Artificial intelligence in clinical pharmacy—A systematic review of current scenario and future perspectives

Abstract

Objective

Medication prescription errors represent a significant and persistent challenge within healthcare systems globally, constituting a primary focus for clinical pharmacy practice. Additional complexities involve the optimization of drug dosing and the implementation of personalized medicine. This review aims to synthesize the current advancements in artificial intelligence (AI) applications within clinical pharmacy and to discuss future directions for the field.

Methods

To present this narration, 30 articles were reviewed in total. The literature search was done using electronic databases, for example, PubMed, Medline, and Google Scholar, with the help of some keywords. Only articles published in peer-reviewed journals were included.

Results

A total of 30 articles that demonstrated the utility of AI-based applications in clinical pharmacy were included for further analysis. Across all included studies, AI was utilized primarily for the detection of adverse drug events, clinical decision support, verification of prescription accuracy, and pharmacometrics. Secondary applications included providing recommendations to pharmacists for medication therapy management and, importantly, predicting the therapeutic response to a given treatment in conjunction with its cost-effectiveness.

Conclusion

Artificial intelligence-based algorithms have been identified as applicable tools for the early detection of adverse drug events and prescription errors, the prediction of individual drug response, and the design of patient-specific treatment plans. Prior to broad clinical implementation, future multicenter, prospective studies employing standardized clinical endpoints, external validation, and cost-effectiveness analyses are required.

Introduction

John McCarthy, an American computer scientist renowned as the father of artificial intelligence (AI), first introduced the term “Artificial Intelligence” during a conference at Dartmouth College, New Hampshire, United States, in 1956. ¹ This landmark event is widely regarded as the foundational moment of AI as a formal discipline. Subsequently, the field advanced at an accelerated pace. In its initial stages, scientists and researchers employed AI primarily for mathematical computations, solving algebraic problems, developing algorithms, and constructing logic-based systems to address challenges traditionally requiring human cognitive capabilities. ²

A significant innovation occurred in 2006 when Geoffrey Hinton and his collaborators proposed the development of neural networks designed to preserve information and prevent the vanishing gradient problem during data training. ³ Among the various algorithms applied in AI, deep learning (DL) techniques have demonstrated exceptional utility. These methods utilize multiple layers of artificial neural networks to automatically generate abstract representations and derive insights from input data, thereby minimizing the need for human intervention. ⁴

Recent advancements in technology have enabled the integration of multiple features alongside vast quantities of data, which can now be effectively utilized by AI systems. ² Contemporary DL algorithms are capable of remarkably emulating human cognitive abilities through the replication and abstraction of complex datasets. ³ Initially developed as a tool for solving mathematical problems and executing complex algorithms, AI has evolved into a system of advanced intelligence capable of surpassing human problem-solving efficiency in certain domains.. ¹

Clinical pharmacy is a specialized field dedicated to the optimization of medication use to promote health, wellness, and disease prevention. ¹ It represents a pharmacist-driven approach to patient care. A central responsibility of the clinical pharmacist is to advance safe, effective, and economically efficient drug therapy to enhance patient outcomes. This goal of optimal patient care is achieved through a range of services, including the design of pharmaceutical care plans, resolution of drug therapy problems, participation in interdisciplinary rounds and consultations, patient education during hospitalization and after discharge, as well as thorough documentation and data management. ²

The evolutionary advancement in the application of AI within healthcare has enabled the utilization of vast, accumulated datasets, presenting significant opportunities to enhance patient care across all domains of clinical practice including pharmacy and allied health disciplines. In response, clinical pharmacy has begun to integrate these innovative technologies to address and advance its core objectives: improving patient access to care, reducing healthcare costs, and optimizing clinician satisfaction. ³

In recent years, numerous studies have investigated the utilization of AI in clinical pharmacy. ChatGPT-based systems have shown to be successful in detecting drug–drug interactions (DDIs) and recommending alternate therapies in medication management. ⁵ Artificial intelligence-driven clinical decision support systems have been utilized to minimize prescribing errors and enhance dosage for patients with intricate medication regimens. ⁶ Machine learning algorithms have been employed to identify potentially inappropriate medications (PIMs) in older populations, enhancing prescription accuracy and safety. ⁷ Moreover, robotic dispensing systems are currently being implemented in hospital pharmacies to enhance workflow efficiency, minimize human error, and elevate satisfaction among patients. ⁸

Artificial intelligence is a broad term used for the computational technique which is capable of performing the tasks which otherwise need human brain cognition such as learning, understanding, language, planning, and problem-solving. ⁴

Although the current surge in AI applications within healthcare services and research is a relatively recent phenomenon, its origins trace back to the 1980s with the development and commercialization of speech recognition software, which relied on statistical predictive modeling. ⁹ During this same period, the automation of hospital pharmacies began, primarily to support drug distribution, manage inventory, and generate financial reports. By the late 2000s, applications of informatics in clinical pharmacy started to emerge, contributing to the formal definition of the pharmacist informaticist role. ¹⁰ However, limited datasets and technical resources hindered progress, resulting in a prolonged period of stagnation in the practical application of AI in healthcare. ¹¹

Now, in last decade, the recent advances in data processing techniques and availability of huge electronic healthcare records have created this resurgence of association of AI with medicine. ¹² The applications of AI in healthcare can be observed in the fields of risk assessments, prediction of hospital readmissions, clinical decision-making, and individual patient's data management. These all applications are supported by the algorithms generated through the process of machine learning that further based on the techniques of Deep Neural Networking (DNN), Generative Adversarial Networks, Reinforcement learning, and Knowledge graphs. ¹³

The integration of these applications with medicine has progressed rapidly and is evident across three distinct levels within the healthcare industry:

1. at the clinician level, primarily through the rapid and accurate interpretation of complex information,

2. at the patient level, by empowering individuals to process and manage their personal health data, and

3. at the health system level, by optimizing workforce capabilities and available resources to minimize medical errors and enhance the efficiency of patient care management ¹⁴ Saudi Arabia's Vision 2030 seeks to modernize healthcare via digital transformation, with AI designated as a critical instrument for enhancing care delivery and efficiency. Although AI projects have been extensively used in diagnostics and telemedicine, their incorporation into clinical pharmacy is currently developing. However, extensive deployment throughout the Kingdom is still restricted highlighting the necessity for additional research and policy assistance aimed at integrating AI into clinical pharmacy practice. ¹⁵

The technical systems associated with AI are already providing good assistance in routine, manual, and repetitive tasks of medical industry. In addition to that, many researchers have been putting their efforts in observing the effects of implementation of different AI or machine learning models on various aspects of clinical pharmacy. ¹³ For example, the automation of drug dispensing and pharmacy administration has been started at various hospital settings. In Jeddah city, Saudi Arabia, Ahmed et al. evaluated the drug dispensing and administration process in central pharmacies across the city. They found 28.6% hospitals were using automated dispensing cabinets. ¹⁶ Recently, in an ongoing study, in Saudi Arabia, the efficacy of robotic pharmacy solution was studied. In a 21-month long study, the investigators found 22% increase in patients’ satisfaction regarding the pharmacy services; pharmacist productivity was increased by 33% with zero observed dispensing errors. ¹⁷

Data, information, and knowledge management; optimal dosing; identification of over- or underprescribed medications; and prediction of drug interaction effects represent key competencies within clinical pharmacy to which AI methods have been applied. Accordingly, this article aims to review published research on the implications of AI in the field of clinical pharmacy. The broader objective is to contribute knowledge that clinical pharmacists can apply to advance both research initiatives and patient care delivery.

Material and methods

Search strategy

A systematic review was conducted in accordance with the PRISMA 2020 guidelines to identify relevant studies on the application of AI in clinical pharmacy. A comprehensive search was performed across seven databases: ScienceDirect, PubMed, Web of Science, Scopus, Directory of Open Access Journals (DOAJ), ProQuest, and Google Scholar.

To ensure specificity and reproducibility, we developed detailed search strings using Boolean operators (AND, OR) and truncation where applicable. Examples of the search strings used are as follows:

• PubMed: (“Artificial Intelligence"[MeSH Terms] OR “AI” OR “Machine Learning” OR “Deep Learning”) AND (“Clinical Pharmacy” OR “Pharmaceutical Care” OR “Pharmacy Practice”) AND (“Current Scenario” OR “Future Perspectives”)

• Scopus/Web of Science: TITLE-ABS-KEY(“Artificial Intelligence” OR “AI” OR “Machine Learning” OR “Deep Learning”) AND TITLE-ABS-KEY(“Clinical Pharmacy” OR “Pharmacy Practice” OR “Pharmaceutical Care”) AND TITLE-ABS-KEY(“Current Scenario” OR “Future Perspectives”)

• Google Scholar/DOAJ (broad exploratory): “Artificial Intelligence” AND “Clinical Pharmacy” AND (“Current Scenario” OR “Future Perspectives”)

Filters were applied to limit results to peer-reviewed articles published in English. Search strategies were tailored to the syntax of each database but were conceptually aligned.

Inclusion and exclusion criteria

Studies were included if they met the following criteria:

1. Addressed the application of AI in clinical pharmacy.

2. Discussed the current scenarios or future perspectives.

3. Were original research articles, reviews, or case studies published in peer-reviewed journals.

Exclusion criteria included

1. Studies not related to clinical pharmacy.

2. Articles lacking full text or not in English.

3. Conference proceedings and case studies not directly relevant to the objectives of the review.

Study selection

A total of 1875 records were retrieved. After removal of duplicates (n = 592) and renounced records (n = 23), 1260 records were screened based on titles and abstracts. Of these, 1098 were excluded for not meeting inclusion criteria.

