A theoretical framework for structuring machine learning algorithms in response to the unique dynamics of the pharmaceutical supply chain

Purpose This study aims to develop a theoretical framework to guide the selection and structuring of machine-learning algorithms (MLAs) for pharmaceutical supply chains and address challenges such as order variability, inventory levels and drug-specific characteristics. Design/methodology/approach This paper analyzes demand data for 150 drugs across 30 pharmacological groups, identifies the attributes that most influence MLA performance and classifies algorithms according to their suitability for variations in demand for different drugs throughout the pharmaceutical supply chain. Findings The proposed framework discovered the key factors of medications that determine the most suitable algorithms for demand prediction, “Sensitivity to order batching” and “Stock level.” Notably, simpler models tend to perform well for drugs with stable demand and high stock levels, while more advanced MLAs provide substantial benefits for highly variable demand categories. Originality/value This study contributes to the literature by offering systematic and actionable principles for selecting MLAs based on pharmaceutical supply-chain dynamics and medical guidelines, providing practical guidance for more precise, AI-driven demand forecasting.