Analytical approaches for medication management for patient safety: a scoping review Authors: Xinyu Yao, Amogh Rao, Rema Padman Published: 2025 ABSTRACT AbstractSafe medication management is a cornerstone of high-quality patient care, yet it remains one of the most complex and challenging components of healthcare delivery. This scoping review investigates analytical approaches for patient safety-focused medication management, with an emphasis on artificial intelligence and decision-theoretic methods. Guided by PRISMA-ScR, we searched PubMed, Web of Science, ACM, INFORMS, IEEE, and CINAHL for English-language studies published through December 2023. We identified 64 unique studies, over half of which were published since 2020, with 27% based in the United States. Across all studies, 49% addressed adverse drug event detection or prediction, 24% focused on medication error detection, and 13% examined medication reconciliation or review. The remainder explored related areas such as polypharmacy management, dosage adjustment, and medication recommendation. Although artificial intelligence methods were widely applied, only 7 studies used decision-theoretic frameworks, and just 5 combined both approaches. Overall, the literature reflects a predominant focus on single-drug event detection but limited research on managing incomplete or inaccurate medication lists across patient encounters. Future work should explore the combination of predictive, prescriptive, and generative analytics by integrating traditional machine learning, generative AI, and decision-theoretic methods to advance a more comprehensive safe medication management.Subject terms: Computational biology and bioinformatics, Drug discovery, Health care, Mathematics and computing, Medical research Introduction Medication management is the systematic process of ensuring patients receive the right medications, in the correct doses, via the appropriate route, at the proper times, and for the intended duration1. This fundamental healthcare practice encompasses the entire continuum of prescribing, dispensing, administration, monitoring, and reconciliation2. Its primary goals are to achieve optimal therapeutic outcomes, prevent harm, and support patient adherence. At its core, effective medication management requires accurate patient medication lists, understanding of drug actions and interactions, and coordination among healthcare providers and patients. Despite its critical importance as a cornerstone of high-quality and safe patient care, medication management remains one of the most complex and challenging components of healthcare delivery. Contemporary healthcare delivery systems continue to suffer from preventable medication errors that lead to serious adverse consequences for the patient, including adverse drug events (ADE), unnecessary hospitalizations, and increased healthcare costs3. Medication-related injuries harm more than 1.3 million people annually in the US alone4. These risks escalate significantly with polypharmacy, defined as the concurrent use of five or more medications, increasing the likelihood of drug interactions, side effects, and non-adherence5,6. The medication management lifecycle faces multifaceted challenges that compromise patient safety. One such challenge is the emergence of harmful prescribing cascades, where medication side effects are misdiagnosed as new medical conditions, leading to the addition of another, potentially avoidable, medication7. Such risks could be further amplified by barriers to medication adherence and discontinuities in care, which increase the likelihood of medication errors8. Within this context, medication reconciliation plays a pivotal role as a formalized safety process to prevent errors during transitions of care. As defined by the Joint Commission, reconciliation involves creating and verifying an accurate medication list through comparison of existing and newly prescribed regimens across care interfaces9,10. Though essential for resolving both clerical errors (omissions, duplications) and clinical mismatches (interactions, inappropriate dosing), reconciliation constitutes merely one element within holistic medication management. A complete approach demands continuous coordination across providers and settings, coupled with sustained patient engagement, to ensure therapeutic regimens remain appropriate as clinical needs evolve and to mitigate preventable ADE risks11,12. Despite this recognition, the whole medication management process remains vulnerable to implementation barriers, including manual data collection burdens, fragmented information systems, and inconsistent stakeholder communication, all contributing to undetected discrepancies and preventable errors13,14. Traditional medication management approaches depend heavily on error-prone manual workflows15, exacerbated by care fragmentation and inadequate information exchange across settings15. Efforts to improve this have primarily focused on enhancing workflow tasks and integrating information technology (IT) solutions to ensure accurate documentation and smooth transitions between care settings6,16. Nevertheless, these process-oriented tasks face continuing challenges with low consistency and standardization of application, fallible memories of patients, incomplete or inaccurate medical histories, poor collaboration during the transition, and failure to maintain the process after initial implementation17–19. Besides, primary care physicians frequently bear primary responsibility for maintaining accurate medication records20. However, medication management extends substantially beyond primary care to include specialist consultations, outpatient procedures, and community pharmacies. Given the growing complexity of medication regimens, comprehensive medication management processes should occur systematically during every clinical encounter to ensure patient safety and appropriate prescribing21. Healthcare digitization offers unprecedented opportunities to address these systemic challenges. Electronic health records, prescription databases, and clinical systems generate rich data streams that enable advanced analytics. In this scoping review, we focus on Artificial Intelligence (AI) and Decision-Theoretic (DT) methodologies as they demonstrate transformative potential for optimizing medication management through automated data extraction and discrepancy