Machine Learning Approaches to Pharmaceutical Demand Forecasting: A Systematic Review

Background and rationale The planning and scheduling of pharmaceutical procurement represent a central challenge for both public and private health systems. Inherent uncertainty in demand complicates the definition of appropriate quantities and efficient supply timelines. In the private sector, various strategies are employed to mitigate financial risk and enhance supply chain resilience, such as consignment-based contracts and flexible replenishment mechanisms. In contrast, in publicly funded health systems operating under more rigid legal and regulatory frameworks—such as the Brazilian public health system—these alternatives are more limited, increasing reliance on well-calibrated demand estimates. Inaccurate forecasting may lead to both supply shortages and excess inventory, directly affecting continuity of care and the efficient use of public resources. In this context, pharmaceutical consumption data constitute a strategic asset for public health analyses, enabling the development and application of machine learning techniques to predict demand based on multiple parameters and historical patterns. In recent years, machine learning and advanced quantitative models have been increasingly applied to demand forecasting problems in various sectors, including healthcare and pharmaceutical supply systems. These approaches encompass a wide range of methods, such as traditional statistical time-series models, supervised machine learning algorithms, and deep learning architectures. Despite the growing body of literature, evidence on how machine learning–based models are applied to pharmaceutical demand forecasting remains fragmented across disciplines, including health informatics, operations research, and supply chain management. A systematic synthesis of empirical studies is necessary to characterize the types of models used, data sources, forecasting horizons, performance metrics, and application contexts. This systematic review aims to synthesize empirical evidence on machine learning–based approaches to pharmaceutical demand forecasting in health and pharmaceutical systems. Objectives The objective of this systematic review is to identify, describe, and synthesize empirical studies that develop or evaluate quantitative models, including machine learning and statistical approaches, for forecasting pharmaceutical demand. Review question What empirical machine learning–based and quantitative models have been applied to forecast pharmaceutical demand, and how are these models characterized in terms of data sources, methodological approaches, forecasting tasks, and performance metrics? Eligibility criteria Inclusion criteria • Empirical studies that develop, implement, or evaluate quantitative models for forecasting pharmaceutical or medicine demand • Applications situated in pharmaceutical systems, health supply chains, or medicine procurement and distribution contexts • Studies using real-world or simulated data for demand forecasting • Peer-reviewed journal articles and full conference papers • Publications in any language (with translation assisted when necessary) Exclusion criteria • Studies focused on individual-level medicine consumption, adherence, or prescribing behavior • Clinical drug trials or pharmacological efficacy studies • Image-based artificial intelligence applications (e.g., radiology, pathology, diagnostics) • Studies applying machine learning exclusively to supply chain risk management, resilience, or disruption detection without an explicit demand forecasting component • Purely theoretical or conceptual papers without empirical application • Narrative reviews, systematic reviews, editorials, commentaries, and grey literature Information sources The following electronic databases were searched: • PubMed/MEDLINE • Scopus • Web of Science • IEEE Xplore • LILACS (via the Virtual Health Library – BVS) The inclusion of LILACS aimed to ensure coverage of Latin American and Caribbean literature relevant to pharmaceutical demand forecasting in health systems. Search strategy Search strategies were defined a priori and applied consistently across all databases using an equivalent conceptual structure. Three core concepts were combined using the Boolean operator AND: (1) forecasting or prediction, (2) pharmaceutical demand, and (3) quantitative predictive modeling methods, including machine learning and statistical time-series approaches. The general search structure was: (forecast OR predict* OR “demand forecast*” OR “demand prediction”) AND (“pharmaceutical demand” OR “medicine demand” OR “drug demand”) AND (“machine learning” OR “deep learning” OR AI OR LSTM OR ARIMA OR neural OR model*)* This conceptual structure was applied across all databases. Minor syntactic adaptations were made to accommodate database-specific requirements, without altering the underlying search logic. • PubMed/MEDLINE: Free-text searches were conducted in title and abstract fields. • Scopus and Web of Science: Searches were applied to title, abstract, and keyword fields. • IEEE Xplore: Searches were conducted in metadata fields (title, abstract, and keywords). • LILACS: The same three-concept structure was applied using Boolean operators, consistent with the strategy used in PubMed. No restrictions on publication year were applied. Studies published in any language were eligible. All search strings, execution dates, and numbers of records retrieved will be reported as supplementary material. Study selection All records retrieved from the databases were imported into Rayyan (QCRI) for reference management and duplicate removal. After removal of duplicates, titles and abstracts were screened according to the predefined eligibility criteria. Full-text articles were then assessed for eligibility. A second reviewer independently verified the selection of studies at the full-text stage. Disagreements were resolved through discussion. Reasons for exclusion at the full-text stage were documented and will be reported in the PRISMA flow diagram. Data extraction Data extraction will be performed using a standardized extraction form developed a priori. Extracted information will include: • Bibliographic information (authors, year, country) • Study context and application setting • Data sources and characteristics • Type of forecasting task and time horizon • Modeling approach (statistical, machine learning, deep learning) • Performance metrics (e.g., RMSE, MAE, MAPE) • Key findings and limitations Data extraction will be conducted by the primary reviewer and verified by the second reviewer. Risk of bias and methodological quality assessment Methodological quality of included studies will be assessed using appropriate Joanna Briggs Institute (JBI) critical appraisal tools, selected according to study design. Quality appraisal will be conducted after full-text inclusion and prior to data synthesis. Results of the appraisal will be reported descriptively and considered in the interpretation of findings. The GRADE approach will not be applied, as this review does not assess intervention effects. Data synthesis A narrative synthesis will be conducted. Findings will be organized according to: • Modeling approaches • Data characteristics • Forecasting objectives • Performance metrics and comparative results Where appropriate, summary tables will be used to facilitate comparison across studies. Quantitative meta-analysis is not planned due to expected heterogeneity in models, data, and outcomes. Registration status At the time of protocol registration, title and abstract screening had been completed. Full-text assessment and data extraction had not yet been performed. Dissemination plans The results of this systematic review will be submitted to a peer-reviewed journal in public health, health systems research, health informatics, or pharmaceutical policy and management.