Hematobot: Retrieval‑augmented large language model‑based assistants for therapy, toxicity and frailty management in multiple myeloma by the spanish society of hematology and hemotherapy (SEHH)

Abstract

Abstract Background Generative artificial intelligence (GenAI) powered by large language models (LLMs) and retrieval-augmented generation (RAG) is reshaping continuous medical education and point-of-care consultation. When free-text output is anchored in a rigorously curated specialty corpora, such assistants provide rapid, guideline-aligned answers, surface primary references on demand and preserve clinician autonomy. Aims To present SEHH-HematoBot, a unified platform launched in June 2025 that houses two complementary AI tutors for Spanish hematologists treating multiple myeloma (MM): (i) Therapy & Toxicity, focused on first- and later-line regimen selection and adverse-event management in MM; and (ii) Frailty & Supportive Care, dedicated to geriatric assessment, hygienic-dietary guidance and complication mitigation in MM. Methods A four-member scientific committee steered development of the Therapy & Toxicity assistant, while a five-member committee oversaw the Frailty & Supportive Care counterpart. Each team performed an exhaustive literature review (ESMO, NCCN, IMWG, pivotal trials, real-world studies), compiling closed corpora current to May 2025. For both assistants, retrieval-augmented pipelines on GPT-4.1 were constructed and governed by comprehensive rule sets that enforce guideline concordance, dosing safeguards and citation integrity. Committees employed iterative prompt engineering and expert review to meet prespecified quality standards. Committees iteratively refined generations through transparent prompt engineering until domain expectations were met. Each assistant was benchmarked with pre-specified, clinically representative question sets covering its remit. All responses received scores >8/10 from every committee member for clinical relevance, completeness and citation quality, meeting the release threshold. The two assistants were then merged into a single portal—Hematobot (https://hematobot.sehh.es, user id: hematobotGSK1, password: GSKchat+2025)—that includes real-time user-feedback widgets for ongoing refinement. Results The Therapy & Toxicity assistant was built on a corpus of 110 curated documents and distilled into a ~14 500-token system prompt that embeds ~110 individual rules. The Frailty & Supportive Care assistant drew from 123 documents, generating a 15 000–18 000-token prompt with >100 rules across 28 domains. Both models answered their benchmarking question sets with scores >8/10 from every committee member and were released to SEHH members in June 2025 within a single Hematobot portal that offers real-time feedback tools. Early qualitative feedback highlights ease of use and confidence in guideline alignment; quantitative usage data collection is ongoing. Conclusions SEHH has demonstrated that small, expert-led teams can rapidly create and validate retrieval-augmented LLM assistants covering distinct yet complementary aspects of MM care and integrate them in a single, society-wide educational platform. Continuous user feedback and automated literature surveillance will drive future enhancements while maintaining guideline fidelity.