detection, predictive risk modeling for adverse events and polypharmacy complications22, workflow optimization across the medication continuum23,24, and continuous learning systems that adapt to evolving patient needs25. These technologies can streamline the medication management process, significantly reducing the time required and minimizing the risk of human error. AI enables proactive identification of high-risk medication combinations and patients susceptible to ADEs25, and DT can provide robust analytical approaches to optimize the clinical workflow and decision-making processes. This includes the development of algorithms and models that can identify potential discrepancies and suggest corrective actions24. Together, advanced AI and DT analytics can transform the way medication management is conducted, making it more reliable, efficient, scalable, and ultimately enhancing patient safety by shifting focus from error correction to error prediction and prevention. Previous literature reviews have examined the effectiveness of medication management interventions in clinical settings26 or the evidence related to screening for medication discrepancies in patient care27,28. However, there has been no explicit review of the problem of improving the medication management for patient safety with AI and DT methods for risk prediction and associated analytics. Other systematic reviews have focused on limited patient cohorts, such as error reduction or interventions to improve medication safety among specific disadvantaged groups29–31, or specific health outcomes32,33, such as length of stay34, adverse drug events35, and hospital readmissions36. These systematic reviews indicate a dearth of generalizable, high-quality evidence on the impact of medication management interventions using advanced modeling and analytics enabled by AI and DT. Due to the fast-growing use of AI and DT in healthcare and other domains, there is a timely opportunity to understand the factors, data sources, and measures used effectively in AI and DT-enabled medication management to improve care quality and health outcomes23. This scoping review aims to map the existing research landscape regarding the application of AI and DT in addressing the medication management problem from a patient safety perspective. We seek to identify and categorize the various methodologies, tools, and frameworks that have been developed and to assess their impact on reducing medication discrepancies, preventing adverse events, and improving patient safety. We explore a wide range of AI and DT applications, from basic rule-based systems to sophisticated machine learning and optimization models, to understand real-world implementation and effectiveness in mitigating medication-related harm. Results Overall characteristics of the studies included in this review We summarize the characteristics of the 64 included studies in Table 1. The temporal distribution of the studies included in this review spans a duration of seventeen years, from 2007 to 2023. Early contributions are seen in 2007 with 2 studies, and the number gradually increases to 11 and 13 studies in 2022 and 2023, respectively. We observe that a significant proportion of the studies (45%) published between 2021 and 2023 reflect a recent surge in medication management-related research activity. This suggests a growing focus and ongoing research in the field of medication errors and adverse events and, thus, medication management for patient safety. Given the rapid increase in the capture and storage of digital data in clinical settings, we also discern a renewed interest in the field with the use of AI-related methodologies. Table 1Key characteristics of selected studiesCharacteristicsn%Publication Year2021–202329452016–202020312011–201512192007–201035Major data sourcesEHRs4367Drug knowledge databases2031ADE reporting systems1117Insurance claims46Data cohort typesOutpatient primary care69Outpatient specialty care1828Surgery/inpatient1930Not specified2742Data duration (year)(0,1]813(1,5]2539(5,10]914(10,20]35Not specified1930Application StatusPrototype35Experiment5078Deployed1117 Key characteristics of selected studies The selected studies primarily relied on Electronic Health Records (EHR) data, with 67% using this source. Drug databases were utilized in almost a third of the studies (31%), showcasing the integration of clinical pharmacology into algorithmic approaches. Adverse Drug Event (ADE) reporting systems featured in 17% of the studies, emphasizing the need to track and mitigate medication-related risks. The data cohort types also varied significantly: 30% of studies focused on surgery/inpatient settings, reflecting the critical need for precise medication management in these complex environments; 28% targeted outpatient specialty care, addressing the diverse medication needs of these patients. However, only 9% of studies covered primary care settings, indicating a lack of sufficient research on medication management in this fundamental area of healthcare delivery. This distribution demonstrates the varied approaches required for effective medication management while highlighting the need for more research in primary care. In terms of data sample sizes, the studies also demonstrated considerable variation, shown in Fig. 1. 14 studies employed sample sizes in the range of 100 to 10,000 patients, reflecting a focus on moderately sized datasets. Smaller datasets with 0 to 100 patients were used in 5 studies, likely for initial exploratory research. Larger datasets, containing 10,000 to 1,000,000 patients, were leveraged in 12 studies, while 8 studies utilized very large datasets ranging from 1,000,000 to 10,000,000 patients. This diversity in data sample sizes highlights the increasing capacity for large-scale data analytics in patient safety-focused medication management-related quantitative research, enhancing the robustness of patient medication management insights. Histogram of patient population size in the included studies (log scale). Regarding data duration, several studies collected data for 1 to 5 years (39%), while only 5% of studies collected data for more than 10 years. The most commonly encountered study duration was two years, represented by 9 studies, suggesting a preference for medium-term studies on this research topic. Although longer-term studies likely provide more comprehensive insights, they