Two reviewers independently screened the titles, abstracts, and full texts of all retrieved records. Disagreements regarding eligibility were first resolved through discussion, and when consensus could not be achieved, a third reviewer was consulted to make the final decision. The selection process followed the PRISMA 2020 guidelines, and the flow of studies is illustrated in Figure 1.

Figure 1.

PRISMA flowchart of the systematic review.

Source: Page MJ, McKenzie JE, Bossuyt PM, Boutron I, Hoffmann TC, Mulrow CD, et al. The PRISMA 2020 statement: An updated guideline for reporting systematic reviews. BMJ. 2021;372:n71. doi: 10.1136/bmj.n71.

Retrieval and eligibility assessment

Full-text retrieval was attempted for 162 studies, which included 47 conference proceedings and 29 case studies. Of these, 76 were unavailable. The remaining 86 studies underwent full-text screening for eligibility, resulting in the exclusion of 56 studies. Exclusions were due to not meeting the inclusion criteria (n = 37) or continued unavailability of the full text (n = 19).

Quality assessment

In accordance with PRISMA guidelines, we assessed the methodological quality of the included studies to evaluate the strength of the evidence and potential for bias. Given the heterogeneity in study designs, the Mixed Methods Appraisal Tool (MMAT, version 2018) was employed for this purpose. The MMAT evaluation consists of two parts: an initial screening section and a subsequent appraisal with criteria specific to five categories of study designs. All items were scored as “Yes,” “No,” or “Can’t tell.” Consistent with the tool's guidelines, studies that received a “No” on both screening questions or a “Can’t tell” on one or more were deemed potentially unsuitable for full MMAT appraisal. A “Yes” rating signifies that the study met the methodological standard for that item, while “Can’t tell” indicates insufficient reporting to make a judgment. Two reviewers conducted the assessments independently; any discrepancies were resolved through consensus or by consulting a third reviewer. In keeping with MMAT guidance, we present the individual item-level assessments and have refrained from calculating an overall summary score.

Final selection

Following systematic screening and eligibility assessment, 30 studies were included in the final review for detailed analysis. Two reviewers independently screened the titles and abstracts of all identified records for relevance and subsequently assessed the full-text articles against the predefined inclusion and exclusion criteria. Discrepancies were resolved through discussion, and when consensus was not reached, a third reviewer adjudicated the decision. The included studies offered important insights into the application of AI in clinical pharmacy, highlighting both current practices and future directions.

Registration and protocol

Registration Information: This review was registered in the PROSPERO International Prospective Register of Systematic Reviews under registration number CRD420251058239.

Available from: https://www.crd.york.ac.uk/PROSPERO/view/CRD420251058239

Protocol Access: A full protocol was submitted to PROSPERO at the time of registration. The protocol will be made publicly available after the completion of the review as per PROSPERO policy.

Amendments to Protocol: No amendments have been made to the registered protocol since its initial submission.

Results

The PRISMA diagram (Figure 1) illustrates the process of study identification, screening, and inclusion. In total, 1875 records were retrieved from various databases. Following the removal of 592 duplicates and 23 renounced records, 1260 records remained for screening. Of these, 1098 were excluded. Full texts of 162 articles were sought, but 76 could not be retrieved (47 conference proceedings and 29 case studies). After assessing 86 articles for eligibility, 56 were excluded (37 not meeting the inclusion criteria and 19 without accessible full text). Ultimately, 30 studies were included in the final review.

Characteristics of the study

Of the 30 studies included in the analysis, the majority were conducted in the United States (n = 12), followed by China (n = 3). Contributions were also made by researchers from Israel, France, the Netherlands, Japan, Portugal, Nigeria, Saudi Arabia, Switzerland, Russia, the United Kingdom, and Australia. As summarized in Table 1, the reviewed articles encompassed a range of methodological approaches, including observational, exploratory, experimental, and retrospective designs, reflecting the methodological diversity within this field of research. The analysis synthesized key details from each study, including authorship, geographical context, research design, study population and sample size, research objectives, forms of AI employed, principal findings, conclusions, and recommendations.

Table 1.

Study characteristics of the included studies.

