Enhancing Postoperative Care in Cosmetic Surgery using Prompt Agents and Chain of Thoughts Strategies to improve Retrieval-Augmented Generation Models

Abstract

PURPOSE: Providing effective postoperative guidance is essential for improving patient outcomes and satisfaction in plastic surgery. Retrieval-Augmented Generation (RAG) models powered by Large Language Models (LLMs) hold significant promise in delivering personalized and accurate postoperative support. However, their success is highly dependent on the optimization of prompts, which shape the model's reasoning and response generation. This work highlights the transformative potential of incorporating Prompt Agents and Chain-of-Thought (COT) strategies to optimize reasoning and elevate AI-driven postoperative care. By leveraging these techniques, we aim to set new standards for accuracy, relevance, and decision-making performance in clinical applications. METHODS: The study evaluated two distinct RAG architectures integrated into Gemini 1.0 Pro using a curated set of 32 postoperative questions covering five common plastic surgery procedures: liposuction (n=7), breast augmentation (n=6), abdominoplasty (n=6), mastopexy (n=6), and blepharoplasty (n=7). The first architecture utilized a naive RAG pipeline as a baseline, while the second introduced an advanced framework with Prompt Agents incorporating COT strategies. This advanced pipeline enhanced reasoning capabilities, allowing the system to generate more precise, accurate, and contextually relevant responses tailored to postoperative queries.Performance was systematically assessed using key metrics, including medical accuracy, relevance, precision, recall, and F1 score, ensuring a robust evaluation of each model's effectiveness in addressing complex patient concerns. The Prompt Agent-COT framework was designed to improve clinical reasoning and ensure alignment with medical standards, particularly in addressing nuanced and detailed postoperative questions. RESULTS: The results demonstrated the superiority of the Prompt Agent-COT framework over the naive RAG pipeline across all performance metrics. The advanced architecture achieved exceptional results, including a precision of 0.81, recall of 1.0, and an F1 score of 0.90. The relevance of its responses was similarly impressive, with 88% rated as highly relevant, 12% as moderately relevant, and 0% as low relevance. The accuracy distribution further underscored its effectiveness, with 81% of responses classified as highly accurate, 19% as moderately accurate, and 0% as low accuracy.In contrast, the naive RAG model underperformed significantly, achieving a precision of 0.50, recall of 0.33, and an F1 score of 0.39. Its relevance distribution showed only 43% of responses rated as highly relevant, 44% as moderately relevant, and 13% as low relevance. Accuracy metrics reflected similar challenges, with 50% of responses falling in the low accuracy range, 25% in the medium range, and only 25% in the high range. These results emphasize the transformative impact of integrating Prompt Agents with COT strategies, showcasing the ability of this advanced framework to outperform baseline models in delivering high-quality responses. CONCLUSION: This study demonstrates the pivotal role of Prompt Agents and Chain-of-Thought strategies in enhancing the reasoning capabilities of RAG-based LLMs for postoperative care. The advanced framework not only improves the precision and accuracy of responses but also ensures greater contextual relevance, tailoring guidance to the unique needs of postoperative patients. By addressing the limitations of naive implementations, this approach provides a transformative solution for delivering superior patient care, paving the way for more reliable, precise, and intelligent AI-driven systems in plastic surgery.© 2026. Plastic Surgery Research Council | All rights reserved |*Source: https://ps-rc.org/meeting/Program/2026/AS65.cgi*

Ähnliche Arbeiten