Healthcare systems can experience product shortages and medical waste at the same time because procurement, expiry monitoring, and redistribution are managed separately. This study develops a circular inventory architecture that links machine-learning risk prediction with constrained inter-facility pooling. The empirical setting is the early United States COVID-19 vaccine program. Public CDC-derived files contain 293,907 vaccine shipment records to 53,041 provider locations and 1,154 usable wastage reports. The analysis constructs 4,240 awardee-manufacturer-week observations and applies strict temporal validation. Logistic regression, random forest, and gradient boosting are compared. The random forest achieved an out-of-time area under the precision-recall curve of 0.672, area under the receiver-operating-characteristic curve of 0.809, and F1 score of 0.626. Shipment frequency, shipment-size dispersion, provider coverage, prior wastage, and dose volume were the most influential predictors. A retrospective redistribution simulation increased the shipment-based fill-rate proxy from 0.815 to 0.864. Risk-based transfer capacity potentially covered 156,097 observed wasted doses, although the public data do not identify wastage cause or completed transfers. The paper contributes an integrated theory of expiry, shortage, and waste, demonstrates an explainable prediction-to-decision workflow, and shows how equity constraints can be added at low incremental scenario cost. Findings support better lot visibility and network coordination, but do not establish causal waste reduction.