Novel Breast Reconstruction Generative Pretrained Transformer Promotes Decisional Confidence and Clinical Efficiency

Summary:

Patients undergoing breast reconstruction face many decisions. A generative pretrained transformer (GPT) tool was custom built to assist patients at their initial reconstructive consultation. Twenty patients undergoing immediate breast reconstruction were enrolled, and 10 were randomized to interact with this GPT for 10 minutes at most before seeing the surgeon who was blinded. Patients then completed the 4-item abbreviated Decisional Conflict Scale SURE survey, the 12-item preoperative Consumer Assessment of Healthcare Providers and Systems survey, and a 3-item GPT-specific survey. Patients spent an average of 6.79 minutes asking an average of 6.5 questions, which were most commonly about recovery (34%), autologous reconstruction (23%), and complications (17%). With a 100% response rate (20 of 20), patients in the GPT cohort experienced less decisional conflict (Decisional Conflict Scale SURE—GPT: 3.9 versus control: 3.1, P = 0.03), had a shorter consultation time (GPT: 21.7 min versus control: 28.4 min, P = 0.02), and were more satisfied with surgical care across multiple domains, including shared decision-making (P = 0.04) and preoperative communication (P = 0.04). All patients (100%, 10 of 10) found this GPT to be very helpful and thought it should be made available for other patients. To our knowledge, custom GPTs built and trained by surgeons have not yet been studied in this setting, making this an innovation in our field. Other reconstructive breast surgeons may consider implementing this publicly available GPT into their practice to increase efficiency and satisfaction in patient care.

Takeaways

Question: Can a custom-built generative pretrained transformer (GPT) help counsel patients going through breast reconstruction?

Findings: This randomized control trial of 20 patients showed that a brief exposure to our breast reconstruction GPT significantly decreased decisional conflict (Decisional Conflict Scale SURE—GPT: 3.9 versus control: 3.1; P = 0.03), decreased consultation time (GPT: 21.7 min versus control: 28.4 min; P = 0.02), and increased satisfaction with shared decision-making (P = 0.04) and preoperative communication (P = 0.04).

Meaning: This publicly available GPT for breast reconstruction can be safely implemented in practice to increase patient satisfaction while decreasing consultation times.

INTRODUCTION

Generative pretrained transformers (GPTs) are large language models (LLMs) that use natural language processing to produce human-like responses to user queries. ChatGPT (OpenAI, San Francisco, CA) has many applications in plastic surgery, including improving readability of postoperative instructions, streamlining patient communication, and crafting operative notes.^1,2

Patients undergoing breast reconstruction face many decisions, including reconstructive method, implant type, and timing of reconstruction.³ Although reconstructive options might be limited by patient or pathology characteristics, patients often experience decisional conflict.⁴ Patient-facing GPTs may help those undergoing breast reconstruction.

The implementation of LLMs such as ChatGPT into patient care requires caution, as they can hallucinate, fabricate, and falsify, potentially compromising patient safety.^5,6 As such, we created a breast reconstruction GPT to assist patients navigating breast reconstruction, available through OpenAI’s ChatGPT interface (https://chatgpt.com/g/g-kvXVNuIp2-breast-reconstruction-gpt) (Fig. 1). The tool was trained using iterative feedback cycles with surgeons to ensure that answers produced accurately reflected institutional standards and included common complications, such as infection, poor wound healing, implant loss, flap failure, and pain. Twenty patients at a single institution were recruited (institutional review board no. STU00222182) and randomized into the GPT (n = 10) and control (n = 10) cohorts. Women aged 18–80 years seeking initial consultation for immediate breast reconstruction were included, whereas those younger than 18 years, older than 80 years, or undergoing delayed breast reconstruction, and men were excluded. The reconstructive surgeon was blinded to patient allotment. Patients in the GPT cohort were given a maximum of 10 minutes before the initial consultation to query the GPT, with no minimum or maximum number of questions required. To ensure accessibility across varying levels of health literacy, the GPT was designed to deliver responses at approximately a sixth- to eighth-grade reading level.⁷ No additional education intervention was provided by the study team before GPT exposure. Both groups received a standard in-person consultation with the surgeon. Patients then completed the 4-item, abbreviated Decisional Conflict Scale SURE survey, the 12-item preoperative Consumer Assessment of Healthcare Providers and Systems (CAHPS) survey, and a 3-item GPT-specific survey (GPT cohort only) using REDCap (Vanderbilt University, Nashville, TN). (See survey, Supplemental Digital Content 1, which displays the preoperative survey, https://links.lww.com/PRSGO/E478.)

Fig. 1.

Breast reconstruction GPT opening dialogue.