are fewer in number due to the extended time commitments and resources required. Most studies were in the experimental phase (80%), which focuses on validation and testing within the context of medication management for patient safety or related topics using AI/DT methods. Fewer studies were in the prototype stage (5%) or had reached deployed status (16%), indicating a continuing need for medication management research to be deployed and evaluated in real-world clinical settings. Geographic regions of the selected research The selected set of papers analyzed in this review reflects a broad international interest in the subject matter. The United States leads with the highest number of studies, contributing 17 papers, closely followed by Sweden with 14 studies. France and the Greater China region also contribute a number of papers, with 6 and 5 studies, respectively. Denmark follows with 3 studies, while several other countries, including Australia, Canada, India, Iran, Japan, and Turkey, each contribute 2 studies. The country of origin was not clearly specified in 5 studies. At the continent level, as shown in Fig. 2, Europe has the highest number of papers related to medication management topics using AI and DT methods, followed by North America, Asia, and Australia, in that order. No related papers were found from Africa and South America. Distribution of the selected studies across different continents. We note that only 3 papers (5%) utilize data from different geographic regions, indicating a lack of geographical diversity for medication management studies37–39. This limitation also presents challenges to the generalizability and applicability of the proposed algorithms and models in the literature. When algorithms are developed and validated using data from a single region, they may not perform as effectively in different regional healthcare settings with varying patient populations, healthcare practices, and data characteristics. This geographical constraint underscores the need for more collaborative, multi-regional studies to enhance the robustness and versatility of medication management solutions. Expanding research to include diverse data sources from multiple regions will help ensure that the developed models are more universally applicable, ultimately leading to improved patient safety and medication management in varied healthcare environments. Topics and analytic approaches used in the selected papers The research topics of the selected papers reveal several key priorities. ADE Detection, at 27%, is the largest area of focus, highlighting its critical role in ensuring medication management by identifying recorded adverse events in the medical history38–54. Medication error detection follows closely at 23%, which emphasizes identifying and correcting errors in medication administration2,6,37,55–65. ADE prediction, representing 22%, reflects proactive efforts to foresee and prevent adverse events before they occur66–79. Polypharmacy management, at 9%, addresses the complexities involved in managing patients who are prescribed multiple medications, which is crucial for reducing drug interactions and improving outcomes80–85. Studies directly addressing medication management account for only 8%, indicating that this critical area remains understudied86–90. This limited focus suggests a need for more comprehensive research to ensure accurate, comprehensive, and consistent medication records across different healthcare settings for every patient, which is essential for minimizing discrepancies and enhancing patient safety. Medication review, constituting 5%, focuses on the regular assessment of patients’ medication regimens to ensure their appropriateness and effectiveness91–93. Finally, Drug dosage adjustment94,95 and medication recommendation96,97 each make up 3% of the selected studies, highlighting the need for tailored dosage modifications and personalized medication advice to optimize treatment. The topic distribution of the selected papers is summarized in Fig. 3. Pie chart of the topics covered in the selected papers. We also summarize the analytic approaches used in the selected literature in Fig. 4. The methods have been categorized into the following groups: tree-based ML models, regression-based ML models, support vector machines, Natural Language Processing (NLP) algorithms, non-NLP deep learning (DL) models, and decision-theoretic methods. The graph shows that the most frequently used methods are tree-based approaches, which are featured in 33 studies. NLP techniques (i.e., word tokenization, sentence representation, entity alignment, light-weighted language models, etc.) are used in 27 studies, highlighting the increasing importance of text and language data in EHRs and clinical notes. However, as of the end of 2023, we identified only one selected paper that applies Large Language Models (LLM) for medication error classification, highlighting a significant research opportunity in medication management for patient safety with LLMs56. Non-NLP DL methods (i.e., Feedforward Neural Networks, Recurrent Neural Networks, Convolution Neural Networks, etc.) are used in 21 studies. Regression-based methods are used in 13 studies, and SVM is featured in 9 studies. Decision-theoretic approaches are employed in 7 studies, illustrating the limited use of formal decision-making frameworks in medication management research. The details of analytic approaches are discussed further in the Discussion Section. Algorithms used in the selected literature. Synthesis of findings and their Implications Beyond a descriptive summary of the findings from the papers, we have integrated the distribution of the methods within each topic area of the papers in a synthesis heatmap plot shown in Fig. 5. In the heatmap, the color intensity of each block corresponds to the prevalence (as a percentage) of a specific method within the papers of a given topic. Our synthesis of methods that were applied within each topic and between topics reveals distinct patterns. Distribution of topics and methods. With AI approaches, tree-based models are the most widely adopted method, with more than 50% of the papers applying this method in nearly all topic areas, which underscores their role as a default choice that balances predictive accuracy with clinical interpretability. The significant cross-topic dominance of tree-based models