Supervised ML consists in labeling and training datasets so that they can predict expected outcomes. The term supervised refers to the fact that the model is provided with the outcome of the variable to be predicted (target variable) once the dataset has been trained. The essence of the process lies in the mapping of a dataset where the input variables (characteristics) and the outcomes (labels) are known, which makes it possible for the computer to learn and generate patterns following an appropriate training process. The labelling of a variable depends on whether it is a categorical variable (e.g., “adverse event” or “non-adverse event”) or a continuous variable (e.g., plasma hemoglobin levels). This allows systems to make accurate predictions about a target variable in new cases.

Supervised ML systems can be used in hospital pharmacy (HP) departments to predict the efficacy or the toxicity of a drug in a given patient, which makes it possible to make informed decisions when the goal is to personalize a treatment or minimize potential side effects. Hospital pharmacists can play an essential role in supervised learning programs, contributing to the identification of outcome variables it could be beneficial to predict, determining what input characteristics may provide useful information to make such predictions, and identifying datasets that contain these characteristics, labels, or outcomes.

Table 1 describes the most common algorithm-based supervised ML techniques used in healthcare.5

Unsupervised ML systems use training data to identify patterns and trends, or to automatically generate classifications into different groups (clustering) based on various unlabeled variables. These kinds of models provide an understanding of input variables, making it possible to classify, though not predict, them. At the same time, the most recent ML techniques allow the use of data-based discoveries, ranging from unknown drug-related adverse events to emerging diseases.4 A specific example of the use of unsupervised ML in HP departments in the clustering of patients according to certain variables (e.g., genetic or clinical) with a view to permitting a personalized choice of the best drug for every new case.

In addition to the discovery of new clinical–pharmacological knowledge, these models may facilitate the adoption of clinical decisions regarding patients' drug therapy as well as promote further development of personalized medicine.

The need of expert clinical and pharmacological knowledge for identifying patterns and evaluate them, and for validating the results obtained, means that hospital pharmacists are ideally placed to participate in, or even lead, unsupervised ML projects.

Reinforced learning systems are trained with data obtained from interactions with the environment or with simulations. While they execute different actions, models receive feedback in the form of rewards or penalties, which facilitates the learning process and the continuous refinement of their parameters.

Unlike supervised ML, previous labeling of the data is not required here as the model learns through continuous exploration and adaptation, depending on the feedback received. This means that the model learns and makes decisions by itself during the learning process.

Reinforced learning algorithms can be used by HP departments to optimize dosing schedules, particularly in patients exhibiting continuous changes in their clinical status. In these cases, the rewards of the system may encompass various effectiveness indicators, such as the reduction of symptoms or changes in some analytical or clinical variable. As systems accumulates more information on the patient and adapt the doses they receive, they develop the ability to administer drugs in an increasingly efficient and personalized manner, accurately catering for the needs of each individual patient.

Artificial neural networks (ANNs), inspired by the functioning of the human nervous system, are ML models whose structure is formed by layers of interconnected nodes, known as artificial neurons. Each neuron corresponds to a computational unit responsible for processing and transforming information and conveying the output to the nodes in the next layers.

These models can be divided into 2 categories, depending on the number of layers in the network. Models with just a few layers of connected neurons are known as non-deep or shallow models, while those with a larger number of layers and greater complexity are called DL models.

In DL models, each layer in the neural network performs mathematical operations and non-linear transformations of the initial data it is provided with. As progress is made toward the more internal layers in the network, mathematical operations become increasingly complex and non-linear. This leads to the creation of more and more abstract representations of the information, which are less and less directly connected with the original data input. For that reason, DL algorithms require larger numbers of operations than traditional ML models, which significantly increases the magnitude of the datasets and the computational power required by the system.6

Unlike standard ML systems, where the models developed allow machines to learn from the data and make decisions based on patterns, DL systems are characterized by capturing complex patterns and highly abstract representations from the data.

