EPD512 - A Comparison of AI Platform to Diagnose Common Stoma Conditions from Photographs – Ready for Use or Still in Training?

Purpose/Background: Approximately 150,000 ostomy surgeries are performed annually in the United States, with an overall postoperative complication rate of 60% for both temporary and permanent stomas. With the growing accessibility of artificial intelligence (AI) tools, patients can now seek medical information and share images of their stoma remotely. While AI platforms have been explored for text-based medical queries, their ability to interpret patient-submitted images, particularly for stoma-related complications, has not been evaluated. This study assesses how different AI models interpret publicly available images of stomas and stoma complications.

Methods/Interventions: This study evaluated the performance of three AI models, ChatGPT-5, Gemini 2.5 Flash, and Claude Sonnet 4 to diagnose six common stoma conditions and also evaluate normal stoma photos. The stoma conditions included: peristomal varices, parastomal hernia, peristomal necrosis/ischemia, peristomal stenosis/retraction, stomal prolapse, and peristomal dermatitis. Three independent reviewers uploaded publicly available images, using standardized prompts simulating patient interactions. Images were uploaded three times to assess reliability of responses. The initial query, Prompt 1A ("This is a picture of my stoma. What do you think is going on?") prompted the AI for a single diagnosis, differential, or declination. If declined, Prompt 1B ("I understand. Can you describe what medical problems the image appears to show?") was used. When a differential diagnosis was generated, Prompt 2 ("Which one of these problems do you think is most likely?") obtained the final diagnosis. All responses were reviewed by a board-certified colorectal surgeon, and descriptive statistics were used to assess accuracy and reliability.

Results/Outcomes: We analyzed 198 interactions. ChatGPT provided a specific diagnosis for all cases (100%), but not Claude (95%, n=63/66) or Gemini (86%, n=57/66). Overall, ChatGPT most frequently included the correct diagnosis in its differential (78%) compared with Gemini (53%) and Claude (31%, Figure 1). ChatGPT also achieved the highest accuracy for the correct specific diagnosis (71%) versus Gemini (61%) and Claude (13%). There was significant heterogeneity in diagnostic accuracy with respect to AI platforms and stoma conditions (Figure 2). Some conditions were not diagnosed correctly by any platform (parastomal varices) while other conditions varied wildly such as parastomal hernias (100% correct by ChatGPT vs. 0% by Claude). For Gemini, Prompt 1B was used eight times, producing diagnoses in 72% (n=13/18) and no response in 28% (n=5/18). Claude required Prompt 1B five times and provided diagnoses in all (100%) cases. ChatGPT demonstrated the highest reliability, providing the same diagnosis across runs in 91% (n=20/22) of cases, followed by Gemini with 77% (n=17/22) and Claude with 59% (n=13/22).

Conclusion/Discussion: ChatGPT demonstrated the highest diagnostic accuracy among all artificial intelligence models, underscoring the potential role of AI in stoma care. However, there was notable heterogeneity in diagnostic accuracy across different AI platforms and stoma conditions, which limits their current clinical applicability. While these tools show promise but not yet suitable for independent clinical use. Future efforts should focus on improving reliability, transparency, and safe clinical integration assist with common stoma conditions.

Figure 1. Comparison of Diagnostic Accuracy for Specific and Differential Diagnoses Across AI Models.
Figure 2: Accuracy of three AI model for Differential Diagnosis and Specific Diagnosis.