Sr. No.	Study author and year	Country	Study design	Study population and sample size (N)	Study objective for using AI	Type of AI used in the study	Result and conclusion	Recommendation
1	Alon Bartal et al., 2024 18	Data from Israel, USA, and others	Exploratory data analysis-based study	The study uses data from social media, clinical research, manufacturers’ reports, and ChatGPT. It does not use a conventional clinical population or N.	To disclose adverse side effects (ASEs) of GLP-1 receptor agonists using social media and AI-based analytics.	Named Entity Recognition (NER) model leveraging NLP with tools like ScispaCy and integration with ChatGPT data.	Identified 381 ASEs related to GLP-1 receptor agonists, with 21 novel ASEs solely from social media. Constructed an ASE network revealing frequently co-occurring ASEs. Results indicate the possibility of improving medication safety procedures and pharmacovigilance.	Use of social media analytics and artificial intelligence to promote proactive pharmacovigilance. To lower the risk of ASE, continuous medication safety monitoring should be encouraged, treatment protocols should be enhanced, and individualized approaches should be looked into.
2	Don Roosan et al., 2024 19	USA	Qualitative Study	45 clinical patient cases categorized into simple (15), complex (15), and very complex (15).	To evaluate the effectiveness of ChatGPT in performing MTM activities including identifying DDIs, recommending alternatives, and devising management plans.	ChatGPT 4.0, an AI-powered big language model, is used for MTM services such DDI analysis, alternative therapy recommendations, and management plan development.. For the purpose of evaluating accuracy, ChatGPT replies were compared with pharmacist-validated responses.	All 39 patient cases were solved by ChatGPT with 100% accuracy, recognizing interactions and providing suggestions and care strategies. Limitations, however, were the use of follow-up questions and the absence of precise dose recommendations.	Improve ChatGPT's ability to integrate with clinical decision support systems (like EHRs) and give users access to extensive reference materials. Educate pharmacy staff about AI-powered solutions that can optimize MTM and enhance patient care.
3	Euibeom Shin et al., 2024 20	USA	Quantitative Descriptive	Simulated situations in a community pharmacy, such as obtaining medication information, interpreting prescriptions, making decisions, and consulting.	To evaluate ChatGPT's effectiveness in a variety of community pharmacy tasks, including retrieving drug information, seeing labeling mistakes, and making multidisciplinary suggestions	Prescription label error detection, retrieving pharmacological information (such as Rituximab, Warfarin, and St John's wort), and multidisciplinary scenarios combining drug management and health guidance were among the simulated tasks used to test ChatGPT-4. The accuracy, concordance, and integration of the outputs with medical guidelines were assessed using reference databases (e.g., DrugBank, PubMed) and tools such as Grammarly.	ChatGPT demonstrated potential for improving pharmacist decision-making, effectively detected labeling problems, and offered practical medication and health advice. However, specific and regional clinical guidelines were found to have shortcomings.	To improve geographic accuracy, create LLMs customized to individual pharmacies. Integrate ChatGPT with real-time pharmacy management systems to increase its functionality and ensure ethical data handling. Carry out extensive validation research.
4	Merel van Nuland et al., 2024 21	Netherlands	Quantitative Descriptive	264 multiple-choice questions in clinical pharmacy domains, contrasted against responses from 458 pharmacists	To compare ChatGPT's performance to that of pharmacists in responding to clinical pharmacy factual knowledge questions.	The accuracy, concordance, validation quality, and repeatability of ChatGPT-4 were evaluated. The questions covered seventeen clinical pharmacy categories. The metrics included correctness of responses, consistency between responses and explanations, and reproducibility across iterations. Using the same set of questions, performance was compared to pharmacist results from 2022.	ChatGPT outperformed pharmacists (66%), achieving 79% accuracy. 72% of the substantiations had an outstanding rating, while the concordance rate was 95%. Reproducibility within and between days was around 90%. One limitation was managing region-specific guidelines.	ChatGPT's usefulness can be improved by adding regional guidelines and clinical scenarios to training datasets. For patient safety, use ChatGPT as an additional tool in pharmacy practice under human supervision. Improve its capacity to respond to intricate queries.
5	Qiaozhi Hu et al., 2024 7	China	Retrospective Study	Beers criteria were used to examine 18,338 dementia-stricken older persons from 75 hospitals spread across eight major Chinese cities.	To use multilabel classification to create the best machine-learning model for detecting potentially inappropriate medications (PIMs) in dementia patients.	Combined six algorithms (CatBoost, XGBoost, RF, etc.) with three multilabel classification techniques (binary relevance, label powerset, and classifier chains). Accuracy (97.93%), precision (95.39%), recall (94.07%), F1 score (95.69%), and Hamming loss (0.0011) were all highest for the CC + CatBoost model. A sizable dataset was used to test the robustness and operational effectiveness of the model metrics.	The CC + CatBoost model successfully recognized 36 PIM types, with the most common alerts being for antipsychotics and antidepressants. For PIM detection, the model performed noticeably faster and more accurately than conventional CDSSs.	For real-time PIM detection, combine machine learning-based models, such as CC + CatBoost, with electronic medical records. To enhance the generalizability of the model, broaden the scope of data collecting to cover more regions and conditions. Train medical professionals in the use of AI in dementia care.
6	Ramya Padmavathy Radha Krishnan et al., 2024 22	Australia	Observational Study	120 hospitalized patients, with data on 11,698 drug pairs generated from medication orders.	To evaluate ChatGPT-3.5's performance against that of professional pharmacists in predicting drug–drug interactions (DDIs) using actual patient data.	Three distinct prompt styles were used to test ChatGPT-3.5: (1) generic wording, (2) the addition of “drug interaction,” and (3) specific references to drugs without dosages. The gold standard was a pharmacist's evaluation. Cohen's kappa for inter-rater agreement, sensitivity, specificity, and ROC-AUC were among the metrics.	ChatGPT-3.5 showed poor agreement with pharmacists (Cohen's kappa: 0.077–0.143) and low sensitivity (0.24). Although ChatGPT had a high specificity (>95%), it had trouble correctly identifying clinically significant DDIs in practical situations.	Enhance ChatGPT by integration with decision support systems and clinical dataset training. Improve prompts and algorithms to improve contextual comprehension of actual clinical situations. Use iterative testing to increase its reliability.
7	Sara Grossman et al., 2024 23	USA	Cross-Sectional Study	39 questions about medications sent to a university drug information service.	To evaluate ChatGPT's performance in comparison to conventional sources in terms of providing precise, thorough, and adequate responses to questions about medications.	Investigator-prepared answers based on scholarly literature were compared with ChatGPT (GPT-3.5) responses to drug-related questions. Accuracy, completeness, relevance, and reference-quality were among the metrics. Questions covered a wide range of subjects, including legal issues, dosage, ADRs, compounding/formulation, and therapies.	Just 26% of queries received satisfactory answers from ChatGPT. The primary flaws were inadequate information (41%), inaccuracy (38%), and a lack of direct responses (38%). References were frequently false or incorrect. The study brought up concerns about using ChatGPT to obtain pharmaceutical information.	Improve ChatGPT's training using domain-specific datasets and put in place measures to ensure reliable and accurate references. Users need to confirm ChatGPT's answers with reliable sources. Create guidelines specifically for AI applications in the healthcare industry.
8	Xiaoru Huang et al., 2024 24	China	Comparative Study	30 questions covering clinical pharmacy areas: prescription review, ADR recognition, ADR causality assessment, drug counseling, and patient education.	To assess ChatGPT's precision and comprehensiveness of responses in clinical pharmacy tasks by comparing it with qualified clinical pharmacists.	Structured tasks on prescription review, patient medication education, ADR recognition, ADR cause evaluation, and drug counseling were used to evaluate ChatGPT (version March 23). Five clinical pharmacists evaluated the performance using a 10-point scale and quality rating system. For inter-rater reliability, statistical analysis comprised intraclass correlation coefficients and paired t-tests.	While ChatGPT did well in drug counseling (8.77/10 vs. 9.50/10, p = 0.0791), it performed significantly lower in tasks involving ADRs and prescription reviews (5.23/10 vs. 9.90/10, p = 0.0089). Managing complicated instances and reasoning were among the limitations.	Increase the number of clinical data and patient situations in ChatGPT's training datasets. Use AI technologies in clinical pharmacy environments to support pharmacists’ work, particularly when it comes to repetitive and data-intensive activities.
9	Zhengliang Liu et al. 2023 25	Data from multiple countries	Simulation-based exploratory study design	It does not use a conventional research population or sample size (N) because it doesn't depend on clinical trials or human subjects. Rather, it uses actual ICU data for simulated clinical scenarios to assess AI's skills (more especially, GPT models like ChatGPT and GPT-4).	To assess whether ChatGPT and GPT-4, two examples of large language models (LLMs), can accurately simulate the duties and responsibilities of clinical pharmacists	The AI Pharmacist made use of ChatGPT and GPT-4, two OpenAI-developed large language models (LLMs).	ChatGPT and GPT-4 replicated key clinical pharmacist functions, such as evaluating ICU patient data, creating individualized drug regimens, and predicting patient outcomes. The AI models showed great promise in anticipating adverse drug effects, improving decision-making, and enhancing medication management.	The study recommends improving AI models to increase their accuracy, encouraging human–AI cooperation in clinical decision-making, updating models frequently with fresh clinical data, ensuring ethical and legal supervision, and growing AI applications in pharmacy, including drug discovery and medication adherence.
10	Wu Xingwei et al., 2022 26	China	Quantitative Descriptive	404 elderly patients with cardiovascular disease hospitalized between 3 and 60 days at a single medical center.	To establish a machine learning-based platform that can forecast potential prescribing omissions (PPO), potentially inappropriate medications (PIM), and potentially inappropriate prescriptions (PIP).	Various feature screening and sampling techniques were part of the data preprocessing. 270 prediction models were produced using 18 machine learning methods (such as logistic regression, random forest, and XGBoost). For feature importance, SHapley Additive exPlanation (SHAP) was used. F1 score, recall, precision, and AUC were among the predictive platform measures.	The prediction platform achieved AUCs of 0.8341 (PIP), 0.7007 (PPO), and 0.7061 (PIM). The study highlighted angina, length of hospital stay, and quantity of drugs as important indicators.	Create prediction tools for broader clinical use to reduce PIP in senior citizens. Use multicenter and larger datasets to validate the model. Improve the accuracy of the data and incorporate predictive platforms into EHR systems.
11	Tomoki Takase et al., 2022 27	Japan	Experimental Study	768-bed hospital, dispensing data from 158,548 prescriptions (preimplementation) and 114,111 prescriptions (postimplementation).	To evaluate robotic dispensing systems’ effectiveness and safety in lowering dispensing errors and enhancing pharmacist workflow.	The integrated robotic dispensing system consists of a bar-coded pharmaceutical support system (Hp-PORIMS®) and automated dispensing robots (Drug Station® and Mini DimeRo®). Collectively, pharmacists and pharmacy support personnel used robotic assistance to prepare and dispense drugs. Three study phases: preimplementation, early postimplementation, and collaborative phase—were used to evaluate dispensing times and mistake rates.	A large reduction in pharmacist workload allowed for a greater focus on clinical care, while the robotic dispensing system reduced overall dispensing errors by almost 80%, with unprevented errors dropping to almost zero.	Encourage hospital pharmacies to implement integrated robotic dispensing systems. For optimal safety and effectiveness, make sure pharmacy support staff are properly trained and maximize robotic features (such as error prevention systems).
12	Attayeb Mohsen et al., 2021 28	Japan	Quantitative Descriptive	14 ADR prediction models were created using adverse event reports (FAERS) and gene expression profiles (Open TG–GATEs).	To create a deep learning system that combines gene expression profiles with adverse drug reaction (ADR) reports in order to predict ADRs.	Gene expression data from Open TG-GATEs and ADR incidence data from FAERS were subjected to deep neural networks. To improve accuracy, the models made use of feature selection, hyperparameter adjustment, and data preprocessing. ADRs such hepatitis fulminant and duodenal ulcers were the focus of predictive models.	The models performed well for ADR prediction, with a mean validation accuracy of 89.4%. The molecular mechanisms behind ADRs have been found via pathway enrichment analysis.	Increase the number of integrated datasets used in drug discovery for ADR prediction. For wider applications, improve data quality, lower noise, and expand models to different ADRs. Encourage cooperation between pharmacologists and AI researchers.
13	Hugo Lopes et al., 2021 29	Portugal	Consensus-Based Study	To determine KPIs pertinent to hospital pharmacy practice, a panel of eight experts—hospital pharmacists—participated in five evaluation rounds.	To establish quantifiable key performance indicators for clinical pharmacy (cpKPIs) and support activity (saKPIs) in Portuguese hospital pharmacies.	The relevance and measurability of indicators were identified and evaluated using a combination of nominal and focus group methodologies. The technique included iterative rounds of expert discussions, relevance/measurability rating, and consensus-based indicator refinement. The final list included 85 KPIs in six categories that matched the statements of the European Association of Hospital Pharmacists (EAHP).	The establishment of a final list of 85 KPIs (40 saKPIs, 39 cpKPIs, and 6 general) allowed for performance evaluation and benchmarking in Portuguese hospital pharmacies. This study established a framework for consistent quality assessment.	Implement the established KPIs nationwide in hospital pharmacies in Portugal. Create an international benchmarking system to compare KPIs and promote ongoing quality improvement in hospital pharmacy practice worldwide.
14	Hisham Momattin et al., 2021 17	Saudi Arabia	Observational Study	Analysis of 10-month outcomes using 21-month pre- and postimplementation data of an automated pharmacy system in a hospital context.	To assess the deployment and results of a robotic pharmacy system with an emphasis on dispensing accuracy, wait times, pharmacist productivity, and patient happiness.	Automated robotic pharmacy system integrated with computerized provider order entry (CPOE) and health information systems (HIS). Wait times, error rates, pharmacist productivity, and the satisfaction of patients were among the metrics. Analysis of the FTE workload, space utilization, and dispensing speed was conducted.	The robotic system boosted patient satisfaction by 93%, cut wait times by 53%, raised pharmacist productivity by 33%, and eliminated dispensing errors. ROI was attained in three and a half years.	Increase the use of robotic pharmacy systems while emphasizing their smooth integration with the hospital's current infrastructure. Ensure that staff receive thorough training and look into additional KPIs like long-term drug adherence and patient education results.
15	Jonathan Salcedo et al., 2021 30	USA	Modeling Study	43 patients using AiCure and 71 patients using DOT for active TB treatment in Los Angeles Country, CA.	To compare the AiCure AI platform's cost-effectiveness to Directly Observed Therapy (DOT) for active tuberculosis treatment	Through smartphone-based monitoring, AiCure verified medication adherence using computer vision and machine learning. The study compared expenditures and health outcomes (QALYs) over a 16-month treatment horizon using a Markov model. Deterministic and probabilistic sensitivity studies were performed to confirm that the cost-effectiveness results were robust.	With 1.05 QALYs at $2668 per patient as opposed to 1.03 QALYs at $4894 per patient under DOT, AiCure was more affordable. 93.5% of simulations showed AiCure to be dominating. Nurse travel and time were minimized, resulting in significant cost savings.	Expand the use of AI-based TB therapy monitoring technologies, such as AiCure. Carry out more research on different kinds of patient populations to increase generalizability. Extend applicability to chronic illnesses that need adherence monitoring, such as more complicated TB cases.
