Screening a national drug register for look-alike names: development, external validation and deployment of an open, explainable model on the Saudi Food and Drug Authority register

Objective. Look-alike drug names are a common, preventable source of wrong-drug errors. Regulators and hospital committees must decide which of millions of name pairs to review, but no openly validated screening tool exists for a national register. We developed and externally validated such a screen and tested it on the Saudi Food and Drug Authority (SFDA) register. Methods. Fourteen orthographic and phonetic similarity measures were combined by regularised logistic regression and gradient boosting, trained on the US Institute for Safe Medication Practices (ISMP) confused-names list (527 pairs) and the ISMP-España list (471 pairs), with name-grouped cross-validation and cross-list external validation. Models trained on one list were then applied to the full 2023 SFDA register (6,541 names): for 262 queries from 131 listed, different-ingredient pairs registered in Saudi Arabia, the partner was ranked against all 6,540 other names (TRIPOD+AI). Results. In external validation, logistic regression reached 80% sensitivity after reviewing 0.93–1.46% of all pairs, versus 1.81–3.38% for the best single measure. In the Saudi register, reviewing the top 20 candidates per name found 56.1% (95% CI 48.1–64.5) of known confusions, versus 46.2% for Levenshtein distance (paired difference 9.9 points, 3.4–16.4) and 48.5% for BI-SIM; 83.5% in simulated 1,000-name formularies. More complex models and hard-negative retraining performed worse. We identified 60 unlisted, highly similar trade-name pairs with different ingredients. Conclusions. A simple, transparent model can screen an entire national register and is ready for local piloting. It supports, but does not replace, expert review. Archive contents: preprint PDF; code; SFDA 2023 register; open review list (top 20 candidates for each of 6,541 names); unlisted high-similarity pairs; results. ISMP lists are not redistributed (copyright); fetch scripts are provided.