Among its multiple applications in healthcare, DL has been proposed as a tool to improve the diagnostic process, enhance the prediction of the efficacy of clinical outcomes, optimize drug therapies, and facilitate treatment personalization on the basis of patients' individual (genetic, clinical, etc.) characteristics. Other potential uses of DL in healthcare are related with developing smart medical records, automating administrative tasks, improving the management and assignment of hospital resources, and even identifying disease patterns among the population.

An analysis of patients' data and clinical characteristics makes it possible to make recommendations based on scientific evidence. Consequently, AI algorithms may help optimize drug therapy, identify the most appropriate pharmacological alternatives, develop more precise dosing schemes, and supervise the effectiveness of treatments.38 The data generated should allow more informed decision-making, improve health outcomes and reduce the incidence of adverse events.

AI is capable of analyzing specific patient data such as genetic profiles, clinical and lab test results, and identifying response biomarkers and subpopulations at risk of drug inefficacy, thus anticipating the patients' response to a given drug, and detecting molecular targets and interactions.39

As regards pharmacokinetic monitoring, AI can be a useful adjunct to traditional systems,40 assisting in the decision-making process (exposure-concentration prediction and dose optimization), in the development of models (covariate selection, model validation, and patient classification), and in the administration of precise doses (selection of the most appropriate model and estimation of parameters). The goal is to prepare more accurate therapeutic plans, predict each individual patient's response to treatment and, as a result, optimize drug selection and dosing, based on the knowledge extracted from large and complex sets of molecular, pharmacokinetic, epidemiological, clinical, and demographic data.

Methods such as DL allow us to predict adverse drug reactions with a good AUC value (around 0.85).41 They also help discover new reactions by analyzing the chemical structure and mechanism of action of the drug, as well as the effects associated with polypharmacy.42

AI is capable of rapidly analyzing the patients' profiles, including their medical history and the drugs they are taking, in order to identify potential drug–drug interactions.43 This is an effective way of avoiding harmful combinations, thus promoting patient safety.

AI-based predictions allow identification of high-risk patients and early optimization of treatments, thereby improving patient safety and reducing hospital readmissions.

The use of conversational agents44 able to interact with patients in a personalized and contextualized way has all but revolutionized patient follow-up as, when combined with other technologies, they can provide health data such as vital signs, patient-reported outcomes, adverse events, etc., which are extremely useful for clinical and pharmacotherapeutical decision-making.

The collection and analysis of patients' health data makes for more proactive interventions, e.g., when a treatment plan needs to be adjusted, and allows clinicians to offer lifestyle recommendations for effectively managing chronic conditions.

AI has also been shown to be an effective tool for evaluating and optimizing adherence.45 AI systems may act as a coach, generating automatic medication reminders, sending text and voice messages and collecting information on patients' lifestyles, beliefs, and opinions.

The use of AI in these scenarios simplifies patient follow-up, provides real-time information and improves medication adherence and persistence, which typically results in better health outcomes.

Generative AI is extremely useful in remote pharmaceutical care. AI-based selection systems46 draw on datasets containing patient symptoms, drug information, and medical records among others to generate personalized questions. These systems allow a more effective characterization of health problems, making recommendations including referring patients to a health center, changing a pharmaceutical appointment, sending a personalized message, or contacting a hospital pharmacist. In these circumstances, it is possible to adopt a proactive stance in order to prevent or at least minimize errors and adverse events.

An analysis of large datasets allows algorithms to predict the demand for medications, identifying expiry dates and optimizing stock levels to allow a stable supply, the reduction of waste and the prevention of shortages, all the while improving operating efficiency and the quality of care.38 Use of AI in HP departments should reduce inefficiencies, freeing up pharmacists' time so that they can focus on patient care and clinical decision-making.

Funding and reimbursement systems have in the last decade moved away from traditional practices based on fixed unit costs to more innovative result- and value-based solutions with an emphasis on the attainment of specific therapeutic goals.