16	Viktoria Jungreithmayr et al., 2021 31	Germany	Retrospective before-and-after study	320 patients (160 pre- and 160 postimplementation of CPOE system)	To evaluate how a computerized physician order entry (CPOE) system affects the standard of medication documentation.	A CPOE system with decision support was implemented without the use of direct AI.	Overall, prescription documentation ratings increased from 57.4% to 89.8% due to the CPOE system; however, several areas, such as allergy documentation and abbreviations, demonstrated deterioration.	Address particular challenges, including documentation, improve prescriber training, streamline processes, and make the CPOE system easier to use.
17	Jennifer Corny et al., 2020 15	France	Observational Study	Model development involved 10,716 patients (133,179 prescription orders), while validation involved 412 patients (3364 orders).	To evaluate how well a hybrid AI system that combines rule-based alerts and machine learning detects high-risk medication errors.	Lumio Medication is a hybrid decision support system that combines rule-based expert systems and machine learning (the LightGBM algorithm). The model was trained on 25 features, such as prescription drug alerts, demographics, medical history, and laboratory data. AUROC, precision-recall curves, and F1 score were among the metrics used in comparison to multicriteria query methods and CDS alert systems.	The system outscored CDS systems and multicriteria query methods, with an AUROC of 0.81 and an AUCPR of 0.75. It collected 74% of high-risk medications with 74% precision and decreased false alerts.	Expand the use of hybrid decision support systems in medical facilities. Use unstructured data to improve performance and integrate with electronic health records in real time. Validate with a variety of patient groups.
18	Bharath Dandala et al., 2019 32	USA	Quantitative Descriptive	1089 clinical notes with ADE annotations that have been deidentified	To identify ADEs in clinical narratives by employing neural networks to jointly model entities and relations..	BiLSTM with attention mechanism for relation extraction, and BiLSTM with CRF model for medical entity recognition. To improve detection performance, other resources (like FAERS) were incorporated. Three methods were used to evaluate the models: joint modeling, sequential modeling, and joint modeling using external resources.	For integrated ADE detection tasks, joint modeling using external resources produced the greatest F-measure (0.662), significantly exceeding conventional methods. This approach improved the context and precision of recognizing complex interactions.	To improve ADE identification, increase the usage of joint modeling techniques using external knowledge sources. Invest on domain-specific NLP enhancements for addressing limitations such as context-specific understanding, ambiguity, and spelling errors.
19	G Segal et al., 2019 33	Israel	Prospective Study	A 16-month monitoring in a single 38-bed inpatient department yielded 3160 patients and 78,017 medication orders.	To assess a probabilistic, machine learning clinical decision support system's (CDSS) clinical accuracy and usefulness in detecting medication errors.	An existing electronic medical record system integrated with a machine-learning-based CDSS that used statistically generated outliers. Based on real-time patient data, the system detected outliers and produced synchronous (real time) and asynchronous (postprescription) notifications for problems such as clinical mismatches, dose errors, and drug overlap. Physician response, clinical relevance, and alert burden were among the metrics evaluated.	With 43% of alerts resulting in revisions to medical orders and 85% of alerts considered clinically justified, the system achieved a low alert burden of 0.4% of prescriptions. The system's potential to lessen error-related alert fatigue was noted by physicians as an advantage.	Encourage additional hospital departments to implement probabilistic machine learning CDSS. Improve the detection of uncommon or newly developing medication error types and integrate greater datasets to improve system algorithms. Teach medical professionals how to properly evaluate alerts.
20	Moses E. Ekpenyong et al., 2019 34	Nigeria	Quantitative Descriptive	Akwa-Ibom HIV database (3168 treatment episodes, 1301 unique patients) and Stanford HIV database (5780 treatment change episodes, 1521 unique patient record)	To provide a hybrid framework for predicting patients’ responses to antiretroviral medication (ART) that combines deep learning, fuzzy logic, and multidimensional scaling (MDS).	A hybrid framework that combines MDS, Deep Neural Networks (DNN), and Interval Type-2 Fuzzy Logic (IT2FL). Drug combinations, viral loads, and CD4 levels were among the features examined. Supervised learning using MDS was used for training in order to reduce errors and visualize clustering. Fuzzy logic outputs and Levenberg–Marquardt algorithm-predicted optimal patient responses were used to calculate performance measurements.	Compared to the Stanford dataset, the Akwa-Ibom dataset displayed superior immunological response and better grouping, demonstrating the hybrid framework's higher predictive accuracy. This strategy performed better than current ART outcome modeling techniques.	Expand the utilization of the hybrid IT2FL-DNN-MDS system for ART in environments with restricted resources. To boost prediction accuracy, increase the diversity of the dataset and examine other variables like adherence and resistance profiles.
21	Xi Yang et al., 2019 35	USA	Quantitative Descriptive	The MADE1.0 challenge dataset contains 1089 deidentified clinical notes with 79,114 entities and 27,175 relations annotated.	To create a natural language processing (NLP) system (MADEx) that can identify drugs, adverse drug events (ADEs), and their connections from clinical notes.	The MADEx system used a combination of Support Vector Machines (SVMs) for relation extraction, long Short-Term Memory (LSTM) networks, and Conditional Random Fields (CRFs) for Named Entity Recognition (NER). The integrated pipeline made the identification of entities and relations in clinical literature possible.	MADEx obtained an F1 score of 0.8466 for relation extraction and 0.8233 for NER. The F1 score for the combined pipeline was 0.6125. The technology showed great promise for drug safety monitoring and pharmacovigilance.	To enhance ADE detection, enhance the NER and relation extraction integration. To efficiently handle complicated entities and relations, create postprocessing rules and investigate joint learning models. Include medical expertise in the framework..
22	Akbar K. Waljee et al., 2018 36	USA	Retrospective Cohort Study	20,368 patients from the Veterans Health Administration (2002–2009), including 351,112 visits.	To develop machine learning models predicting hospitalization and corticosteroid use in patients with IBD within six to twelve months.	Random Forest (RF) and logistic regression (baseline) models were used. Laboratory values, corticosteroid use, and prior hospitalizations were all incorporated into RF models. Using ensemble trees to determine variable importance, the RF model's sensitivity, specificity, and AUROC were verified. Other models extended outcomes to 12 months for sensitivity analysis and did not include immunosuppressive use.	With AUROC scores of 0.85–0.87 for 6-month predictions and 0.90 for 12-month predictions, RF models performed significantly better than logistic regression. Age, albumin levels, platelet counts, and previous results were the most predictive factors. These models have great potential for individualized treatment plans.	Use risk stratification methods based on machine learning in clinical settings to find high-risk patients with IBD. Extend the validation of the model to a variety of populations. Connect these models to EHRs so that decisions may be made at the point of care in real time.
23	Daniel M. Bean et al., 2017 37	UK	Quantitative Descriptive	Electronic health records (EHRs) were used to validate 3144 ADRs and 524 medications in the knowledge graph for ten spelected ADRs.	To use an EHR to test predictions and a knowledge graph-based machine learning algorithm for predicting unknown ADRs.	A knowledge graph comprising medications, adverse drug reactions, indications, and targets was created. To anticipate unknown ADRs, a machine learning method employed feature weighting and enrichment tests. NLP-extracted ADR mentions from EHRs were used to validate the predicted associations, which were then compared to random models and conventional techniques (logistic regression, SVM, and decision trees).	The prediction algorithm performed better than conventional techniques, with a validation percentage much higher than random models (92.3%). Hyperprolactinemia and hypersalivation were among the ADRs for which high-confidence predictions were confirmed. The technique showed great promise for predicting postmarketing adverse drug reactions.	Increase the integration of clinical data with knowledge graphs for practical uses. Concentrate on improving validation methods and utilizing advanced NLP pipelines to increase the precision of ADR detection and prediction..
24	Daniel L. Labovitz et al., 2017 38	USA	Randomized Controlled Trial	28 ischemic stroke patients on anticoagulants were randomly assigned to one of two groups: AI monitoring (n = 15) or control (n = 13).	To evaluate if visual confirmation of medication administration using an AI-based smartphone application (AiCure) improves adherence to anticoagulant therapy.	The AiCure platform used AI and computer vision to verify medicine consumption using a smartphone. The intervention group's patients were tracked in real time, and pill counts and plasma collection were used to measure adherence. For missed or late dosages, automated notifications and reminders were provided. Standard care was given to the control group. Usability assessments, plasma drug levels, and adherence rates were among the metrics.	According to plasma drug concentration levels, the AI monitoring group achieved significantly greater adherence rates (90.5%) than the control group (50%). The AI platform demonstrated promise in improving direct oral anticoagulant adherence. Patients successfully used the application, which received great usability ratings.	Expand the use of AI-based adherence platforms in more therapeutic domains where high drug adherence is necessary. Enhance patient EHR integration and investigate scalability and long-term usability across a range of demographics. Address smartphone access issues in low-tech or elderly groups.
25	Earle E. Bain et al., 2017 39	USA, UK, Russia	Phase 2 Clinical Trial	431 individuals with schizophrenia, 75 of whom were in the AI substudy at ten US sites.	To evaluate the effectiveness of modified directly observed therapy (mDOT) and an AI platform (AiCure) for tracking drug adherence.	The AiCure software used AI and computer vision to visually verify medicine consumption on mobile devices. mDOT adherence rates and pharmacokinetic data were compared with AI-generated adherence rates. Real-time alerts, timestamps, and facial recognition were among the features. Cumulative adherence and the relationship between AI data and pharmacokinetic samples were indications of performance metrics.	Subjects under AI monitoring had better adherence (89.7%) than those under mDOT (71.9%). By identifying patterns of nonadherence and enabling targeted action, the AI platform enhanced real-time intervention capabilities. One limitation was the low correlation (r = 0.33) between pharmacokinetic data and AI adherence rates.	Further application of AI adherence tools in normal care and clinical studies. To increase data correlation with pharmacokinetics and improve AI algorithms. Explore applications for longer trial durations and a variety of patient populations.
26	Gordon D. Schiff et al., 2015 40	USA	Mixed-method analysis and testing	63,040 CPOE-related cases out of 1.04 million medication error reports were examined. Testing for vulnerabilities in 13 CPOE systems.	To develop a taxonomy, examine CPOE-related pharmaceutical errors, and evaluate how susceptible the current systems are to these mistakes.	Error vulnerabilities in Computerized Physician Order Entry (CPOE) systems were examined through user testing and real-world settings to determine how simple it is to enter incorrect or troublesome orders.	Although CPOE systems reduced some errors, they were still quite prone to others; 28% of incorrect orders were submitted easily, and 79.5% of erroneous orders were successfully entered. It is essential that CPOE systems be carefully monitored and designed..	Iterative vulnerability testing, consistent error reporting, and strict error monitoring can all help to lower risks and improve the safety of the CPOE system.
27	Hossein Rahmani et al., 2015 41	Netherlands, UK	Quantitative Descriptive	146 drugs having at least five target proteins and adverse effects were found in DrugBank and SIDER.	To create a network-based approach that uses connections between medications to anticipate adverse drug reactions (ADRs).	A multilabel classification model based on networks is called Augmented Random Walk with Restarts (ARWAR). For predicting ADRs, the method involves adding side-effect nodes and edges to the Human Drug Network (HDN) and then using Random Walk with Restarts.	With an average 20% improvement in F-measure, ARWAR performed better than the Majority Rule Method (MRM). New ADRs for medications without previous ADR annotations were among the biologically significant predictions the approach produced.	Use ARWAR to anticipate adverse drug reactions in early drug discovery. To improve prediction accuracy, investigate integration with clinical datasets. Improve coverage for drugs with fewer target proteins and incorporate patient-specific data to overcome limitations.
28	Emmanuel Bresso et al., 2013 42	France	Quantitative Descriptive	554 drugs from the SIDER and DrugBank databases with related adverse effects.	To utilize machine learning to integrate drug properties and relational background knowledge in order to understand and predict drug side-effect profiles (SEPs).	SEPs were characterized utilizing pharmacological attributes and prior knowledge (e.g., pathways, protein interactions, and GO terms) using Decision Trees (DTs) and Inductive Logic Programming (ILP). Maximal frequent itemset (MFI) mining was used to extract SEPs, and cross-validation and direct testing with novel medications were used to assess the models.	While DTs offered superior specificity, ILP achieved higher sensitivity (predicting fewer false negatives). SEP prediction was greatly enhanced by prior knowledge. The findings encouraged the early detection of adverse effects during the drug-development process.	For early SEP identification, incorporate relational machine learning into drug discovery. For better predictions, make use of in-depth prior knowledge about pathways and interactions. Incorporate clinical data to models to further improve models
29	Aurel Cami et al., 2011 43	USA	Retrospective Study	A data snapshot was used to validate 809 medications and 852 adverse drug events (ADEs) from a 2005 database snapshot.	To develop and validate a model for detecting unknown adverse drug events (ADEs) that is based on predictive networks.	Logistic regression was used to create predictive pharmacosafety networks (PPNs) that incorporated intrinsic drug data with taxonomy and drug-ADE relationships. To predict new ADEs, the model integrated ADE taxonomy, network structure, and pharmacological characteristics.	The model showed great potential for early ADE prediction with an AUROC of 0.87, sensitivity of 0.42, and specificity of 0.95. Network techniques are a useful tool for generating hypotheses and conducting drug safety research.	Enhance the use of pharmacosafety networks in drug safety monitoring in Premarketing and Postmarketing surveillance. To increase prediction accuracy, concentrate on using larger datasets and investigating more factors. Encourage cooperation between pharmacologists and AI experts..
30	F. Hammann et al., 2010 44	Switzerland	Quantitative Descriptive	507 compounds with CNS, liver, kidney, and allergy ADR classifications were taken from the Swiss drug registry.	To develop decision tree models that use chemical structure and activity relationships for predicting adverse drug reactions (ADRs).	To find the molecular characteristics linked to ADRs in the CNS, liver, kidneys, and allergy endpoints, decision tree models (CART and CHAID) were employed. To identify high-risk and low-risk chemicals, descriptors such polar surface area, lipophilicity, and molecular complexity were employed.	High predictive accuracy was demonstrated by the models for CNS ADRs (89.7%), liver (90.2%), kidney (88.6%), and allergy (78.9%). The outcomes demonstrated how decision tree models can effectively and computationally anticipate ADR risks.	Early in the drug discovery process, use decision tree-based models to screen for the possibility of adverse drug reactions. For increased accuracy, enhance models using more descriptors and investigate how to integrate them with other machine learning technologies.

