Poster #12: Kamalakkannan Ravi PhD

Reclassification Is All You Missed: Surfacing Variant Reinterpretation Need in Online Health Communities at Population Scale

Kamalakkannan Ravi, PhD (Boston Children's Hospital and Harvard Medical School); Mengshu Nie, MA (Boston Children's Hospital and Harvard Medical School); Annapurna Poduri, MD, MPH (Boston Children's Hospital and Harvard Medical School); Cong Liu, PhD (Boston Children's Hospital and Harvard Medical School)

Genetic variants are routinely reclassified in ClinVar as evidence accumulates; a result reported as uncertain may later become pathogenic or likely pathogenic. Patient recontact is inconsistent, with responsibility diffused among laboratories, clinicians, and patients, and updated interpretations often fail to reach the patient. Many affected individuals discuss prior genetic results in public online health communities that clinical recontact does not reach. Methods: We built PRISM, a three-agent large language model pipeline. A triage agent classifies each of 3.9 million public posts (2008 to 2026) from 32 health-related Reddit communities under a fixed instruction prefix, returning a structured label with a confidence and a one-sentence rationale. An extraction agent identifies the gene, variant, and test date, then programmatically resolves the variant's ClinVar classification as of the test date against the current record. An outreach agent drafts a disclaimer-bearing, plain-language message for each candidate reclassification; the study is a no-contact dry run and no message was sent. Each stage was validated on synthetic control sets with known answer keys before any real-post run, and a clinician reviewed a held-out set of drafts. Results: Reinterpretation-relevant discussion is rare (0.288 percent) but concentrated in specialist genetics communities (6.30 versus 0.10 percent in general medical-advice communities) and has grown roughly fivefold since 2014. PRISM flagged 11,253 reinterpretation-relevant posts from 9,395 distinct authors and surfaced 123 candidate reclassifications, 5 from uncertain to pathogenic or likely pathogenic. A clean end-to-end run cost about 700 dollars. Of 30 held-out drafts, the clinician judged 27 factually grounded and 15 fit to send. Conclusion: Population-scale surfacing of variant reinterpretation need from public online communities is feasible and inexpensive; factual accuracy is necessary but not sufficient for clinically acceptable outreach. The work is under peer review, and a stakeholder study of acceptable use is planned.