directly reflects clinicians’ preference for transparent models, despite the availability and often better performance of newer deep learning models. Their outputs, including data splits for training and testing and feature importance results, are easily interpretable and actionable, enabling clinicians to understand why an alert or prediction is generated and what action may be best for the specific context. This aligns with safety-critical requirements in clinical decision-making, where accountability and trust are paramount98. Moreover, we find that most tree-based models in the selected papers are applied to datasets with fewer than 100 features, facilitating better tree-based search as well as reflecting data limitations that hinder the training of more complex deep learning models. NLP methods cluster strongly around ADE detection, prediction, and reconciliation, reflecting reliance on unstructured clinical text, while non-NLP-based deep learning appears sporadically and remains a secondary approach that is applied when ample data is available. This finding implies that “black-box” deep learning methods, while powerful, face adoption barriers because their recommendations cannot be easily explained. The strong presence of interpretable models therefore reflects not only methodological tradition but also the practical need for clinician acceptance and integration into real-world workflows. Finally, regression methods and SVM are used only occasionally, suggesting their diminishing role compared to tree-based methods when the number of features is limited. Decision-theoretic methods are applied in only 7 of the 64 papers, highlighting their overall scarcity of studies. Within this limited set, 5 studies combine AI and DT, showing that hybridization is possible but currently restricted to narrow areas such as medication review, reconciliation and polypharmacy. Detection-oriented tasks like ADE detection and prediction remain almost entirely AI-driven. Several barriers help explain both the small footprint of DT and its limited integration with AI. Expertise in AI and operations research often remains siloed, and meaningful collaboration in patient management is typically feasible only within large, resource-rich health systems or academic medical settings99. The two approaches traditionally demand different data types: AI can process large and noisy EHR or text corpora, whereas DT requires structured and validated parameters. A natural integration paradigm, predict-then-optimize, illustrates the tension where predictive models generate forecasts from unstructured data, which are then passed to optimization frameworks for prescriptive decision-making, but ensuring consistency between predictive accuracy and optimization feasibility is nontrivial. Evaluation adds a further burden, as hybrid systems must undergo both retrospective validation of predictive models and prospective simulation of decision processes. Finally, health IT infrastructures tend to separate predictive tools from prescriptive ones, reinforcing institutional inertia. Together, these factors contribute to why DT overall is rarely applied and why AI-DT hybrids, although proportionally common within DT, remain few in absolute number. New methodological approaches that combine the two for decision-aware learning are increasingly being developed and evaluated in many domains100,101. Empirical studies show improved performance compared to predict-then-optimize approaches and is a promising area of research for improving medication management, particularly for incorporating the solutions into clinical workflows102–104. In summary, our result confirms the diverse methodologies and topics explored in existing studies. However, there remains a significant need for the integration of advanced algorithms into clinical information systems. Additionally, the generalizability of these algorithms is limited due to the reliance on data from single or dual EHR systems within specific geographic regions, posing challenges for broader application across different healthcare facilities. Discussion The review findings show that AI-based approaches are revolutionizing the field of medication management, with their great potential to enhance precision, efficiency, and accuracy compared to traditional manual methods6,41,57,80,86,105. Historically, the medication management process has been labor-intensive and prone to human error due to its reliance on manual checks59,106,107. The integration of AI technologies is transforming this landscape by introducing more reliable, consistent, and streamlined workflows, significantly improving the overall effectiveness of medication management and thereby bolstering patient safety. A key application of AI in medication management is through NLP. Given that EHRs are rich in patient data, much of which is traditionally stored in unstructured formats such as physician notes, discharge summaries, and clinical narratives, extracting relevant medication information from these documents is a time-consuming and error-prone task42,54,91. NLP algorithms can transform this aspect of medication management by automating the interpretation and extraction of essential information from unstructured text, such as drug names, dosages, frequencies, administration routes, conditions, and adverse events86. This structured, extracted data can then be seamlessly integrated with other EHR components, enabling the identification of discrepancies, such as mismatches between physician notes and formal prescription orders, and allowing for immediate corrective actions68,80. The automation provided by NLP not only reduces the workload on healthcare providers but also significantly enhances the accuracy and speed of the medication management process56,87. Among various NLP algorithms, Conditional Random Fields (CRFs) have historically been a dominant approach for detecting ADE and medication discrepancies within clinical texts through named entity recognition and relation extraction in the medical domain46,86. Although laying the groundwork for sequence labeling, and continuing to be used as a valuable tool, CRFs have been largely supplanted by or combined with more advanced models, such as self-attention models, that offer greater accuracy and contextual understanding. Recent advancements have shifted focus toward deep learning models, particularly those based