Publication
Retrieval-Augmented Generation (RAG) AI for End-of-Life Support in Pediatric HCT: Bridging Information Gaps for Patients and Caregivers on Palliative Care
Siddiqui K, Al-Musa A, Alsaedi H, Ghemlas I, AlAnazi A, Al-Ahmari A, Ramiz S, Sadler K, Virk T, Ayas M, Khan S · Transplantation and Cellular Therapy · 2026
Pediatric patients who have undergone hematopoietic cell transplantation (HCT) often face complex, life-limiting complications or disease relapse, leading to the difficult decision to transition to end-of-life (EOL) and palliative care (PC). In this context, providing immediate, trustworthy, and accessible information to patients and their caregivers is paramount for shared decision-making and optimal healthcare delivery. Retrieval-Augmented Generation (RAG) is an advanced Artificial Intelligence (AI) technique that combines the fluency of Large Language Models (LLMs) with the factuality of a curated knowledge base (KB). This approach is ideally suited to handle the dynamic, specialized, and highly sensitive information needs unique to the post-transplant, refractory disease population on PC. Objective To develop and validate a reliable, RAG-enabled AI agent capable of providing 24/7 on-demand information support to pediatric HCT patients and their parents facing end-of-life decisions and registered with palliative care services. Methods An expert-vetted KB was meticulously curated from approved hospital documentation, patient education materials specific to HCT, and FAQs about EOL and palliative care management (e.g., symptom control, DNAR status, medication regimens). This KB was used to train an LLM hosted on a prototype server, creating a user-facing AI Agent. The agent’s performance was rigorously assessed through an iterative testing and refinement process. Human evaluators, including clinicians and PC specialists, used a comprehensive suite of test questions and a matrix of six benchmark categories to assess the agent's responses. Results The AI Agent was powered by a fine-tuned GPT-4.1 nano LLM, featuring multi-language capabilities including Arabic. After three iterations of refinement, the agent demonstrated exceptional performance across all measured domains: Accuracy (Factual Accuracy and Relevance) reached 100%; Comprehensiveness (Completeness, Depth) 98%; Coherence (Logical Flow, Readability) 98%; User Experience (Conversationality, Personalization, and User Friendliness) 98%; Performance (Response Time, Scalability, Robustness) 98%; and RAG Specific metrics (Retrieval Effectiveness, Generation Quality) achieved 100%. Conclusion The validated RAG-enabled AI agent exhibits excellent performance and factuality based on human clinical evaluation. This technology represents a viable and scalable option for deploying 24/7 on-demand information support specifically for the highly vulnerable population of pediatric HCT recipients who have transitioned to palliative care. Integrating this agent can significantly enhance caregiver support, improve patient-reported outcome measures, and standardize the delivery of sensitive, specialized end-of-life information within the transplant center.
Synced from the TRENDS in Pediatric Palliative Care Zotero library, curated by The Siden Research Team. ICPCN does not host or verify the full text.