Machine Learning Approaches for Pharmaceutical Demand Forecasting: A Bibliometric and Systematic Review of Methods and Research Trends

In the healthcare sector, ensuring the availability of pharmaceutical stock is a critical concern that requires both cost-efficiency and sustainability in procurement and usage [1]. A significant challenge in this domain is demand uncertainty, which complicates inventory management and requires consideration of various risk factors when estimating future requirements [2]. Consequently, demand forecasting has become an essential tool in healthcare logistics and supply chain operations [3]. Forecasting methodologies are increasingly prominent across industries, particularly in pharmaceutical supply chains, where mathematical models are used to predict inventory trends [4]. These models often incorporate variables such as the number of healthcare facilities, patient counts, and daily consumption fluctuations to mitigate demand uncertainty [5]. Forecasting is now recognized as a strategic tool in sales and operations planning, enabling adjustments to production capacity, strategic positioning, and budgeting. It also enhances operational efficiency and mitigates external risks by providing insights into future demand patterns [6].

Despite its widespread application, forecasting in the pharmaceutical industry presents unique challenges due to the sector's sensitivity to fluctuations in political, social, and economic conditions. Accurate forecasting is crucial for effective planning and resource allocation, especially in pharmaceutical supply chains, where misalignment between supply and demand can have severe consequences. These consequences include disruptions in patient care, inefficiencies in drug distribution, and potential public health risks [7, 8]. Currently, many healthcare institutions rely on rudimentary forecasting methods, such as extrapolating annual demand from average monthly usage. Although simple, this approach often yields inaccurate projections, resulting in two primary issues: overstocking and stockouts. Overstocking leads to drug expiration and waste, while stockouts prevent patients from accessing necessary medications [9]. Both scenarios incur material losses, such as disposal costs and emergency procurement, as well as non-material consequences, including compromised patient safety and diminished trust in healthcare systems [10].

Given the financial and operational risks associated with ineffective drug demand planning, there is a pressing need for more accurate and responsive forecasting models. Recent advancements in artificial intelligence (AI) and ML have introduced sophisticated techniques that address the limitations of traditional statistical models. These techniques include linear regression (LR), decision trees, support vector machines (SVMs), random forests (RFs), recurrent neural networks (RNNs), and long short-term memory (LSTM) networks [11, 12]. Time series forecasting remains a foundational approach to demand prediction, relying on historical data collected at regular intervals to estimate future trends [7]. However, the complexity of pharmaceutical demand, characterized by nonlinear patterns, seasonal variations, and external disruptions, necessitates more robust models. ML techniques, particularly those capable of handling large datasets and capturing intricate temporal dynamics, offer promising alternatives.

Recent literature highlights emerging trends in AI and ML within the pharmaceutical sector. Notable contributions include a 2025 review of current and future trends in AI and LR for pharmaceutical supply security [13], a bibliometric overview of computer-aided drug design [14], a review of topic models, and a bibliometric analysis in healthcare applications using AI [15], Additionally, studies have examined bibliometric analysis of healthcare supply chain opportunities and challenges [16], a social network analyses in AI-enabled drug discovery and development [17], and a LR and bibliometric reviews of green and sustainable logistics [18].

This review critically examines the application of LR to pharmaceutical demand forecasting by integrating bibliometric analysis with an evaluation of the employed forecasting methods, a combination rarely addressed in the existing literature. Despite the growing use of LR algorithms such as LR, LSTM, AI, and RFs, current studies often lack methodological integration that connects algorithmic performance with broader research trends and thematic evolution. This gap limits the ability to identify systemic strengths, weaknesses, and opportunities for innovation in forecasting accuracy, responsiveness, and scalability. By mapping scholarly contributions and comparing data sources and algorithmic approaches, this review seeks to uncover these deficiencies and propose future research directions to advance healthcare supply chain management through more robust, data-driven strategies that mitigate demand uncertainty and optimize pharmaceutical logistics.

Figure 4 illustrates the Forecasting Methods Employed in Pharmaceutical Demand Forecasting through LR. The reviewed literature reveals a broad spectrum of forecasting techniques applied to pharmaceutical drug demand, ranging from conventional statistical models to advanced LR algorithms. Among the most frequently employed methods are RF, LSTM, Artificial Neural Network (ANN), XGBoost, and LR. RF is widely acknowledged for its robustness in handling high-dimensional, nonlinear data; studies by Nabizadeh et al. [38], Mbonyinshuti et al. [9], and Haoudi et al. [39] confirm its efficacy. LSTM, a deep learning model designed explicitly for time series forecasting, has consistently demonstrated superior performance over traditional models, such as ARIMA, in capturing temporal dependencies, as noted by Mbonyinshuti et al. [12]. ANNs, used in various configurations, have proven effective for modelling complex relationships in data [40, 41], while XGBoost, a gradient boosting framework, has demonstrated high predictive accuracy in studies by Fourkiotis and Tsadiras [11].