on transformer architectures like BERT (Bidirectional Encoder Representations from Transformers)45–47,56,91,108. With deep contextual understanding, these advanced NLP models can help discern that certain medications should not be co-prescribed or recognize potential omissions in a medication list based on a patient’s medical history. They can also handle complex cases involving multiple medications and conditions with greater accuracy, reducing the risk of limited oversight and ensuring that patients receive the correct medications. Beyond NLP-related models, statistical and machine learning approaches such as regression models and tree-based techniques are integral to predictive analytics and decision-making in medication management41,58,59,76. Regression models, traditionally used for predicting continuous outcomes, can also be employed for classification, helping to identify the likelihood of various outcomes. In the context of medication management, regression models can predict the probability of experiencing potential adverse drug events and quantify the impact of various patient factors, such as age, weight, and existing health conditions, allowing for more nuanced and personalized predictions67,92. Tree-based models, including ID3 decision trees, demonstrate particular efficacy in medication management through their hierarchical partitioning of feature spaces. By recursively segmenting data according to optimal predictive thresholds, these models excel at classification tasks essential for therapeutic decision support43,81,109. Their inherent capability to model complex interactions between dependent variables (e.g., treatment outcomes) and multidimensional predictors, including patient demographics, comorbidity profiles, and medication history, provides critical insights into variable importance and interaction dynamics. Similarly, regression models could reveal a higher risk of adverse reactions in a particular patient group, guiding safer prescribing practices76,95. These models are particularly valuable for modeling relationships between a dependent variable and multiple independent variables, providing insights into the strength and nature of these relationships, which are crucial in analyzing large datasets that include patient demographics, comorbidities, and historical medication usage55,57,80. By uncovering patterns and correlations that may not be immediately obvious to clinicians, these models can forecast potential adverse events, flag possible drug interactions, and recommend appropriate medication adjustments. Moreover, both tree-based and regression models are often preferred for their interpretability compared to more complex neural networks110. Their transparent decision-making process allows healthcare providers to understand and trust the rationale behind AI-generated recommendations111. Incorporating these models into the medication management process may enable healthcare providers to anticipate risks, customize treatment plans, and suggest safer alternatives, ultimately improving patient safety and treatment outcomes. As for the selected decision-theoretic papers, most of them are rule-based decision support models, which do not cover common DT techniques, such as mathematical optimization and simulation84,92. Although the literature on the decision-theoretic models in medication management is limited, DT has the potential to significantly impact medication management by providing data-driven, optimized approaches to streamline the process, prioritize high-risk patients for manual review, optimize the allocation of healthcare resources, and reduce the cognitive load on clinicians6,23,105,112. While AI models can predict high-risk scenarios, DT methods could optimize the response by designing efficient workflows to mitigate those risks113. This could lead to faster detection of medication errors, minimizing adverse drug events, and improving patient safety114. However, there are limited real-world deployments and applications of the models reported in the literature on medication management reveals an opportunity to translate theoretical insights into clinical improvements for patient safety. The literature indicates that Clinical Decision Support Systems (CDSS) applied in medication management can primarily be categorized into two areas: risk alerts and drug recommendations. In risk alert applications, typical scenarios include polypharmacy management and error prevention. As for polypharmacy, CDSS tools alert clinicians to potential drug-drug interactions or duplications, thereby reducing side effects and improving patient outcomes72,80,84. Error prevention tools, such as MedAware, focus on identifying outliers in prescriptions based on patient-specific clinical data, preventing medication errors, and enhancing patient safety61,62,64. Besides, AI-powered prescription alert systems are also shown to significantly reduce irrelevant alerts compared to rule-based Computerized Physician Order Entry (CPOE) systems, streamlining the clinical decision-making process61,84. In drug recommendation applications, e-prescribing systems integrate recommendations to reconcile potential discrepancies and flag adverse interactions, especially in long-term care97. Advanced AI-based tools further improve accuracy by offering real-time predictions that can account for allergy information and drug incompatibilities60,87,90. However, integrating these tools into existing clinical workflows remains challenging, limiting broader applicability and clinician acceptance. Concerns around the transparency of AI algorithms and the risk of alert fatigue also complicate their implementation. In addition to literature-based models, we also found some commercial medication management tools that have been integrated into healthcare systems, enhancing patient safety by reducing manual data entry and providing a consistent view of patient medication histories. Tools like DrFirst’s Rcopia 115 and Cureatr’s Medication Management Software116 leverage AI to enhance data usability and alert capabilities. Rcopia uses AI to interpret free-text prescriptions, while Cureatr provides real-time alerts for drug interactions. However, their effectiveness depends on data quality, and complex clinical judgments will still require clinician oversight. Other solutions, such as Avicenna Medical Systems’ MedRec117 and Health Care Systems’ HCS