With a 100% response rate (20 of 20), the GPT cohort experienced less decisional conflict, indicating increased confidence in reconstructive decisions (Decisional Conflict Scale SURE—GPT: 3.9 versus control: 3.1; P = 0.03) (Table 1). Patients spent an average of 6.79 minutes asking an average of 6.5 questions to the breast reconstruction GPT before consultation with the surgeon. Common question topics included recovery (34%, 22 of 65 questions), autologous reconstruction (23%, 15 of 65 questions), and complications (17%, 11 of 65 questions). Supplemental Digital Content 2 contains the transcripts from all patient interactions with our GPT. (See Supplemental Digital Content 2, which displays all transcripts from patient interactions with the custom breast reconstruction GPT, https://links.lww.com/PRSGO/E479.) Consultation time was significantly shorter in the GPT cohort (GPT: 21.7 min versus control: 28.4 min; P = 0.02). The GPT cohort was also more likely to report the highest satisfaction with surgical care across multiple domains, including shared decision-making (CAHPS top-box answer proportion—GPT: 100% [10 of 10] versus control: 80% [8 of 10]; P = 0.04) and preoperative communication (CAHPS top-box answer proportion—GPT: 100% [10 of 10] versus control: 80% [8 of 10]; P = 0.04). All (10 of 10) patients strongly agreed that the breast reconstruction GPT was helpful, and all (10 of 10) patients strongly agreed that the tool should be made available for other patients.

Table 1.

Comparison of GPT and Control Cohorts

	GPT (n = 10)	Control (n = 10)	P
Race, n (%)
White	7 (70)	6 (60)	0.91
African American	2 (20)	2 (20)	0.99
Asian	1 (10)	1 (10)	0.99
Hispanic/ Latino	0 (0)	1 (10)	0.91
Age (mean ± SD), y	52.1 ± 6.9	47.3 ± 5.8	0.68
Reconstruction timing, n (%)
Immediate	10 (100)	10 (100)	0.99
Delayed	0 (0)	0 (0)	0.99
Decisional conflict score	3.9	3.1	0.03
Consultation time with surgeon, min	21.7	28.4	0.02
CAHPS shared decision-making top-box answer, n (%)	10 (100)	8 (80)	0.04
CAHPS preoperative communication top-box answer, n (%)	10 (100)	8 (80)	0.04
CAHPS preparation for surgery top-box answer, n (%)	9 (90)	9 (90)	0.99

^*Compared to the control group, the GPT group was significantly more confident in decisions, had a signficantly shorter consultation time, and was significantly more likely to select the most satisfied (top-box) answer on the Consumer Assessment of Healthcare Providers Systems (CAHPS) survey regarding shared decision making and preoperative communication.

DISCUSSION

This custom-built GPT created to assist patients navigating breast reconstruction decreased decisional conflict, increased satisfaction, and decreased consultation time. To our knowledge, this is the first study to assess a surgeon-created, publicly available GPT in the preoperative setting.

First, although GPTs such as the generic ChatGPT or Google’s Gemini are useful, providers should consider a tailored approach to LLM implementation using customized GPTs, especially when patient facing to reduce the risk of misinformation or hallucinations.^5,6 The platform of ChatGPT allows users to create their own GPTs with simple text instructions, allowing even those without extensive coding skills or background knowledge of LLMs. This platform is advantageous because it allows creators to upload published articles to the knowledge source of GPT, which it uses to answer queries about complication rates and patient outcomes.

Second, we found that the downtime patients’ experience before seeing a surgeon can be repurposed with this GPT. This likely helped patients feel their time was used productively, thus increasing satisfaction. Although this tool did not replace any aspect of the patient–surgeon interaction, its previsit use gave patients a better background understanding of breast reconstruction options, recovery timelines, and common complications. A more refined understanding of the reconstructive process enabled patients to ask more specific questions earlier in the surgeon’s visit. Many educational tools have been evaluated in the setting of breast reconstruction.^8,9 ChatGPT 3.5 can answer breast reconstruction–related questions well, but can sometimes hallucinate and lack depth.¹⁰ Our GPT expanded on the generic ChatGPT’s answers by adding institution-specific details without relaying any false information. Other electronic educational tools that decrease decisional conflict are more time consuming, requiring an average of 25 minutes.¹¹ Although our GPT did not replace the provider, it enabled patients to enter the consultation better informed, potentially reducing anxiety and allowing more focused discussions. Finally, providers should proceed with caution when expanding the use of this technology in settings without validation. Although our findings lend support to GPT use in a controlled, preconsultation environment, unsupervised access could lead patients to rely on GPTs for questions better answered by providers. Additionally, we did not examine its use for delayed reconstruction or long-term follow-up, which differs from immediate reconstruction.

This study has limitations. Our small sample size (n = 20) limited generalizability, although randomization and surgeon blinding strengthened our findings. The GPT was not formally validated through benchmarking but underwent extensive iterative review by the authors to ensure accuracy of responses. Patient literacy was not formally accessed, and the observed benefits in decisional confidence and satisfaction may reflect an indirect effect of feeling more prepared before consultation, rather than the GPT interaction alone. Nonetheless, these findings demonstrate feasibility, safety, and early efficacy. Our novel GPT is available to the public on OpenAI’s ChatGPT interface (https://chatgpt.com/g/g-kvXVNuIp2-breast-reconstruction-gpt) and easily accessible for future study and replication. Broader evaluation and comparison with generic models are warranted to further assess safety and use in diverse patient populations.

DISCLOSURE

The authors have no financial interest to declare in relation to the content of this article.