Research indicates that automated robotic dispensing systems, deployed in nations such as Saudi Arabia and Japan, significantly reduce medication dispensing errors. These technologies enhance the precision and efficiency of pharmacy operations, thereby improving overall patient satisfaction. Concurrently, digital health platforms, exemplified by AiCure, have proven effective in augmenting medication adherence, with notable success in managing complex treatment regimens for conditions including anticoagulation therapy and tuberculosis.

Furthermore, predictive analytics has emerged as a critical component in proactively identifying prescription inaccuracies and adverse drug reactions (ADRs), substantially augmenting patient safety and supporting clinical decision-making. Investigations from China have demonstrated the efficacy of sophisticated machine learning algorithms, including CatBoost and hybrid models, in the accurate detection of PIMs and the prediction of ADRs.

The findings from the included studies underscore the transformative potential of AI in advancing healthcare delivery and enhancing patient outcomes. Collectively, the evidence suggests that AI can serve as a valuable adjunct in clinical pharmacy, supporting healthcare professionals in providing more precise, effective, and patient-centered interventions. By aligning AI applications with the specific needs of diverse healthcare settings, these technologies may contribute to more consistent, reliable, and high-quality patient care on a global scale.

Comparative outcomes analysis

Across included studies, reported outcomes included: (i) DDI detection accuracy ranging 24–100% in ChatGPT-based medication therapy management (MTM) evaluations; (ii) machine learning models achieving AUROC up to 0.90 for hospitalization risk and PIM detection; and (iii) robotic dispensing associated with ≈80% lower dispensing errors and >30% higher pharmacist productivity. Reporting of patient-level clinical outcomes and long-term implementation effects was limited.

Quality assessment

Among the 30 included studies, the majority were quantitative descriptive in nature (n = 24), followed by quantitative nonrandomized designs (n = 3), mixed-methods studies (n = 2), and a single randomized controlled trial (n = 1). This variation reflects the diverse methodological approaches used to investigate AI applications in clinical pharmacy. The results of quality assessment are presented in Table 2.

Table 2.

Mixed methods appraisal tool (MMAT), version 2018.

Category of study designs	Methodological quality criteria	Alon Bartal et al., 2024	Don Roosan et al., 2024	Euibeom Shin et al., 2024	Merel van Nuland et al., 2024	Qiaozhi Hu et al., 2024	Ramya Padmavathy Radha Krishnan et al., 2024	Sara Grossman et al., 2024	Xiaoru Huang et al., 2024	Zhengliang Liu et al. 2023	Wu Xingwei et al., 2022	Tomoki Takase et al., 2022	Attayeb Mohsen et al., 2021	Hugo Lopes et al., 2021	Hisham Momattin et al., 2021	Jonathan Salcedo et al., 2021
Screening questions	S1. Are there clear research questions?	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
	S2. Do the collected data allow to address the research questions?	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Quantitative nonrandomized	3.1. Are the participants representative of the target population?	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	Can’t tell	N/A	N/A	Can’t tell	N/A
	3.2. Are measurements appropriate regarding both the outcome and intervention (or exposure)?	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	Yes	N/A	N/A	Yes	N/A
	3.3. Are there complete outcome data?	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	Can’t tell	N/A	N/A	Can’t tell	N/A
	3.4. Are the confounders accounted for in the design and analysis?	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	No	N/A	N/A	No	N/A
	3.5. During the study period, is the intervention administered (or exposure occurred) as intended?	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	Yes	N/A	N/A	Yes	N/A
Quantitative descriptive	4.1. Is the sampling strategy relevant to address the research question?	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	N/A	Yes	N/A	N/A	Yes
	4.2. Is the sample representative of the target population?	Can’t tell	No	No	No	Can’t tell	Can’t tell	No	No	No	Can’t tell	N/A	Can’t tell	N/A	N/A	Can’t tell
	4.3. Are the measurements appropriate?	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Can’t tell	Yes	N/A	Yes	N/A	N/A	Yes
	4.4. Is the risk of nonresponse bias low?	No	Can’t tell	NA	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A
	4.5. Is the statistical analysis appropriate to answer the research question?	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	N/A	Yes	N/A	N/A	Yes
Mixed methods	5.1. Is there an adequate rationale for using a mixed methods design to address the research question?	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	Yes	N/A	N/A
	5.2. Are the different components of the study effectively integrated to answer the research question?	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	Yes	N/A	N/A
	5.3. Are the outputs of the integration of qualitative and quantitative components adequately interpreted?	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	Yes	N/A	N/A
	5.4. Are divergences and inconsistencies between quantitative and qualitative results adequately addressed?	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	Yes	N/A	N/A
	5.5. Do the different components of the study adhere to the quality criteria of each tradition of the methods involved?	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	N/A	Yes	N/A	N/A

The MMAT appraisal highlighted three common concerns across nonrandomized and descriptive studies: (i) sample representativeness was often unclear or low, (ii) confounders were frequently not accounted for, and (iii) completeness of outcome data was sometimes unclear. By contrast, measurements and statistical analyses were usually appropriate. Overall, the evidence quality ranged from average to good, with no study rated as low quality.