Medication Reconciliation118, focus on data integration and workflow automation. MedRec, used extensively within Veterans Affairs hospitals, consolidates data across sources but lacks predictive capabilities, limiting proactive discrepancy identification. HCS’s tool emphasizes efficiency by automating medication history retrieval using static algorithms, but it may lack flexibility in adapting to unique clinical cases. Tools such as MEDITECH’s Expanse Pharmacy119 and WellSky Medication Management120 emphasize workflow customization and support, though extensive training is often necessary for full functionality. Systems like Better Meds121 and First Databank’s MedKnowledge122 offer comprehensive data integration and drug information, but challenges with disparate technologies can affect their interoperability in some settings. Platforms such as RXNT123 offer extensive functionalities and support for medication management. However, limitations in interoperability with other electronic health systems can hinder their seamless integration and usability in diverse clinical environments. Overall, both the selected literature and commercial tools underscore the importance of technology-enabled solutions in improving medication safety and prescribing efficiency. Despite their benefits, challenges such as EHR interoperability and data quality remain, necessitating ongoing validation to ensure responsiveness to evolving clinical needs. Continuous real-world application is essential to bridge the gap between theoretical models and practical implementation for improving the medication management process. From clinician perspectives, AI-powered tools can automate data extraction from EHRs and reduce administrative tasks for clinicians, allowing them to focus more on clinical decisions42,66,83. Although several of the published studies focus on medication error management, a majority of them concentrate on single drugs or a narrow panel prescribed within a specific specialty124,125. Ideally, medication management should account for all medications a patient may be taking, including over-the-counter (OTC) and also unprescribed substances126. A clinical decision support system that could predict what non-prescribed medications a patient is likely to take would be highly beneficial, which can enhance the comprehensiveness of medication management and reduce the risk of missing important interactions127. Almost 72% of the studies reviewed centered on ADE detection/prediction or medical error detection, which is important but represents only one step in the overall medication management process. A very small number of studies explored the complete process, highlighting a gap in the research86–90. Additionally, the challenge of alert fatigue, a common issue when introducing technological solutions to clinical workflows, was not adequately addressed in most studies128. This concern is consistent with our finding that 80% of methods remain at the experimental stage: many algorithms have been validated retrospectively but not deployed, partly because generating too many non-specific alerts risks rejection by clinicians. The lack of robust strategies to improve alert specificity and integrate decision support for medication review and reconciliation seamlessly into existing workflows has limited translation into real-world practice, as an effective medication management solution must balance comprehensive alerts with minimizing unnecessary interruptions to clinical care. This is a promising area for active research using the increasing availability of highly granular clinical, consumer and administrative data for analysis and decision support. Moreover, this review identifies a gap in the use of DT methodologies that can optimize clinical workflows and reduce the burden on clinicians. Decision-theoretic models remain underutilized23. More research is necessary, not just on developing sophisticated algorithms, but on how they can be seamlessly integrated into everyday clinical workflows, making medication management solutions practical for adoption and daily use129,130. Despite significant advancements in both AI and DT, this review highlights some challenges of their development and integration in the medication management pipeline remain limited. The challenges result from the complex and fragmented nature of healthcare systems. One major barrier to effectively integrating both AI and DT in this management process is data quality131. Incomplete, inaccurate, or biased EHR data can reduce the accuracy of ML models, which may lead to false positives or negatives when identifying medication discrepancies132. This helps to explain why most studies have focused on narrowly scoped tasks such as single-drug ADE detection, which can often be performed using limited data from a single source or a small feature set measured in constrained usage settings, such as the use and monitoring of anticoagulants for ADE prediction133,134. By contrast, tasks such as medication reconciliation and comprehensive medication review require integrating and validating information across multiple provider systems, encounter types, patient characteristics, and data formats. These tasks are therefore far more vulnerable to missing or inconsistent records, hindered by time-consuming and non-standardized management and review processes that complicate data collection, challenged by human and technological factors in documentation and handoffs, and further constrained by underspecified clinical workflows135. As a result, researchers have tended to avoid reconciliation tasks in favor of simpler detection problems, leaving this clinically critical area underexplored. Similarly, with inaccurate input data in the DT models, the optimization programs will yield a suboptimal decision for the real-world system136. This highlights the need for robust data governance and continuous improvements in healthcare data practices to ensure that AI and DT can work together effectively. Besides, integrating AI and DT systems into the existing EHR workflows faces transparency and interoperability issues. These interoperability barriers are especially acute in primary care and community care settings, where patient records are dispersed across multiple outpatient providers, specialists, pharmacies, and fragmented health IT systems, which makes robust data collection and analysis difficult137. The