Neuropathvision: a Knowledge Graph-Enhanced LLM Agent for Patient-Centered Interpretation of Neuropathology Reports

Abstract

Neuropathology reports are written for clinicians, making them difficult for patients to understand. Recent advances in large language models (LLMs) suggest a path to plain-language explanations for lay readers. In this study, we present NeuropathVision, a knowledge graph-enhanced LLM agent that generates patient-centered interpretations of neuropathology reports. The agent combines semantic retrieval from authoritative sources with a glioma-focused domain knowledge graph capturing entities and relations. It applies prompt-level style constraints such as clarity, transparent uncertainty, and empathy. To evaluate NeuropathVision, a nine-item patient-centered readability assessment form was developed covering four dimensions: Accuracy, Clarity, Empathy & Respect, and Treatment & Prognosis. Twenty cases were rated independently by two senior neuropathologists. All interpretations met the prespecified pass criteria, and the pooled mean score was 96.78. Exact agreement on item-level ratings was 86.7 %. Clarity and empathy scores were high, reflecting LLM strengths in following style guidance. The knowledge graph contributed structured anchors and consistency checks that supported accuracy. These results indicate the feasibility and acceptability of NeuropathVision for producing patient-centered neuropathology interpretations. Future work will add statementlevel citations and evaluate patient-reported outcomes.

Ähnliche Arbeiten