A Hybrid Pharmacovigilance Method for National-Scale Comorbidity Discovery: Association Rules with FDA-Approved PRR/Chi-square and EBGM Validation.

Background Clinicians need scalable, statistically rigorous maps of disease–disease co-occurrence to support screening, safer prescribing, and differential diagnosis. Spontaneous-report data such as FAERS are national in scope but prone to bias, so discovery requires conservative, interpretable validation. Methods We developed a three-stage pipeline—Association Rule Mining for candidate triage, disproportionality testing with PRR and χ² (criteria: PRR ≥ 2.0, χ² ≥ 4.0, p < 0.05, a ≥ 3), and Empirical-Bayes shrinkage via EBGM (low-count rule: if a < 10 and EB05 < 2, reject). The pipeline was applied to FAERS (n = 393,130; 50 index conditions) and externally evaluated with a lab-test corpus (n = 2,461). Results The workflow yielded 25,083 validated condition–condition associations after ≈ 99% overall rejection of naïve pairs (≈ 85% specificity among retained links). Statistical strength was high: 97.1% of associations met p < 0.001. Risk tiering showed 43.5% of links in a High-Risk band (PRR ≥ 10), with clinically coherent hubs (e.g., hypertension, rheumatoid arthritis, Crohn’s disease, psoriasis). Internal five-fold analysis indicated strong reproducibility, and cross-corpus comparison supported generalizability, with directionally consistent effects and high rejection concordance for spurious pairs. End-to-end performance supports real-time use (< 50 ms for cached retrieval; ~2–3 s on-demand). Conclusions A staged, FDA-aligned pipeline (Association Rules → PRR/χ² → EBGM) converts spontaneous reports into a defensible, reproducible comorbidity network at national scale. The approach reduces false positives without sacrificing sensitivity, aligns with familiar pharmacovigilance statistics, and is production-ready for deployment in clinical decision support (medocsecondopinion.com) while providing a clear queue of novel, testable hypotheses for follow-up in longitudinal data.