Figure 4. The forecasting methods employed

The selection of forecasting methods is primarily influenced by the nature of the dataset and the complexity of demand patterns. Simpler models, such as LR, are often applied to datasets with linear trends; however, they are generally outperformed by more sophisticated models when dealing with dynamic, seasonal, or nonlinear data. LR models were particularly favored for their adaptability and precision in handling variability and uncertainty in pharmaceutical demand forecasting. This methodological diversity highlights the importance of tailoring forecasting techniques to specific data characteristics and operational requirements to achieve the most accurate predictive outcomes.

3.1 The long short-term memory

LSTM networks are particularly advantageous for modelling complex temporal patterns due to their ability to learn nonlinear relationships and dynamically adapt to factors such as market trends, seasonal fluctuations, and economic indicators [28, 42]. For instance, traditional statistical methods often fail to capture the intricate dynamics of pharmaceutical demand. Comparative studies consistently demonstrate that LSTM networks outperform traditional models, such as ARIMA, particularly when external variables are incorporated into the forecasting process, resulting in greater accuracy [34].

The utility of LSTM models is further enhanced when combined with external variables. Chen [42] highlighted that incorporating variables such as promotional activities, market conditions, and broader economic indicators significantly improves the accuracy and robustness of LSTM models. This integration reflects the growing recognition within the pharmaceutical industry of the need to use comprehensive datasets to improve demand predictions [19, 34]. Practical applications further underscore LSTM's relevance in pharmaceutical demand forecasting. Yani and Aamer [28] noted that adopting LR techniques, particularly LSTM, has led to substantial improvements in forecast precision, thereby helping mitigate common disruptions in pharmaceutical supply chains. These findings align with broader research emphasizing the transformative potential of AI and LR in the health and pharmaceutical sectors, suggesting that improved demand forecasting can enhance the availability of medicines and patient outcomes [11, 34]. Additionally, LSTM's versatility extends beyond basic demand forecasting; it can be tailored to optimize performance across different types of pharmaceutical products by adapting its architecture and training parameters to the unique characteristics of each segment [43]. This adaptability makes LSTM a practical and strategic tool in pharmaceutical logistics and the optimization of patient care.

A pivotal finding from the discussion is that LSTM networks not only outperform traditional statistical models such as ARIMA in pharmaceutical demand forecasting but also demonstrate transformative capabilities when integrated with external variables, such as promotional activities and economic indicators. This synergy enables LSTM models to dynamically adapt to complex, nonlinear market dynamics, significantly enhancing forecast accuracy and resilience against supply chain disruptions. Their architectural adaptability to different pharmaceutical product types further underscores their strategic value, solidifying LTSM as a cornerstone technology in the evolution of data-driven healthcare logistics and patient care optimization.

3.2 Artificial neural networks

ANN has gained increasing recognition for its effectiveness in pharmaceutical demand forecasting. The integration of AI and ML, including ANNs, into forecasting systems holds significant potential to enhance accuracy, optimize inventory levels, and improve overall supply chain efficiency in the pharmaceutical sector. Rathipriya et al. indicate that various neural network models, including Radial Basis Function (RBF), Generalized Regression Neural Network (GRNN), and Probabilistic Neural Network (PNN), are commonly applied for time-series pharmaceutical data forecasting [10]. These models capture nonlinear relationships in the data, enabling more precise demand predictions than traditional statistical methods. Furthermore, the study emphasizes the advantages of deep learning approaches, an evolution of ANN, which can improve forecasting accuracy by uncovering complex data patterns that simpler models often overlook.