Using the MMAT (2018), all included studies clearly stated their research questions and collected data that appropriately addressed those questions. Among quantitative descriptive studies, most applied relevant sampling strategies (e.g., Alon Bartal et al., Don Roosan et al., Euibeom Shin et al., Merel van Nuland et al., Qiaozhi Hu et al., Ramya Padmavathy Radha Krishnan et al., Sara Grossman et al., Xiaoru Huang et al., Zhengliang Liu et al., Wu Xingwei et al., Attayeb Mohsen et al., and Jonathan Salcedo et al.), although several had limitations in representativeness of the target population, which was frequently rated as “can’t tell” or “no.” Measurements were generally appropriate across these studies, and statistical analyses were consistently suitable. However, nonresponse bias was noted in a few studies (e.g., Alon Bartal et al.).

For quantitative nonrandomized designs, outcome measurement and fidelity of intervention administration were reported adequately (e.g., Tomoki Takase et al. and Hisham Momattin et al.), but control of confounding variables was generally lacking. Only one randomized controlled trial (Daniel L. Labovitz et al.) provided sufficient information on randomization, group comparability, completeness of outcome data, and blinding of outcome assessors.

The single mixed-methods study (Hugo Lopes et al.) demonstrated a clear rationale for the mixed-methods design, effective integration of components, and adequate interpretation of findings, while also addressing divergences between qualitative and quantitative results.

Overall, while methodological quality was acceptable across the included studies, the most common limitations related to representativeness of study samples and insufficient consideration of confounding factors.

Discussion

This review synthesizes current evidence on the application of AI across various domains of clinical pharmacy, with particular emphasis on its impact on medication safety, clinical decision support, and patient adherence. While several AI-driven tools demonstrate strong predictive performance in retrospective analyses or simulated environments, their external validation, integration into routine workflows, and demonstration of patient-level outcomes remain limited and inconsistent.

Artificial intelligence holds significant promise for enhancing clinical pharmacy services by improving precision, efficiency, and decision-making. As healthcare systems become increasingly complex, AI-based solutions are being applied to support pharmacists in diverse tasks, including the detection of ADRs at an individual patient level, optimization of medication management, identification of PIMs, automated prescription verification, and pharmacometric modeling. By leveraging machine learning algorithms, natural language processing, and predictive analytics, these tools have the potential to improve patient outcomes, reduce medication errors, and optimize resource utilization.

Despite these advancements, several challenges remain. Many AI applications are developed and validated in controlled or single-center settings, limiting their generalizability. Moreover, the integration of AI tools into existing clinical workflows is often insufficiently addressed, which can hinder adoption and limit measurable impact on clinical practice. Importantly, few studies report outcomes at the patient level, underscoring the need for prospective evaluations and real-world implementation studies.

In this review, AI applications were categorized based on prevalent themes and areas of clinical significance identified across 30 studies, including adverse drug event (ADE) detection, clinical decision support, and pharmacometrics. These categories were selected based on their frequency in the literature, documented clinical relevance, and potential impact on pharmacy practice. The sorting of AI applications in clinical pharmacy was based on prevalent themes and focal points identified in the 30 papers taken into account. These categories such as ADEs, clinical decision support, and pharmacometrics were selected based on their frequency, clinical significance, and documented impact in the literature:

Adverse drug events and side effects

In this review, the term “ADEs” is used as a broad umbrella term that includes ADRs and adverse side effects (ASEs), unless otherwise specified. The majority of medical errors during hospitalization are the result of invasive procedures, hospital-acquired infections, and the use of drugs and medical devices. However, adverse events can manifest as mistakes during any stage of medical treatment. Prescribing drugs and their administration are more frequently occurring sources of errors. ⁴⁵ However, this part of medication errors can be potentially prevented by clinical pharmacists’ medication review, which is currently a Gold standard for drug verification. On the other hand, the process is time-consuming and has consistency issues.

In the present review, analysis of the 30 selected studies highlighted the primary roles of AI in clinical pharmacy, including the identification of DDIs, prediction of adverse drug effects based on patient demographics, and detection of side effects (SEs). A recent study by Bartal et al. underscores the transformative potential of AI in detecting ADEs and SEs by leveraging diverse public data sources, including social media platforms and large language models (LLMs) such as ChatGPT. Utilizing a Named Entity Recognition (NER) model, AI was able to identify 21 previously unreported ASEs associated with GLP-1 receptor agonists, including irritability and numbness. This AI-driven methodology not only enhances drug safety monitoring but also facilitates rapid identification of latent risks, supporting proactive post-market pharmacovigilance and ultimately improving patient safety. ¹⁸

A recent study by Roosan et al. highlights the pivotal role of AI in enhancing clinical pharmacy, particularly in the context of MTM. The study evaluated the performance of ChatGPT 4.0 in detecting DDIs, recommending alternative therapies, and developing comprehensive management strategies. Across 45 patient cases categorized as easy, complex, and very complex, ChatGPT demonstrated 100% accuracy in identifying DDIs, providing actionable recommendations, and formulating detailed care plans. These findings underscore the potential of AI-driven tools to support pharmacists in optimizing therapy, reducing medication errors, and improving patient outcomes. ¹⁹

The study highlights AI's potential to improve clinical decision-making, but certain limitations, like the inability to make accurate dose recommendations or ask follow-up questions. Artificial intelligence-driven approaches like ChatGPT may contribute to improvements in patient safety and care standards in clinical pharmacy practice; however, real-world evidence and further evaluation are required. ¹⁹

A study by Radha Krishnan et al. highlights the limitations of AI models, such as ChatGPT-3.5, in clinical pharmacy applications. The study found that while the model exhibited high specificity, it demonstrated gaps in sensitivity and consistency when identifying clinically significant DDIs compared to pharmacist evaluations. These findings emphasize the need to enhance AI tools through comprehensive training on clinically relevant datasets, optimization of prompt strategies, and integration into structured clinical decision support systems. Such improvements are critical to increasing contextual awareness, reliability, and safe application of AI in real-world clinical settings. ²²

A study by Mohsen et al. demonstrates the potential of DL to predict ADRs by integrating gene expression profiles with adverse event data. Using advanced neural network models, the study elucidated the molecular mechanisms underlying ADRs through pathway enrichment analysis and achieved a high predictive accuracy of 89.4%. These findings underscore the importance of developing comprehensive, high-quality integrated datasets, and designing models capable of handling diverse ADRs. Furthermore, the study highlights the need for close collaboration between pharmacologists and AI researchers to refine these systems, thereby enhancing their applicability in drug safety monitoring and pharmacological discovery. ²⁸

Corny et al. conducted a study to assess the accuracy of an algorithm-based medication review system. A total of 10716 (with 133179 prescription orders) patients’ data were included in the study for medication review. After reviewing, 20% of patients’ prescriptions were found with the risk of a potential medication error which was either overdosing or underdosing or noncompliance with the drug formulary. Simultaneously, the data were used to generate an algorithm with a rule based expert system. In the process of validation of their model, 412 patients (with 3364 prescription orders) dataset was randomly selected. In 42% of patients, pharmaceutical intervention was suggested by the algorithm while demonstrating 74% accuracy. ¹⁵

In a similar study, Segal et al. developed a machine learning based computerized decision support system (CDSS), named MedAware, for finding potential medication based errors and generating the real time alerts. Out of total alerts generated, 89% were found accurate, 85% were clinically valid, and 80% were clinically useful. It was observed with low alert burden and low false alarm rate as compared to existing CDS (Clinical Decision Support) system. Hence, the prescribing doctor could focus on the alerts and change their behavior of prescribing medication after following 43% of errors. Another benefit of using this system was its ability to do surveillance of post prescribing ADEs. Over 60% of alerts were generated after the medication was already prescribed to the patients. ³³

A 2019 study by Bharath Dandala highlights the advancements of AI in detecting ADEs from clinical narratives using neural network-based approaches. The study employed joint modeling strategies, combining BiLSTM-CRF for medical entity recognition with BiLSTM incorporating attention mechanisms for relation extraction. The performance of these models was further enhanced by integrating external resources such as the FDA Adverse Event Reporting System (FAERS), achieving the highest F-measure of 0.662 for ADE identification. These findings underscore the importance of domain-specific natural language processing (NLP) enhancements to address challenges such as contextual understanding, ambiguity, and spelling variations, paving the way for more accurate and reliable recognition of ADEs from clinical texts. ³²

