Rule-Based NLP Extracts Health Outcomes from YouTube Comments

The authors published a cross-sectional methods study in J Med Internet Res that develops and validates a rule-based natural language processing framework for extracting self-reported health outcomes from Healthcasting YouTube comment threads. According to the PubMed abstract, the peer-reviewed record analyzed 43,111 unique comments from 110 videos across 11 Therapeutic Carbohydrate Restriction (TCR) channels, collected with the YouTube Data API v3 and spanning November 2013 to January 2026. The JMIR preprint version reports a larger corpus of 209,661 comments and 37,742 unique authors and documents a four-phase development workflow and manual validation (n=500).

Per the preprint and PubMed abstract, the pipeline development included exploratory corpus characterisation, iterative construction of a 35-aspect hierarchical health outcome ontology, and a precision-optimised rule-based classifier with multiple validation studies (authors report three construction phases and five validation studies in the abstract). The study emphasises achieving high precision when extracting first-person outcome statements such as weight change and biomarker normalisation from noisy, unstructured comment text.

Rule-based approaches remain viable where label scarcity and the need for precision outweigh the generalisation strengths of large supervised models. For practitioners: rule-based ontologies combined with targeted manual validation can yield high-precision signals useful for outcome surveillance, cohort identification, or hypothesis generation from public social-media corpora.

Industry observers and researchers have been exploring social media as a complementary, real-world evidence source for patient-reported outcomes. The study provides a documented workflow and ontology that other teams can adapt for high-precision needs or when training data for supervised models is limited.

Observers should track whether the authors release the outcome ontology, classification rules, or annotated validation sets, and whether future work compares the rule-based pipeline against transformer-based classifiers for recall-precision tradeoffs. Also watch for replication on other healthcasting communities beyond TCR.

These methods and corpus figures are taken from the PubMed abstract for the J Med Internet Res article (PMID 42077206) and the JMIR preprint (preprint #94855).