PRISM: AI agents that surface unmet genomic care needs from social media

Kamalakkannan Ravi, PhD; Mengshu Nie, MA; Annapurna Poduri, MD, MPH; Cong Liu, PhD
Division of Genetics and Genomics, Boston Children’s Hospital; Harvard Medical School.
CliGenAI Lab.

People discuss genetic results publicly, often before a clinic sees them. PRISM, an LLM-based pipeline, reads public posts, checks ClinVar, the public NIH database, and drafts a message for review. The AI never decides whether a variant causes disease.

Key numbers

  • 3,900,607 public Reddit posts from 32 communities (2008 to 2026)
  • 11,253 flagged as relevant to reinterpretation
  • 123 candidate classification differences (112 to conflicting, 5 uncertain to pathogenic or likely pathogenic, 6 other); candidates are not confirmed reclassifications
  • A clinician reviewed 30 selected drafts: 27 stated the change correctly; only 15 were fit to send
  • About $694 in AI processing fees for one full screening pass

No messages were sent. No users were contacted. Who should press send? Our next step is a stakeholder study with patients, families, clinicians, and labs.

Download

One-page summary (PDF)

Short video (coming soon)

Connect