challenge is exacerbated by the diversity of health conditions treated in primary care settings, with corresponding medication management challenges, as well as the variety of payment models for reimbursement of medication review and management. This structural fragmentation and complexity contribute to why only 9% of the selected studies were conducted in primary care settings, compared with the relatively improved consistency of single-site inpatient systems. Besides, the “black-box” nature of many AI models raises concerns about reliability, which is critical for clinical decision-making138. The lack of clear model explanations can reduce trust in these systems, reinforcing the need for explainable AI methods, alongside DT-driven optimization, to enhance their clinical utility. There are also limitations to our review itself. Although we searched for articles across 5 large databases, we only considered the published literature and, therefore, may have failed to capture unpublished studies that might also contain relevant content. As considerable studies on patient safety-focused medication management and AI/DT algorithms are disseminated through conferences, posters, and other non-traditional means, which are often not indexed in standard bibliographic databases, our focus on bibliographic databases may have introduced selection bias by excluding potentially valuable insights found in these alternative sources. Furthermore, the majority of the identified publications evaluated algorithms at the single-drug or chemical level and were consequently excluded from our review. While we believe that this exclusion does not fundamentally alter the overall content or methodologies of the medication management efforts examined, it may impact the perceived comprehensiveness of our findings. This filtering process, while emphasizing studies involving multiple medications and symptoms, may inadvertently overlook valuable insights specific to single-drug evaluations, which could contribute to a more nuanced understanding of AI and DT applications in medication management. Besides, the executed search strategy required iteration due to the evolving terminology in AI/DT-driven biomedicine research, necessitating the inclusion of new terms and definitions over time. From a research perspective, future research should also delve deeper into advanced AI-DT paradigms. While our review primarily mapped traditional machine learning approaches, including tree-based, regression-based, NLP, and deep learning, emerging areas such as generative AI and LLMs were underrepresented, largely due to the limited number of studies available at the time. As LLMs continue to evolve, they offer a promising avenue for exploration, particularly in healthcare, where their natural language reasoning and ability to integrate structured and unstructured medication data can not only enhance predictive modeling but also provide complementary evaluation to support clinical decision-making. Importantly, DT methods such as mathematical programming, rule-based modeling, and simulation can be paired with these advanced AI models to move beyond prediction toward prescriptive decision-making in medication management. Pursuing these integrative approaches is essential for translating advances in AI into practical, trustworthy tools that address patient safety in routine care. In conclusion, this scoping review advances the existing body of literature by systematically reviewing studies that have applied analytical methods, especially AI and DT approaches, to improve medication management, and by evaluating their current level of adoption in healthcare practice. However, despite notable progress, we notice that research integrating AI and DT within a unified framework remains scarce, and few studies have progressed beyond conceptual frameworks to deploy these analytical approaches in clinical settings. This gap underscores the need for future research that not only explores the joint application of AI and DT but also strengthens their integration into everyday healthcare workflows. Continued interdisciplinary collaboration and empirical validation are essential to address these challenges and fully realize the potential of these technologies in enhancing patient safety and healthcare quality. Methods We conducted the scoping review following the guideline Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews (PRISMA-ScR). The PRISMA-ScR checklist is shown in Supporting Information 1. The published literature was searched using strategies created with the help of a university librarian for studies that utilized AI or DT in the medication management context for patient safety. We used the Sysrev platform to conduct the screening process of the papers139. Databases and search strategy We searched PubMed, Web of Science, ACM, INFORMS, IEEE, and CINAHL databases through December 2023. These databases were selected because of their comprehensive coverage of relevant academic disciplines, including medicine, healthcare, computer science, decision analytics, and operations research. To ensure the inclusiveness and breadth of our search, we designed the search query to be as broad as possible. We adopted concepts related to medication management, such as medication error, medication discrepancies, medication reconciliation, adverse drug events, polypharmacy, etc., to cover all pertinent aspects. To find all papers that incorporate AI or DT in patient safety-focused medication management, we carefully established a combination of terminologies to search in the paper title and abstract. The search strategy was as follows: (ANY medication-management-related terms in the title or abstract) AND ((ANY AI-related terms in the title or abstract) OR (ANY DT-related terms in the title or abstract)), where Medication-management-related terms include: ‘medication error*’, ‘medication discrepanc*’, ‘medication reconciliation*’, ‘medication accuracy’, ‘adverse drug event*’, ‘polypharmacy’, ‘pharmacovigilance’ Medication-management-related terms include: ‘medication error*’, ‘medication discrepanc*’, ‘medication reconciliation*’, ‘medication accuracy’, ‘adverse drug event*’, ‘polypharmacy’, ‘pharmacovigilance’ AI-related terms include: ‘artificial intelligence’, ‘machine learning’,’deep learning’, ‘reinforcement learning’, ‘neural network*’, ‘encoder*’, ‘decoder*’, ‘autoencoder*’, ‘generative