BACKGROUND: YouTube is increasingly used for Healthcasting, the sharing of evidence-based dietary and lifestyle interventions by domain experts. In the metabolic health domain, channels focused on Therapeutic Carbohydrate Restriction (TCR) have accumulated audiences of millions. A distinctive feature is the comment section, where viewers share first-person accounts of health changes, constituting a unique source of real-world outcome data at scale. However, extracting structured health information from unstructured comments presents computational challenges. OBJECTIVE: To develop and validate a precision-optimized computational framework for extracting self-reported health outcomes from Healthcasting YouTube comments, and to characterize the prevalence, distribution across health aspects, and channel-level variation of reported outcomes across a large-scale metabolic health corpus. METHODS: This observational, cross-sectional study analyzed 43,111 unique YouTube comments from 110 videos across 11 TCR-focused Healthcasting channels (37,458 unique authors; data span November 2013 to January 2026; collected via YouTube Data API v3). The methodology comprises three construction phases and five validation studies. The construction phases are (1) exploratory corpus characterization, (2) iterative development of a 35-aspect hierarchical health outcome ontology, and (3) precision-optimized rule-based classification, validated through precision validation (n=500 stratified sample), recall estimation (n=510), external validation on five held-out channels (n=12,653 comments), LLM-assisted inter-rater reliability assessment, and transformer baseline comparison against BERT and RoBERTa classifiers. A supplementary Aspect-Based Sentiment Analysis contextualized the positive-only design. RESULTS: The framework identified 1,790 positive health outcome reports (1790/43111, 4.15% prevalence), achieving 97.6% (488/500) precision (95% CI 95.7%-98.6%) and estimated 56.2% recall (95% CI 43.4%-67.9%). Positive outcomes were distributed across 35 health aspects and 18 named disease conditions extending beyond weight loss: pain and inflammation reduction (1137/6674, 17.0%), type 2 diabetes improvement (977/6674, 14.6%), skin health (784/6674, 11.8%), and psychological well-being (731/6674, 11.0%). Over half (3355/6674, 50.3%) spanned multiple research objectives. Significant channel-level variation was observed (χ²=927.5, P<.001), with positive outcome rates ranging from 1.32% to 10.40% (OR 8.68, 95% CI 7.10-10.61). Transformer baselines achieved higher recall but lower precision, confirming the design advantage for high-confidence corpus generation. A supplementary Aspect-Based Sentiment Analysis indicated a positive-to-negative ratio of approximately 4.6:1 (n=1,003), with negative experiences (11.9%) predominantly involving gastrointestinal and cardiovascular concerns. CONCLUSIONS: This study presents, to our knowledge, the first validated rule-based framework for extracting self-reported metabolic health outcomes from Healthcasting YouTube comments at corpus scale. Unlike existing recall-oriented social media health classifiers, the precision-optimized design achieves the confidence threshold required for outcomes research without manual review. These findings demonstrate that expert-led health content comment sections constitute a scalable, complementary data source for monitoring real-world engagement with dietary interventions, with implications for public health surveillance, platform design, and health communication research.
Ribeiro et al. (Fri,) studied this question.