Yang and his fellow researchers in 2019 demonstrate how well NLP systems such as MADEx can identify ADEs from clinical notes, obtaining strong F1 scores for connection extraction and NER. To increase ADE identification in clinical practice, nevertheless, integration issues decreased pipeline performance overall, highlighting the necessity of better entity–relation integration, joint learning models, and the integration of domain-specific expertise. ³⁵

A study by Bean et al. illustrates the application of knowledge graph-based machine learning algorithms to predict previously unreported ADRs using electronic health record (EHR) data. This approach outperformed traditional methods, achieving a validation accuracy of 92.3% and successfully predicting high-confidence ADRs, including hyperprolactinemia and hypersalivation. The findings highlight the potential of integrating clinical data with knowledge graphs, alongside advancements in natural language processing (NLP) techniques, to enhance the detection and prediction of ADRs, thereby supporting more effective pharmacovigilance and patient safety strategies. ³⁷

In contrast to a study by Rahmani et al. presents a network-based approach, Augmented Random Walk with Restarts (ARWAR), for predicting ADRs. By enhancing the Human Drug Network with side-effect nodes and edges, ARWAR achieved a 20% improvement in F-measure over traditional methods. For medications without previous notes, it also predicted biologically important ADRs. In order to improve prediction accuracy and application in early drug discovery, future research should concentrate on integrating clinical datasets, expanding coverage for medications with fewer target proteins, and incorporating patient-specific data. ⁴¹

Bresso and his fellow researchers in 2013 emphasize how machine learning can be used to predict drug side-effect profiles (SEPs) by combining relational information and pharmacological characteristics. Decision Trees (DTs) provided higher specificity, while Inductive Logic Programming (ILP) offered better sensitivity, with prior knowledge significantly improving SEP prediction. With future improvements concentrating on utilizing extensive pathway or interaction data and including clinical datasets to further refine predictions, the strategy exhibits potential for early adverse effect detection. ⁴²

Predictive pharmacosafety networks (PPNs) offer significant potential for the early detection of previously unrecognized ADEs, as demonstrated by a study by Cami et al. By integrating pharmacological data with ADE taxonomies and relational information, the model achieved high specificity (0.95), moderate sensitivity (0.42), and an area under the receiver operating characteristic curve (AUROC) of 0.87. These findings underscore the promise of network-based approaches in drug safety research while highlighting the need for future improvements, including the utilization of larger datasets, incorporation of additional predictive variables, and closer collaboration between pharmacologists and AI specialists to enhance prediction accuracy and clinical applicability. ⁴³

The application of decision tree models (CART and CHAID) for predicting ADR hazards based on molecular properties including polar surface area, lipophilicity, and chemical complexity has been highlighted in another research article by Hammann et al. As computational techniques for early ADR screening, the models demonstrated good prediction accuracy for CNS (89.7%), liver (90.2%), kidney (88.6%), and allergy (78.9%) ADRs. To further improve predicted accuracy, future research should concentrate on adding more descriptors and combining these models with other machine learning approaches. ⁴⁴

Despite the promising results reported, many studies are limited by small sample sizes and restricted external validation, which constrain the generalizability of their findings. ChatGPT-based evaluations, in particular, often rely on simulated patient scenarios rather than real-world clinical environments. Additionally, numerous AI models have been assessed primarily in retrospective or experimental settings, frequently without longitudinal follow-up, raising concerns about their true clinical effectiveness. This reliance on limited datasets or the absence of pharmacist-based comparative standards increases the risk of overestimating the readiness and reliability of AI tools for routine clinical practice.

Clinical decision support and personalized therapeutic strategies

Several studies have evaluated clinical decision support systems alongside therapy prediction tools, and therefore, these themes are discussed together to highlight their complementary roles and combined significance in enhancing clinical decision-making. There have been positive advancements in the application of AI to treatment response prediction and clinical decision support. According to studies such as Segal et al. ³³ and Hu et al., ⁷ machine learning can help detect PIMs and prevent medication errors. Segal et al. demonstrated the effectiveness of a probabilistic CDSS in identifying real-time medication issues with a minimal alert burden, whereas Qiaozhi Hu used a CatBoost-based multilabel classification model to achieve excellent predictive accuracy.

Additionally, with AUROC scores of 0.85–0.90, the study by Waljee et al. shows how machine learning models, such as Random Forests, can predict hospitalizations and corticosteroid use in patients with inflammatory bowel disease (IBD). These models provide up the possibility to individualized treatment regimens in addition to helping with risk stratification. ³⁶

A study by Labovitz et al. demonstrated that AI-based systems, such as AiCure, significantly improved adherence to anticoagulant therapy compared with standard care, highlighting the potential of AI-driven interventions in optimizing medication adherence monitoring. This demonstrates how AI may improve patient engagement and increase therapeutic efficacy. ³⁸

Although decision support algorithms have demonstrated impressive performance metrics, many studies were conducted using controlled datasets with limited demographic diversity. Additionally, the integration of these systems into existing hospital infrastructure was not consistently addressed, raising concerns about their practical applicability and compatibility with real-world clinical workflows.

Personalized medication and optimal dosing suggestions

The potential for personalized medication and optimal dosing strategies has greatly increased with the introduction of AI into clinical practice. Artificial intelligence-powered tools and models demonstrated potential in customizing treatments to meet the needs of each patient, improving therapeutic results and adherence.

The 2017 study by Labovitz and his fellow researchers emphasizes how AI-based tools such as AiCure can help patients adhere to their anticoagulant treatment. The platform used personalized interventions, reminders, and real-time tracking to achieve adherence rates of 90.5% in the intervention group and 50% in the control group. These results highlight AI's capacity to modify medication regimens in response to unique behavioral patterns, encouraging adherence and ensuring the best possible treatment outcomes. ³⁸

Similarly, a study by Waljee et al. demonstrated that machine learning models, specifically Random Forests, can accurately predict key clinical outcomes in patients with IBD, including hospitalizations and corticosteroid use. By incorporating patient-specific variables such as age, albumin levels, and platelet counts, these models achieved AUROC scores of up to 0.90, illustrating their potential to optimize dosing, enable risk stratification, and support data-driven clinical decision-making. ³⁶

Additional advancements are highlighted in the study by Liu et al., where LLMs, including ChatGPT and GPT-4, were shown to replicate key clinical pharmacist functions. These models were capable of predicting therapeutic outcomes, designing individualized medication regimens, and analyzing data from intensive care unit patients. Such applications illustrate the potential of AI to optimize drug therapy using real-time clinical data, thereby supporting precision medicine and personalized pharmacotherapy. ²⁵

Although many studies discussed in this section report high predictive accuracy, only a limited number assess downstream clinical outcomes, such as actual therapeutic efficacy or the reduction of adverse events. Sample sizes are often small, and tools like ChatGPT have not been rigorously compared against standard clinical judgment. Furthermore, the reliance on retrospective datasets, absence of randomized clinical trials, and frequent exclusion of high-risk populations highlight the need for caution when interpreting these findings and underscore the importance of prospective validation in diverse patient cohorts.

Accuracy in prescription

Patient safety critically depends on the accuracy of prescriptions, and the integration of AI into clinical workflows has demonstrated promising advances in this domain. Several studies have explored the use of AI and machine learning techniques to enhance clinical decision-making, reduce medication errors, and improve prescribing practices. In a study by Radha Krishnan et al., ChatGPT-3.5 was evaluated for its ability to predict DDIs using real patient data. While the model exhibited high specificity (>95%), its sensitivity (0.24) and concordance with pharmacist assessments were modest. These findings underscore the current limitations of AI in managing complex prescription scenarios and highlight the need for more robust algorithms to ensure comprehensive DDI detection in clinical practice. ²²

In contrast, van Nuland et al. demonstrated that ChatGPT-4 outperformed pharmacists in answering standard clinical pharmacy questions, achieving an overall accuracy of 79% compared with 66% for pharmacists. However, the study also highlighted challenges in addressing complex prescription scenarios and managing region-specific recommendations, indicating that AI systems must be adapted to local clinical practices and regulatory frameworks to achieve optimal performance and practical applicability. ²¹

Platforms based on machine learning have also demonstrated an enormous amount of potential. A probabilistic CDSS coupled with an EHR was evaluated by Segal et al. The system obtained 85% clinical justification for alerts, with 43% of these resulting in prescription changes. ³³ Similarly Xingwei et al. created a machine learning technology that achieved an AUC of 0.8341 for PIP detection in order to identify prescription errors, inappropriate medications, and prescriptions in elderly cardiovascular patients. ²⁶

To identify high-risk medication errors, the hybrid decision support system Lumio Medication (Corny et al.) integrated machine learning with rule-based alerts. It identified 74% of high-risk medications with 74% precision, achieving an AUROC of 0.81. This method improved the precision of prescription error detection and reduced false alerts by incorporating advanced analytics. ¹⁵

Although AI systems such as ChatGPT-4 outperform pharmacists in certain tasks, their limitations in regional adaptation and context-specific clinical knowledge remain inadequately addressed. Furthermore, many models have evaluated DDI detection in isolation rather than within the context of complex polypharmacy, limiting their external validity. Several tools also fail to report false negatives transparently, which poses potential safety risks if these systems are implemented in clinical practice prematurely.

Pharmacometrics

The application of AI and machine learning models has significantly advanced pharmacometrics, the quantitative assessment of drug efficacy and safety. Studies have demonstrated that these approaches can enhance patient-specific outcomes, optimize dosing regimens, and predict both therapeutic responses and potential adverse effects, thereby supporting more precise and individualized pharmacotherapy.

