional staffing. OpenAI ChatGPT-4o was selected because it is multilingual, cloud-hosted, and continuously updated, eliminating the need for local infrastructure or specialized coding expertise. Prior work has also shown accuracy in answering positron emission tomography/computed tomography (PET/CT) related queries, supporting its validity for the present task.[7]

This study demonstrates that ChatGPT-4o, a publicly accessible large language model(LLM) chatbot, can deliver clinically acceptable answers to common patient enquiries in nuclear myocardial viability imaging. Mean scores for the preparation prompts – appropriateness (1.50 ± 0.76), helpfulness (1.64 ± 0.63), empathy (1.43 ± 0.76), and consistency (1.14 ± 0.67) – indicate that most responses were both generally correct and practically useful, although reassurance and internal uniformity were less robust than technical accuracy. For the report-interpretation prompts, ChatGPT retained high “Appropriateness” (1.67 ± 0.49) and “Helpfulness” (2.00 ± 0.00). “Empathy” and “Consistency” were better (1.58 ± 0.51 and 1.67 ± 0.49, respectively) in report-related responses [Table 3]. Reference validity was the weakest domain overall [Figure 1], with only a single verifiable citation identified in report-related responses and two hallucinated references recorded. Nevertheless, none of the generated responses contained advice that might jeopardise patient safety if followed.

Table 3.

Quantitative metrics of domains tested

Domain	Preparation prompts	Report prompts
Appropriateness	1.50±0.76	1.67±0.49
Helpfulness	1.64±0.63	2.00±0.00
Empathy	1.43±0.76	1.58±0.51
Consistency	1.14±0.67	1.67±0.49
Reference validity	0.79±0.41	1.00±0.00

Figure 1.

Radar plot comparing the mean performance of ChatGPT-4o on prescan preparation prompts (blue line) versus report-interpretation prompts (orange line) across five evaluation domains: Appropriateness, helpfulness, empathy, consistency and reference validity. Scores are plotted on a 0–2 scale. Larger values indicate stronger performance in the corresponding domain, demonstrating overall better performance in report prompts

The favorable appropriateness and helpfulness metrics suggest ChatGPT could offload routine telephone or desk enquiries, thereby preserving a nuclear medicine technologist’s or physician’s time and reducing scan cancellations due to inappropriate or wrong patient preparation. Its helpfulness score for report explanations highlights its ability to convert perfusion–metabolism jargon into layman’s language, potentially improving patient comprehension and satisfaction. Empathy and Consistency, however, lag behind content accuracy, underscoring the need for tone calibration and stability checks before unsupervised deployment. In addition, lack of appropriate citations underscores a critical gap in reference integrity, emphasizing the necessity of integrating verifiable established department protocol sheets and evidence-based sources to support the chatbot’s responses.

Despite lack of valid references and two contextually irrelevant statements (hallucinations), no response had guidance which may be harmful to the patient, supporting the chatbot’s suitability for non-diagnostic information tasks. The study by Aiumtrakul et al. demonstrated that AI chatbots such as ChatGPT, Bing Chat, and Bard have inconsistent citation accuracy in the medical field, highlighting the need for rigorous vetting before their use.[8] Future research should benchmark ChatGPT against emerging open-source LLMs and domain-tuned healthcare agents.

Customization of existing LLMs through retrieval-augmented generation or instruction tuning on society guidelines and multilingual patient leaflets could enhance consistency and reference validity while retaining ChatGPT’s conversational fluency. In parallel, aspirational departments may fine-tune lighter customised open models on their own QnA archives or societal recommendations mixed with their own markdown documents to create locally hosted chatbots that cite institution-specific protocols.

Strengths of this study are the pragmatic design that evaluates both patient-centric preparation and report-related queries. Use of consensus ratings by a blend of experienced evaluators and representatives from more tech-savvy emerging generations, as well as a threefold regeneration capacity that reveals consistency, further enhances the credibility. However, the reliability and utility of this output could have been improved further by incorporating vernacular language in addition to English and involving multiple centers. Some other limitations of this study include a modest prompt list and absence of inter-rater statistics.

Conclusions

ChatGPT-4o provides appropriate, helpful, and safe replies to most patient questions in nuclear myocardial viability imaging, but its inconsistent, less empathetic responses and improper reference citation mandate supervised use before incorporating it into routine practices. Fine-tuning on authoritative nuclear medicine content or the deployment of specialty-specific tailored models are rational next steps toward an ideal chatbot for patient-centerd care.

Conflicts of interest

There are no conflicts of interest.

Funding Statement

Nil.