In a broader context, Angula and Dongo's systematic review further emphasizes the transformative role of AI and ML in enhancing medication demand forecasting in public pharmaceutical systems. Their research suggests that incorporating ANN into forecasting models can improve cost efficiency, accessibility, and inventory management [34]. This work highlights the impact of AI on public health outcomes by enabling more efficient resource allocation and reducing waste. This finding suggests that incorporating ANN into demand forecasting frameworks can significantly alleviate management challenges faced by healthcare systems. Yani and Aamer emphasize the ongoing need for advanced ML techniques in the pharmaceutical supply chain. Their research indicates that deploying LR algorithms, particularly ANN, can help mitigate supply chain disruptions [28]. Although forecasting challenges persist in this industry—including unpredictable demand fluctuations, a structured application of ANNs has the potential to offer adaptable solutions that enhance supply chain resilience.

Moreover, the conceptual framework proposed by Adeleke and Samuel underscores the importance of AI-driven predictive analytics for optimizing pharmaceutical supply chains [19]. Their findings demonstrate that integrating LR techniques, including ANN methodologies, into supply chain management strategies enhances decision-making and operational efficiency. This ongoing evolution of supply chain practices, amid stringent regulations and logistical constraints, illustrates the adaptability and efficacy of ANN in this context. Additionally, Fourkiotis and Tsadiras explore the application of LR and statistical forecasting methods for pharmaceutical sales predictions, underscoring the range of methodologies, including ANN, that can improve accuracy [11]. They posit that combining traditional statistical methods with advanced neural network techniques enables businesses to address market demands better. The literature indicates a growing consensus that a transformational wave is emerging through the use of AI and ANN in pharmaceutical demand forecasting. As supply chain management becomes increasingly complex due to external factors, the need for robust forecasting methods becomes paramount. This sentiment is supported by recent studies advocating for the adoption of ANN as a standard practice in future pharmaceutical demand forecasting [44-46].

A profound insight from the discussion is that ANN, particularly advanced variants such as RBF, GRNN, and PNN, are not merely enhancing pharmaceutical demand forecasting—they are redefining it by enabling systems to learn and adapt to complex, nonlinear patterns that traditional models often overlook. This evolution is especially critical in public health contexts, where accurate forecasting directly influences the accessibility, cost efficiency, and inventory management of medical services. The convergence of ANN with deep learning and predictive analytics frameworks empowers healthcare systems to anticipate disruptions, optimize supply chains, and make data-driven decisions within regulatory and logistical constraints. Consequently, ANN is emerging not just as a forecasting tool but as a strategic asset that transforms pharmaceutical operations and improves patient outcomes.

3.3 Extreme gradient boosting

Extreme Gradient Boosting (XGBoost) has emerged as a robust LR technique, particularly effective for demand forecasting within the pharmaceutical industry. The growing complexity of pharmaceutical supply chains, driven by factors such as regulatory compliance, demand variability, and logistics, necessitates accurate predictive analytics, where techniques like XGBoost can deliver substantial benefits. XGBoost is particularly well-suited for forecasting with limited datasets, making it advantageous for pharmaceutical companies that may lack extensive historical data. For instance, Abdurohman and Putrada [46] highlighted XGBoost's effectiveness even with short datasets, generally requiring less data than other deep learning methods, such as LSTM. This feature aligns well with the pharmaceutical domain, where data may often be sporadic due to the nature of drug production and market demands. Additionally, the model's inherent capacity to handle various data types, including categorical and continuous variables, enhances its adaptability across diverse forecasting scenarios [19].

Moreover, the pharmaceutical industry is experiencing unprecedented fluctuations in demand, as noted by Adeleke and Samuel [19]. Advanced AI-driven frameworks, such as those based on XGBoost, can improve decision-making and risk management, enabling pharmaceutical companies to adapt more rapidly to supply chain disruptions. Various studies have highlighted XGBoost's strong performance in scenarios characterized by uncertainty, thereby reinforcing its relevance for pharmaceutical demand forecasting, as reported by Mishra et al. [47]. Recent research has revealed that integrating XGBoost within broader supply chain management theories can significantly streamline operations. For instance, Mishra et al. conducted comparative analyses that established XGBoost as a leading algorithm for demand forecasting under uncertainty, outperforming several traditional methods. Furthermore, Pan [48] demonstrated that improvements in forecasting capabilities can directly improve inventory management and reduce costs, critical outcomes for pharmaceutical companies navigating volatile markets.

Given its significant advantages, it is critical to understand the specific contexts in which XGBoost excels. The combination of XGBoost with other models (such as SSA-XGBoost) has demonstrated improvements in forecasting accuracy by capturing complex patterns embedded in historical demand data [49]. These hybrid approaches provide pharmaceutical companies with deeper insights into demand trends, enabling them to optimize inventory levels more effectively and enhance service delivery. Angula and Dongo's [34] systematic review underscores the transformative potential of AI and LR for medication supply forecasting in public pharmaceutical systems, a subsector that stands to benefit significantly from improved demand predictability. As a central component of such AI methodologies, XGBoost plays a crucial role in enhancing efficiency and resource allocation across public and private pharmacy networks.

