Beyond ratings: a condition-aware natural language processing framework for drug effectiveness decomposition, regulatory benchmarking, safety signal detection, and adversarial robustness assessment using patient-generated reviews

Patient-generated drug reviews offer a large source of real-world evidence on perceived medication effectiveness, yet existing computational analyses predominantly model effectiveness as a single drug-level estimate, ignoring variation across therapeutic indications. This study introduces an integrated four-module natural language processing (NLP) framework that extends patient review analytics from aggregate rating prediction toward condition-aware pharmacovigilance support. Module 1 decomposes drug effectiveness across 1880 drug–condition pairs (623 drugs; 131 conditions), showing that the same medication can differ by up to 4.18 rating points across indications (mean cross-condition range = 0.98 rating points on a 1–10 scale among the 533 drugs observed in at least two conditions). A condition-aware model combining term frequency–inverse document frequency (TF-IDF) features with linear support vector regression (SVR) achieves a coefficient of determination ( R 2 ) of 0.749 (95% confidence interval (CI) 0.731–0.767), comparable to a drug-level model ( R 2 = 0.744), while providing indication-specific estimates. Item-based collaborative filtering imputes unobserved drug–condition pairs with a hold-out root mean squared error (RMSE) of 0.716, no better than a simple additive drug + condition baseline. Module 2 benchmarks patient-derived scores against documented U.S. Food and Drug Administration (FDA) safety actions for 230 matched drugs. FDA-actioned drugs were rated slightly higher (Cohen’s d = − 0.297, 95% CI − 0.431 to − 0.163; area under the receiver operating characteristic curve (AUC) = 0.592), but this difference was largely explained by their much higher review volume: it disappeared with unshrunk ratings ( d = − 0.045) and was attenuated after covariate adjustment (odds ratio 1.12, 95% CI 0.95–1.31), although a small difference persisted among drugs with at least 50 reviews ( d = − 0.285). Module 3 constructs a composite safety concern score from four lexicon-based dimensions. The score discriminated FDA-actioned drugs modestly (AUC = 0.568, 95% CI 0.516–0.619), including among drugs with at least 50 reviews (AUC = 0.593) and with corrected lexicon matching (AUC = 0.574); discontinuation language was the most consistent discriminating feature. The module flags 141 “blind-spot” drugs combining high ratings with high safety-concern language; many flags reflected lexicon terms that name the treated condition, so blind spots are hypothesis-generating only. Module 4 evaluates robustness to simple text perturbations and template-based fake review injection. TF-IDF with SVR and TF-IDF with Ridge regression yield robustness indices of 0.991 and 0.994, respectively; approximately 50 (SVR) and 61 (Ridge; extrapolated) injected reviews were required to shift a prediction by one rating point for sparsely reviewed target drugs. Condition-aware effectiveness estimates and linguistic safety signals can complement, but not replace, established pharmacovigilance sources.