Applying NLP to medical records may provide precise information on the activity or the effectiveness/safety of a given drug vis-à-vis its competitors in the real world. Moreover, AI can also be used to determine the risk inherent in a certain agreement or contract.47

The new Center for Artificial Intelligence in Pharmaceutics (CIAM), a pioneering initiative in Europe coordinated by the Health and Social Care Consortium of Catalonia and developed by Catalonia's Department of Health in collaboration with the Agency for Health Quality and Evaluation, the TIC-Salud Social Foundation and the Catalan Health Service, has recently been established. CIAM's mission is to use AI to optimize all processes related to effectiveness, patient safety, and the sustainability of the health system.

The design of virtual agents44 may provide patients with the information they need as well as meet their informative needs with respect to their medications. This should help them make better informed decisions on how to manage their health.

AI solutions may be designed to monitor adherence, reduce the waiting times associated with the dispensing of drugs, collect data on patients' habits and opinions, and remotely manage their drug regimens in order to improve convenience and their overall experience. AI also empowers patients to effectively manage their health, thereby improving their health outcomes.

All the above should result in more effective pharmacist-patient communication and in the availability of valuable information on patients' preferences, concerns, and levels of satisfaction, which is bound to result in more personalized and patient-centered care.

Generative AI solutions are set to become a valuable research and scientific publishing tool that plays a key role in the process of reading, drafting, analyzing, and proofreading scientific papers.

Their participation is also crucial in the realm of specialized health training as a vehicle to evaluate the students' progression in real time. This should enable identification of areas for improvement, the offering of guidance, the development of personalized training programs, the prediction of future performance, and the detection of potential problems.48 Thanks to AI, training will become a fully personalized experience providing healthcare professionals with better skills and more in-depth knowledge, which will unquestionably result in improved patient outcomes.

AI is able to analyze large volumes of clinical and research data, to identify patterns, facilitate patient recruitment and accelerate the various processes involved in clinical trials.49 It will therefore facilitate the identification of potential subjects and the analysis of results, and it will contribute to the development of evidence-based practices, boosting the advancement of medical research and promoting research.

Methods such as DL are being applied to the reuse of already-approved medications for new uses and indications50 thanks to the analysis of genetic profiles, the search of new drug–protein or drug–receptor interactions, and the discovery of new molecular mechanisms, among others. At the same time, clinical researchers will use DL to discover new adverse events and drug–drug interactions, and to identify new phenotypes and response biomarkers.

However, in spite of AI's huge potential benefits, use cases in HP departments are still scarce. Most applications consist in support systems for the review of prescription orders and in the detection and/or prediction of adverse drug reactions.51–59Table 2 contains 6 real-use cases applicable to HP departments.

Hospital pharmacists with an understanding of AI will be able to contribute significant added value and face new challenges from an innovative perspective. Specifically, they will know how to:

• •

  Identify clinical, logistic, and operational challenges and problems that may be addressed by AI.

• •

  Lead the development and clinical validation of AI models.

• •

  Be active members of research teams working on the application of AI to healthcare.

• •

  Liaise between other clinicians and data scientists working on AI projects.

• •

  Critically evaluate drug-related models, algorithms and AI systems.

As conclusions, AI has become an invaluable tool for analyzing large volumes of data and obtaining results that may support decision-making, optimize human efforts, and make the best use of time and money, contributing to saving, and improving the quality of patients' lives.

The next generation of AI technologies applied to HP will be based on the creation of expert systems able to identify and prevent medication-related problems using the patient data stored in pharmaceutical databases and in external data repositories. Aided by advancements such as robotization and workflow improvements, AI technologies will make it possible for pharmacists to identify serious medication-related problems in a way that is closer to patients and to the clinical team.

AI is poised to play a decisive role in the transformation of some of the tasks performed by HP departments, such as dispensing and medication management, contributing to the delivery of a wider range of services that should result in more effective patient care.

Although the use cases of AI in the field of HP are still few and far between, it is to be expected that AI systems will play an increasingly important role in optimizing pharmaceutical care and improving safety and efficiency in the hospital setting. Further, and higher quality research is, however, required to examine the usefulness of AI in the HP setting to determine whether it can indeed become a viable tool at the service of HP departments.