A particularly compelling finding from this analysis is the recognition of XGBoost as a highly effective forecasting tool, particularly in the pharmaceutical industry. This effectiveness stems from its ability to deliver accurate predictions even with limited and heterogeneous datasets, a common challenge in pharmaceutical supply chains. Unlike deep learning models that typically require extensive datasets, XGBoost thrives in data-constrained environments, making it especially valuable for forecasting in emerging markets or niche drug categories. Its flexibility in handling both categorical and continuous variables, along with its robustness in the face of uncertain and volatile demand conditions, positions XGBoost not only as a technical solution but also as a strategic enabler of agile decision-making. When integrated into hybrid models and broader supply chain frameworks, XGBoost empowers pharmaceutical companies to optimize inventory, reduce operational costs, and enhance service delivery, marking a significant shift toward intelligent, data-driven operations in healthcare logistics.

3.4 Linear regression

LR is a fundamental statistical technique for demand forecasting, particularly within the pharmaceutical industry. Its simplicity, interpretability, and computational efficiency make it a widely adopted method for predicting demand based on historical data. The foundational principle of LR is to model the relationship between independent variables and a dependent variable, using the least-squares method to minimize discrepancies between observed and predicted values [50]. In the pharmaceutical sector, LR effectively addresses challenges, such as optimizing inventory levels and anticipating market fluctuations. Studies demonstrate that incorporating key variables, such as pricing, promotional efforts, and seasonal trends, into LR models can significantly improve forecasting accuracy [51]. For instance, research has shown that accurate demand forecasting directly affects logistics costs and resource planning in the pharmaceutical supply chain, underscoring the importance of robust predictive models [28].

The practical application of LR in pharmaceutical demand forecasting is supported by empirical evidence. Research indicates that LR can accurately predict sales performance by analyzing historical data, achieving high accuracy rates up to 98.505% in some instances [52]. Moreover, the method's transparency and ease of explanation make it appealing to stakeholders who require clear insights into demand drivers [53]. By using LR, companies can improve their demand predictions and align production and supply chain strategies accordingly [54]. While LR offers substantial advantages, it is essential to acknowledge its limitations, particularly those related to the assumptions of linearity and error independence. Advances in LR, such as integrating momentum-based gradient descent, have been proposed to accelerate convergence and improve forecasting performance [55]. Furthermore, comparisons with more complex LR methods suggest that LR can outperform these models in specific scenarios, especially when the underlying relationships in the data are approximately linear [56].

A key insight from this discussion is that, despite its simplicity, LR remains a powerful and highly interpretable tool for pharmaceutical demand forecasting, especially when the data relationships are linear. Its ability to incorporate key variables, such as pricing, promotions, and seasonality, allows actionable insights that directly influence inventory optimization and cost management. Notably, empirical evidence indicates that LR can achieve remarkably high accuracy up to 98.505% in specific pharmaceutical contexts, challenging the assumption that only complex LR models can deliver high-performance forecasting. Moreover, enhancements such as momentum-based gradient descent further enhance its effectiveness, suggesting that LR, when applied thoughtfully, can compete with more advanced techniques while maintaining transparency and building stakeholder trust in decision-making processes.

3.5 The random forest

RF method has become a key LR technique for pharmaceutical demand forecasting, providing innovative solutions to the industry's complex challenges, including fluctuating demand, regulatory constraints, and the need for precision in inventory management. This discussion synthesizes recent findings and applications of RF in this domain, highlighting its strengths and emerging trends. As a powerful ensemble learning technique, RF improves forecasting accuracy by combining the predictions of multiple decision trees. Research by Yani and Aamer [28] emphasized that accurate demand forecasting in the pharmaceutical supply chain significantly enhances recovery planning and resilience, asserting that RF is one of the leading LR algorithms for this purpose [28]. Additionally, empirical analysis by Taparia et al. [51] indicated that RF outperformed other regression algorithms in retail contexts, achieving the lowest mean absolute percentage error (MAPE) of 8%, which is crucial in demand forecasting. This capability is increasingly recognized in the pharmaceutical sector, as RF can adapt to the nonlinear relationships characterizing demand patterns, which is especially valuable.

