Knowledge-augmented Agentic AI for psychiatric medication info

Knowledge-augmented Agentic AI, a provenance-aware knowledge-graph multi-agent framework for psychiatric medication information seeking, was submitted to arXiv on 24 Jun 2026 by Huizi Yu and co-authors. The system unifies 466,525 Reddit posts, 60,782 WebMD reviews and twenty years of U.S. FDA Adverse Event Reporting System records across nine antidepressants.

What did the researchers build?

The paper describes a provenance-aware, knowledge-graph-based multi-agent framework that integrates three distinct sources of safety knowledge: large-scale patient narratives from Reddit and WebMD, and authoritative regulatory data from twenty years of FDA adverse-event reports. A Neo4j knowledge graph grounds entities in ATC-N, ICD-10, and MedDRA vocabularies and preserves provenance so every claim remains traceable and regulatory facts remain distinct from patient experience.

The architecture connects an LLM-based entity-recognition pipeline, normalization to clinical vocabularies, and storage in Neo4j. The authors position the system to keep community-sourced signals and regulatory records separate while enabling cross-source queries and auditing of data provenance.

How good is the extraction and what do the data show?

The LLM entity-recognition pipeline was benchmarked against physician annotations and achieved top F1 scores of 0.969 for medications and 0.973 for conditions, showing high extraction accuracy on those entity types. The study reports a Jaccard similarity between the two community platforms of up to 0.905, indicating strong concordance between Reddit and WebMD but less overlap with FDA records.

The dataset scale is concrete: 466,525 Reddit posts and 60,782 WebMD reviews for nine antidepressants, joined with twenty years of FDA Adverse Event Reporting System records. The authors highlight timing differences as well: for sertraline, many adverse events appeared in community sources hundreds of days before the corresponding FDA date, suggesting patient forums can surface experience-near signals earlier than regulatory reports.

Why does this matter?

Source-aware integration matters because psychiatric medication information mixes authoritative but abstract regulatory reports with experience-near patient narratives, and conflating them can amplify fear, nocebo responses, and non-adherence. By preserving provenance and separating regulatory facts from patient experience, the framework aims to make medication safety information more auditable and less likely to mislead patients and clinicians.

This approach also treats patient-generated data as a partly independent safety signal: the high Jaccard overlap between community platforms coupled with lower concordance with FDA records implies community sources carry distinct information that could complement, but should not replace, regulatory evidence.

What to watch

The authors state usefulness and patient benefit must be tested prospectively; the next milestone to watch is a prospective study or deployment that measures whether the provenance-aware outputs reduce misinterpretation, nocebo effects, or improve adherence in real-world users. Also watch for published details about the nine antidepressants' specific signal timelines and any code or Neo4j schema the authors release.

References and concrete data points in the paper include: 466,525 Reddit posts, 60,782 WebMD reviews, twenty years of U.S. FDA Adverse Event Reporting System records, nine antidepressants, LLM NER F1 scores 0.969 (medications) and 0.973 (conditions), and a Jaccard similarity between community platforms up to 0.905. The arXiv submission identifier is arXiv:2606.26205 (submitted 24 Jun 2026).