MedGraph introduces graph memory for medical dialogue
Original title: MedGraph: Maintaining Patient State with Graph Memory for Long-Term Medical Dialogue
MedGraph uses graph memory to maintain patient state, outperforming baselines by 10 points on MedMemoryBench.
The paper examines errors in five memory systems for medical assistants. MedGraph extracts dated facts, attributes them to persons, and links records into chains. It ranks chains for retrieval and includes safety facts. On the benchmark, it achieves highest accuracy with GPT and Qwen backbones.
- MedGraph addresses outdated states, person health facts, and missing safety facts
- Construction extracts dated facts and builds chains for medication and diagnosis
- Retrieval reads chains at specific times and adds dialogue text
- Improves accuracy by 10.5 points over strongest baseline with GPT backbone
- Releases MedMemoryBench for evaluation
From the original article
1 min read at arxiv.orgLLM-based medical assistants that follow a patient across visits need a memory that keeps the patient's clinical state correct as it changes. We examine the errors of five memory systems on longitudinal medical dialogue and find three recurring causes: the retrieved evidence gives an outdated state, includes another person's health facts, or lacks a safety fact the answer depends on.
By Donghua Zhang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.