adversarial network*’, ‘bayes*’, ‘decision tree*’, ‘classifier*’, ‘xgboost’, ‘random forest*’, ‘support vector machine’, ‘bagging’, ‘boosting’, ‘natural language processing’, ‘NLP’, ‘large language model*’, ‘LLM’, ‘generative AI’, ‘GPT’ AI-related terms include: ‘artificial intelligence’, ‘machine learning’,’deep learning’, ‘reinforcement learning’, ‘neural network*’, ‘encoder*’, ‘decoder*’, ‘autoencoder*’, ‘generative adversarial network*’, ‘bayes*’, ‘decision tree*’, ‘classifier*’, ‘xgboost’, ‘random forest*’, ‘support vector machine’, ‘bagging’, ‘boosting’, ‘natural language processing’, ‘NLP’, ‘large language model*’, ‘LLM’, ‘generative AI’, ‘GPT’ DT-related terms include: ‘operation* research’, ‘operation* management’,’integer programming’, ‘linear programming’, ‘mathematical programming’, ‘stochastic programming’, ‘stochastic optimization’, ‘stochastic process’, ‘robust optimization’, ‘Markov’, ‘heuristic*’, ‘metaheuristic*’ DT-related terms include: ‘operation* research’, ‘operation* management’,’integer programming’, ‘linear programming’, ‘mathematical programming’, ‘stochastic programming’, ‘stochastic optimization’, ‘stochastic process’, ‘robust optimization’, ‘Markov’, ‘heuristic*’, ‘metaheuristic*’ A total of 2000 records were identified by this search strategy. Duplicate records were identified using the automatic deduplicator in SR-Accelerator online review software140. We removed 990 duplicates which resulted in 1100 unique studies for further screening. Selection of studies and eligibility criteria To be eligible for inclusion, published research had to meet all of the following 6 criteria: Peer-reviewed original research with full text published in English. Peer-reviewed original research with full text published in English. Integrated AI or DT analytical methods into medication management-related tasks for patient safety purposes. Integrated AI or DT analytical methods into medication management-related tasks for patient safety purposes. Examined the impact or performance of integrating AI or DT into medication management-related tasks. Examined the impact or performance of integrating AI or DT into medication management-related tasks. Quantitative analysis. Quantitative analysis. Not limited to the biochemical level of drug reactions. Not limited to the biochemical level of drug reactions. Involved multiple conditions and diseases and not focused on a single medication. Involved multiple conditions and diseases and not focused on a single medication. Based on these criteria, the articles underwent two exclusion steps. Initially, a review of titles and abstracts led to the exclusion of 990 articles. The title and abstract screening excluded systematic reviews/meta-analyses, opinions, commentaries, perspectives, vision articles, guidelines, protocol articles, qualitative studies, exploratory and conceptual studies, and non-English language studies. This was followed by a review of 110 full-text articles, resulting in the exclusion of 46 articles. During the full-text screening, we excluded conference posters. We further checked the relevance of the paper to our topic and only kept studies that focused on leveraging AI/DT in patient safety-focused medication management-related problems. During both exclusion steps, two researchers independently labeled the articles. One researcher was a doctoral student majoring in information systems with extensive training in machine learning, optimization, and healthcare, while the other was a graduate student in bioinformatics with a medical degree and clinical practice experience. We used the Sysrev online literature review platform to independently label and record the decisions139. Conflicts in labeling were resolved by an experienced researcher specializing in healthcare informatics, analytics and operations. This rigorous process resulted in the inclusion of 64 scientific articles published in the literature. Fig. 6 shows the search and screening process of this scoping review. The details of the selected studies are shown in Supporting Information 2. Flowchart of literature search and screening process. Data collection and analysis We collected the 64 articles and extracted different data elements from each, such as year of publication, country of origin, study target, data source, facility and specialty of the data source, data collection start and end time, duration of the data sample, sample size, features included in the data with the total number and feature categories, applied algorithms and their types, outcome measures, outcome performance, and the status of the proposed method. Due to the differences in the included studies and the lack of standardized reporting, we did not conduct a meta-analysis. Instead, we used narrative synthesis to summarize our findings descriptively. To facilitate this process, we utilized ChatGPT141 to extract and summarize relevant information, followed by a meticulous manual review and refinement to ensure accuracy of our analysis. Supplementary information Supplementary Information Acknowledgements This work was supported in part by a research grant (to XY and AR) from Carnegie Mellon University's partnership with the Jewish Healthcare Foundation for the Initiative for Patient Safety Research (IPSR) and the Center for Innovation in Health. We thank Sarah Young at Carnegie Mellon University Libraries for assistance with the literature search resources. Author contributions R.P. conceptualized the idea for this study. X.Y. and R.P. designed it and framed reference search strategies. X.Y searched the databases and conducted record deduplication. X.Y. and A.R. performed screening, study selection, data extraction, and data analysis, with extensive input and feedback from R.P. X.Y. contributed to the original draft, with input on the clinical perspectives from A.R. R.P. performed review and editing. R.P. supervised the study. All authors approved the final draft for submission. Data availability Data is provided within the manuscript or supplementary information files. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Supplementary information The online version contains supplementary material available at 10.1038/s44401-025-00052-1. References Associated Data Supplementary Materials Supplementary Information Data Availability Statement Data is provided within the manuscript or supplementary information files.