Random Forest models were used to predict hospitalization and corticosteroid use in patients with IBD, according to a 2018 study by Waljee et al. The models obtained AUROC scores as high as 0.90 by incorporating patient-specific variables such as lab values and previous hospital stays, offering important information when designing personalized treatment regimens. These results demonstrate how AI can maximize medication efficacy while reducing hazards. ³⁶

Similarly, Ekpenyong et al. proposed a hybrid framework combining fuzzy logic with DNNs to predict outcomes of antiretroviral therapy. By incorporating patient-specific parameters such as viral load and CD4 counts, this approach demonstrated higher predictive accuracy compared with traditional methods. The integration of pharmacometric parameters in these models ensures greater precision in therapeutic decision-making and supports individualized treatment strategies. ³⁴

Additionally, Labovitz et al. used AI platforms to track anticoagulant therapy adherence and correlate it with pharmacokinetic outcomes. These applications demonstrate how important it is to combine real-time monitoring with pharmacometric details to improve therapeutic results and adherence. ³⁸

Despite the strong predictive potential of AI in pharmacometrics, many models have relied on narrowly restricted datasets or surrogate outcomes. Real-world integration remains limited, and evaluation in multiethnic or resource-constrained settings is scarce. Furthermore, insufficient transparency in methodological aspects such as variable selection and handling of missing data reduces confidence in the robustness and generalizability of these models across diverse patient populations.

Community and dispensing pharmacy

Although limited in number, the studies in this category reveal emerging trends that merit brief discussion. Automation and AI have revolutionized community and dispensing pharmacy procedures, improving accuracy and efficiency while assisting pharmacists in making decisions. The study by Shin et al. demonstrates how ChatGPT and other AI tools can help with tasks in community pharmacies. The requirement for geographically customized AI models that are adapted to community pharmacy settings is highlighted by limitations associated with region-specific guidelines. ²⁰

Further developments are seen in studies such as Takase et al. and Momattin et al., which found that robotic systems greatly decreased dispensing errors and increased workflow efficiency, demonstrated further developments in pharmacy dispensing. These findings demonstrate the way automation might reduce human error and free up pharmacists to concentrate more on patient-centered activities.^17,27

Despite these advancements, challenges remain in the seamless integration of robotic and AI technologies into community pharmacy practice. Successful adoption will require additional research focused on localized adaptations, rigorous validation, and comprehensive training for pharmacy staff. By addressing these gaps, AI and automation have the potential to enhance the quality of community and dispensing pharmacy services, promoting safer, more efficient, and patient-centered care. Additionally, most evaluations are conducted in controlled environments, and long-term effects on staffing needs and pharmacist roles remain underexplored.

Computerized prescriber order entry

This category includes fewer studies but provides important insights into how AI-enhanced computerized prescriber order entry (CPOE) systems can impact prescription quality and safety. The section remains concise due to the limited but relevant scope of available research. With the goal of reducing prescription errors and improving the overall standard of medication management, CPOE systems have emerged as a crucial part of modern healthcare. The present review provides important knowledge on CPOE's efficacy and points out areas that still require development.

A study by Jungreithmayr et al. evaluated the impact of a CPOE system on hospital medication documentation standards. The study reported a significant improvement in prescription documentation scores, increasing from 57.4% to 89.8%. However, persistent issues such as incomplete allergy documentation and frequent abbreviation errors underscore the need for ongoing prescriber training and continuous system optimization to ensure accurate and comprehensive medication records. ³¹

Similarly, Schiff et al. analyzed over one million prescription error reports to assess vulnerabilities in CPOE systems. The study found that, while CPOE implementation reduced certain types of errors, 28% of incorrect orders were still successfully recorded, indicating persistent system vulnerabilities. These findings underscore the importance of iterative vulnerability testing, consistent error reporting, and ongoing system optimization to enhance medication safety and address remaining gaps in CPOE performance. ⁴⁰

Together, these studies show how CPOE systems can improve prescriber accuracy and reduce errors. However, in order to take full advantage of their benefits, ongoing issues including usability problems, insufficient data entries, and system vulnerabilities must be resolved. In order to ensure optimal efficiency and user compliance, future initiatives should concentrate on integrating real-time feedback mechanisms, advanced decision assistance tools, and frequent system audits.

Computerized prescriber order entry systems can improve patient safety in a variety of clinical settings, reduce errors, and improve prescribing workflows by addressing these limitations.

Alert fatigue, inadequate user training, and insufficient interface testing persist in undermining efficacy. Furthermore, numerous systems operate independently, without effective connection with comprehensive EHRs or decision support systems.

The cost effectiveness and other benefits

Artificial intelligence and automated systems have proven to be very cost-effective and operationally beneficial when used in pharmacy practice, especially when it comes to decreasing errors, streamlining processes, and enhancing patient outcomes. In the current review, a number of studies emphasize these features.

Salcedo et al. study showed how cost-effective the AiCure AI technology is for tracking drug adherence during TB treatment. AiCure maintained greater adherence rates (93.5%) while lowering expenses per patient from $4894 to $2668 when compared to Directly Observed Therapy. This demonstrates that by eliminating the need for time-consuming monitoring, AI technologies not only enhance health results but also result in significant cost savings. ³⁰

Likewise, Momattin et al. investigated the implementation of a robotic pharmacy system in a hospital setting. The system not only reduced dispensing errors but also increased pharmacist productivity by 33%, decreased patient wait times by 53%, and improved patient satisfaction by 93%. The achievement of return on investment within just three and a half years further highlights the cost-effectiveness and operational benefits of integrating robotic technologies into pharmacy practice. ¹⁷

The benefits go beyond cost savings. For example, robotic dispensing devices save dispensing errors by 80%, freeing up pharmacists to concentrate on clinical treatment in a study conducted by Takase et al. By reallocating resources, pharmacy staff burnout is decreased and the quality of care is improved. ²⁷

Despite these advantages, the integration of AI and robotic systems can be costly and often requires extensive training as well as modifications to existing workflows. To optimize their impact, future research should explore strategies to reduce implementation costs, evaluate long-term benefits across diverse healthcare settings, and enhance training programs for pharmacy staff to ensure effective and safe utilization.

Overall, automation and AI technologies provide cost-effective solutions with multiple benefits, including enhanced productivity, improved patient safety, and more efficient resource utilization. Their integration into pharmacy practice has the potential to transform healthcare delivery while maintaining financial sustainability and supporting high-quality patient care.

Critical appraisal and generalizability

The body of evidence in this review shows promising performance for several AI tools; however, methodological limitations and external-validity concerns restrict how confidently the findings can be generalized. First, many evaluations were retrospective or based on simulated or convenience scenarios (e.g., LLM/ChatGPT tasks), which are vulnerable to spectrum and selection biases and may overestimate clinical readiness compared with prospective, real-world use. Several studies also reported small samples or lacked longitudinal follow-up, limiting the stability of estimates and the assessment of sustained effectiveness.

The MMAT appraisal further indicates recurring risks: unclear representativeness, incomplete outcome data in some nonrandomized designs, and limited control of confounding all of which can bias effect estimates. In contrast, measurement approaches and statistical analyses were generally appropriate, and no included study was categorized as low quality overall.

Chronological Evolution of AI Applications

The application of AI in clinical pharmacy has demonstrated a distinct evolution throughout time. Initial research 2010–2015 concentrated on rule-based systems and computerized order entries. From 2016 to 2020, machine learning started to emerge in medication safety surveillance and risk prediction models. Since 2021, big language models such as ChatGPT and GPT-4 have been progressively investigated for clinical decision support, drug evaluation, and individualized treatment. This timeline demonstrates a transition from static automation to dynamic, real-time decision-making tools, highlighting AI's development and adaptation to intricate pharmacy settings.

Future directions perspectives

As AI continues to evolve, emerging trends such as explainable AI, real-time predictive analytics, and integration with wearable technologies are gaining traction. However, challenges remain in ensuring data transparency, minimizing algorithmic bias, and developing ethical guidelines for AI deployment in pharmacy settings. Policymakers and academic institutions must collaboratively define frameworks for safe, equitable, and evidence-based AI implementation in clinical pharmacy.

Conclusion

This study investigated the multifaceted applications of AI within clinical pharmacy, including ADE detection, clinical decision support, personalized medicine, pharmacometrics, and operational automation. The reviewed literature indicates that AI has significant potential to augment pharmacist-led care by improving accuracy, efficiency, and patient outcomes.

However, several limitations were identified across the current body of evidence. Many studies relied on retrospective or simulated datasets, lacked external validation, or were conducted within narrowly defined populations. The integration of real-world data into pharmaceutical workflows was often limited or unassessed, and the long-term therapeutic impacts of AI interventions remain largely unquantified. These shortcomings constrain the generalizability and translational potential of many existing AI models.

This review itself has limitations, including heterogeneity in the methodological quality of included studies, a restriction to English-language publications, and the potential for publication bias. Future research should prioritize large-scale, prospective clinical trials in diverse real-world settings, improve model transparency and interpretability, and focus on the seamless integration of AI within existing EHR systems. The establishment of clear regulatory frameworks, dedicated training for healthcare professionals, and robust ethical oversight will be critical to ensuring the safe and effective adoption of AI in pharmacy practice.

In conclusion, while AI is not a substitute for clinical pharmacist expertise, it represents a powerful adjunct tool. Its successful implementation to advance patient care is contingent upon rigorous validation, responsible deployment, and continuous evaluation.