Answers are grounded in this paper and cite their sources.
Syndrome patterns in chronic gastritis are usually compared across stages by counting patterns or their syndrome elements. Frequency comparison cannot tell whether an element becomes more common because its prevalence changed or because patients combine elements differently. This protocol supplies two inference layers for that question. Composite diagnoses are decomposed into eight binary elements by a two-tier rule set (explicit pattern table, then keyword rules). Pairwise co-occurrence is measured by the lift; the lift is normalized by the marginals but not invariant to them, so its margin sensitivity is quantified rather than assumed away: a margins-only counterfactual moves the key lift by +0.16, a dependence-only counterfactual by +1.87. Layer one is a patient-level stage-label permutation test of no distributional difference (two-sided, 5,000 permutations); that null is stronger than equality of lift, and its significance is not attributable to relational reorganization when margins differ. Layer two, carrying the confirmatory weight, is a margin-preserving parametric bootstrap test: each stage keeps its observed marginal prevalences while latent associations are pooled; it reports a margin-adjusted residual with an interval under Benjamini-Yekutieli control across the 28-edge family. Weighted networks (w = clip(lift - 1, 0, 3)) yield strength centrality, Onnela clustering, and edge density. In the representative simulated cohort (400 patients per stage; seed 20260913), the qi deficiency-blood stasis edge moved +1.00 in lift and +1.33 (95% interval 0.93 to 1.74) beyond its margin-induced component; the drift statistic D moved +2.93 (2.06 to 3.91) beyond a component of -1.03. A descriptive interaction diagnostic accompanies the pipeline; its calibrated criterion fires in 29 of 30 no-injection replicates, so it is not a test. All analyses use simulated data only; the stage contrasts are cross-sectional; a double-annotation audit is specified before real records are used.
No takes yet. Share an insight, caveat, or question.
Li et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: