RAG-GenomicSumm: A Retrieval-Augmented Generation Framework for Plain-Language Summarization of Clinical Genomic Reports via an Event-Driven Microservices Pipeline

Abstract

Clinical genomic reports generated from next-generation sequencing contain dense technical terminology poorly understood by non-specialist clinicians, genetic counselors, and patients. Large language models offer an opportunity to automate plain-language summarization but are constrained by hallucination, knowledge staleness, and variant-level factual inaccuracy. This study developed and evaluated RAG-GenomicSumm, a production-grade framework combining Retrieval-Augmented Generation over a curated genomic knowledge index (ClinVar, gnomAD, OMIM, dbSNP, COSMIC) with an event-driven microservices architecture built on Apache Kafka and containerized via Kubernetes. The pipeline ingests VCF or structured JSON reports, enriches them with grounded variant annotations, generates layered summaries (technical, clinical, patient-facing), and streams outputs to downstream consumers with sub-200 ms latency. RAG-GenomicSumm achieved a ROUGE-L score (Recall-Oriented Understudy for Gisting Evaluation, longest common subsequence variant) of 0.724 and a BERTScore F1 (contextual-embedding-based semantic similarity score) of 0.891 on 250 de-identified oncology and hereditary cancer reports. Hallucination rate was 4.2% (95% CI, 2.3-7.5%), compared to 18.7% (95% CI, 14.4-24.0%) for the base GPT-4o baseline without retrieval (p < 0.001). The retrieval accuracy reached P@5 = 0.81 (95% CI, 75.7-85.4%) and nDCG@10 = 0.79 (95% CI, 73.5-83.6%). The Kafka pipeline processed 12,400 variant events per second under simulated peak load, with 99.96% throughput retention. These results demonstrate that grounding LLM inference in authoritative variant-annotation indexes within a scalable streaming architecture substantially improves clinical genomic report summarization quality, reduces hallucinations, and meets real-world throughput requirements.