Moreover, RF's adaptability is particularly suited for pharmaceutical supply chains, which are often influenced by external factors such as public health crises and seasonal variation trends. Adeleke and Samuel [19] elaborated on this adaptability, advocating for the integration of AI-driven analytics, specifically RF, into pharmaceutical supply chain management to optimize operational efficiency. This perspective aligns with a systematic review by Angula and Dongo [34], who identified RF as a prevalent model for predicting medication quantity demand in public health systems, highlighting its relevance for ensuring reliable supply in complex healthcare environments. The predictive power of RF is further enhanced by incorporating other ML techniques into hybrid models. For instance, Wang et al. [57] proposed a hybrid model integrating RF with LSTM networks, which significantly improves forecasting accuracy for utility demand. Furthermore, combining RF with neural networks has shown promise in effectively handling time-series data, a common challenge in pharmaceutical demand forecasting [58]. Research confirmed that RF models have delivered superior results, particularly in high-stakes situations such as pandemics or the management of critical medications during shortages [30]. The urgency of efficient demand forecasting in healthcare settings, especially hospitals, is underscored by Tufa et al. [59], who noted that pooled demand management strategies are essential for optimizing inventory control and reducing waste, particularly for essential medications, for which availability is crucial to public health [59].

A key takeaway from this discussion is that RF has emerged as a cornerstone in pharmaceutical demand forecasting due to its exceptional ability to model complex, nonlinear relationships and adapt to volatile market conditions. Its ensemble structure, which aggregates multiple decision trees, not only enhances predictive accuracy but also mitigates the risk of overfitting, making it particularly effective in volatile healthcare environments, such as those experienced during pandemics or drug shortages. Moreover, RF's flexibility is further enhanced when integrated with hybrid models, such as those incorporating neural networks or LSTM, enabling the precise handling of time-series data and external variables. This adaptability makes RF an invaluable tool for optimizing inventory, minimizing waste, and ensuring the timely availability of critical medications, thereby positioning it as a key enabler of data-driven healthcare logistics and policy planning.

3.6 Comparison of previous research based on machine learning algorithms and data sources

Table 1 presents a comparison of previous research using LR algorithms and various data sources. The integration of LR and forecasting techniques into pharmaceutical demand prediction has led to transformative advancements across healthcare and supply chain management. Studies by Nabizadeh et al. [38] and Fourkiotis and Tsadiras [11], for example, demonstrate how predictive models can accurately forecast hospital demands and sales trends. These models not only help reduce waste and prevent shortages but also enable more responsive and agile strategies to cope with fluctuating demand. Notably, the growing use of hybrid approaches that combine statistical methods with LR offers an ideal balance of accuracy and interpretability, key factors for decision-makers in the healthcare sector. The benefits of predictive analytics extend beyond demand forecasting to areas such as environmental and operational efficiency. Kermet-Said et al. [40] applied neural networks to wastewater treatment, demonstrating how AI can optimize pharmaceutical plant operations while ensuring regulatory compliance. Similarly, Li et al. [60] introduced a decomposition-ensemble methodology to forecast emergency medicine demand, which proves invaluable during health crises. These studies illustrate the versatility of LR, not only in predicting demand but also in supporting sustainability and enhancing emergency preparedness.

On the supply chain front, research by Emmanuel et al. [61] and Siddiqui et al. [62] highlights the transformative potential of LR in logistics, particularly during disease outbreaks. By integrating epidemiological data and hybrid models, these approaches enable dynamic, adaptable supply chains that respond to changing real-world conditions. Emmanuel et al.'s [61] systematic review further emphasizes the growing role of AI in building resilient pharmaceutical networks that can withstand disruptions and ensure uninterrupted care. Finally, the landscape of predictive analytics in healthcare continues to evolve rapidly. Surveys by Chen et al. [63] and Badawy et al. [64] illustrate the potential of deep learning in handling long time-series data and complex healthcare scenarios. These technologies are not just tools; they are becoming strategic assets. From traditional medicine forecasting in Indonesia to optimizing insulin distribution, the studies collectively suggest a future in which data-driven insights guide every aspect of pharmaceutical planning, leading to a more efficient, equitable, and intelligent healthcare system.

Table 1. Comparison of previous research based on LR algorithms and data sources

Note: LR = linear regression; ML = machine learning; RF = random forest; DL = deep learning; NN = neural network; COD = Chemical Oxygen Demand.