Patient2Paper at #SMM4H-HeaRD 2026: Retrieval-Augmented Few-Shot Generation for Clinical Note Synthesis

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing, Health & Medical Research · Depth: Advanced, quick

Summary

The Patient2Paper system achieved 3rd place with an average score of 0.51 in the MedSynth Dial2Note shared task at SMM4H-HEARD 2026, focusing on retrieval-augmented few-shot generation for clinical note synthesis. Across 28 configurations, the study identified retrieval design, specifically a hybrid BM25 and medical-domain dense fusion via RRF, and prompt presentation format, using few-shot examples as conversation turns, as the primary drivers for quality. Surprisingly, model scale, including Llama 3.2:3B, Llama 3.1:8B, and GPT-4o mini, showed limited impact on performance. The final submission utilized GPT-4o mini with k=3 few-shot examples, retrieved by RRF over BioLORD-2023 embeddings, with a full ablation study detailing performance gains and diminishing returns.

Key takeaway

For NLP engineers developing clinical note synthesis systems, prioritize robust retrieval mechanisms and effective prompt engineering over simply scaling up language models. Your efforts should focus on designing hybrid retrieval strategies, like BM25 combined with dense embeddings fused via RRF, and presenting few-shot examples as natural conversation turns. This approach can yield significant performance gains, as demonstrated by the 3rd place finish at SMM4H-HEARD 2026, even with smaller models like GPT-4o mini.

Key insights

Hybrid retrieval and prompt formatting are key for few-shot clinical note synthesis, outweighing model scale.

Principles

• Retrieval design significantly impacts generation quality.
• Prompt format as conversation turns is crucial.
• Model scale has limited impact in this task.

Method

The system uses retrieval-augmented few-shot generation, employing a hybrid BM25 and medical-domain dense retrieval fused via Reciprocal Rank Fusion (RRF) to select examples for GPT-4o mini.

In practice

• Implement hybrid BM25 + dense retrieval.
• Format few-shot examples as conversation turns.
• Consider RRF for retrieval fusion.

Topics

• Clinical Note Synthesis
• Retrieval-Augmented Generation
• Few-Shot Learning
• Hybrid Retrieval
• GPT-4o mini
• BioLORD-2023
• SMM4H-HEARD 2026

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Severity-Aware Curriculum Learning with Multi-Model Response Selection for Medical Text Generation — Severity-aware curriculum learning with multi-model response selection significantly improves medical text generation quality and relevance.
• Benchmarking Retrieval Strategies for Biomedical Retrieval-Augmented Generation: A Controlled Empirical Study — Cross-Encoder Reranking significantly improves biomedical RAG performance by enhancing contextual precision.
• Supervision versus Demonstration-Based In-Context Learning for Multiword Expression Classification — Prompt sensitivity and demonstration quality critically impact LLM performance in metalinguistic classification tasks.
• Predicting relations between SOAP note sections: The value of incorporating a clinical information model — Fine-tuning a RoBERTa-base model on manually annotated clinical data improves NER for SOAP note sections.
• Few-Shot Biomedical Relation Extraction with Large Language Models: A Viable Alternative to Supervised Learning? — LLMs offer a viable few-shot alternative for biomedical relation extraction, outperforming prior few-shot methods, especially for rare relations.
• Measuring the sensitivity of LLM-based structured extraction to prompt, model, and schema choices in clinical discharge summaries — LLM-based clinical data extraction sensitivity varies significantly with model and schema, especially regarding "absence-versus-silence" distinctions.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.