Introduction: Semi-supervised learning may reduce labeled-outcome requirements, but confident pseudo-labels can propagate error. We assessed survey-aware self-training for screening smoking-adjusted low hemoglobin using NHANES August 2021–August 2023. Method: The measured cohort included 5,731 adults, of whom 655 met the outcome definition. Predictors comprised demographic, anthropometric, and blood-pressure measures. Outcomes were deliberately withheld from measured respondents after outcome-stratified splitting and mask construction. Five masks provided each fitted model with 20% of training labels and 20% of validation labels; shared tuning used labels from all five masks. Naturally missing-hemoglobin respondents were reserved for descriptive diagnostics. The primary contrast was survey-weighted test average precision (AP) for self-training minus matched supervised boosting, with a 2,000-replicate design-and-mask bootstrap interval. Results and Discussion: The weighted test prevalence was 0.0990. Mean AP was 0.1555 for supervision and 0.1539 for self-training (difference, -0.0016; 95% CI, -0.0076 to 0.0024). Mean Brier scores were 0.08969 and 0.08986 (difference, 0.00017; 95% CI, -0.00105 to 0.00188). Four masks admitted no pseudo-labels and produced identical predictions; the fifth admitted 258 and reduced AP by 0.0080. Conclusion: The comparison was inconclusive. These findings concern the evaluated boosting and admission rules, and the per-model budgets do not represent the total outcome information used in study design and tuning
Copyrights © 2026