TY  - JOUR
T1  - Why Better Predictions Can Still Produce Worse Pharmacy Decisions
A1  - Emma Wilson
A1  - Frederik De Jong
A1  - Lara Peters
JF  - Archives of Pharmacy Practice
JO  - Arch Pharm Pract
SN  - 2320-5210
Y1  - 2025
VL  - 16
IS  - 2
DO  - 10.51847/699h8jFRoX
SP  - 44
EP  - 52
N2  - Artificial intelligence in pharmacy is frequently evaluated through prediction-centred measures, including discrimination, calibration, sensitivity, specificity, and alert reduction. These measures are necessary for characterizing model behaviour but do not establish whether the model improves medication-related decisions. A prediction becomes consequential only when it is interpreted by a professional, connected to an available action, introduced at an appropriate time, and applied within a workflow in which errors have asymmetric consequences. This article critically examines the prediction-to-decision translation gap: the possibility that technically improved predictions may produce unchanged or worse pharmacy decisions. Six mechanisms are distinguished: target mismatch, automation bias, treatment leakage, poor timing, actionability failure, and unequal error consequences. Decision usefulness is proposed as a conditional property arising from alignment among the prediction target, patient and medication context, uncertainty communication, professional interpretation, action feasibility, and the consequences of acting or not acting. The article develops a proposed Prediction–Interpretation–Action Framework that separates model output from professional interpretation and medication action through three translation gates: target validity, interpretation integrity, and action suitability. Evaluation should therefore extend beyond model performance to include human–AI interaction, decision consequences, workflow burden, medication safety, subgroup effects, governance, and post-implementation change. The framework is an original conceptual synthesis rather than a validated clinical model, procurement standard, or regulatory instrument. Its propositions require prospective testing across pharmacy settings, medication tasks, professional roles, patient populations, and technology configurations before decision readiness can be inferred.
UR  - https://archivepp.com/article/why-better-predictions-can-still-produce-worse-pharmacy-decisions-zlfbntni0rxhhuc
ER  -