An artificial intelligence

toolkit for pharmacy

An introduction and
resource guide for
pharmacists

March 2025
p2 | An artificial intelligence toolkit for pharmacy

Copyright 2025 International Pharmaceutical Federation (FIP)

International Pharmaceutical Federation (FIP)

Andries Bickerweg 5
2517 JP The Hague
The Netherlands
www.fip.org

All rights reserved. No part of this publication may be stored in any retrieval system or transcribed by any form or
means – electronic, mechanical, recording, or otherwise without citation of the source. FIP shall not be held liable for
any damages incurred resulting from the use of any data and information from this report. All measures have been
taken to ensure accuracy of the data and information presented in this report.

Editor:
Dr Whitley Yi Chair of the FIP Technology Advisory Group Artificial Intelligence Working Group

Co-editors:
Dr Ardalan Mirzaei, Asia Pacific Representative in the FIP Health and Medicines Information Section
Mr Brent Sin Hidge, FIP TAG Artificial Intelligence Working Group
Dr Mariana Guia, FIP TAG Artificial Intelligence Working Group
Dr Paul Voigt, FIP TAG Artificial Intelligence Working Group

Recommended citation:
International Pharmaceutical Federation (FIP). An Artificial Intelligence Toolkit for Pharmacy: An introduction and
resource guide for pharmacists. The Hague: International Pharmaceutical Federation; 2025
 An artificial intelligence toolkit for pharmacy| p3

Contents
Acknowledgements ................................................................................................................... 4
1 Executive summary ............................................................................................................. 5
2 Preface ............................................................................................................................... 6
2.1 Purpose of the toolkit ............................................................................................................... 6
2.2 AI & the FIP Development Goals................................................................................................ 7
3 Background ........................................................................................................................ 8
3.1 The fundamentals of AI ............................................................................................................. 8
3.1.1 Learning styles for AI ....................................................................................................................................9
3.1.2 Types of AI tools .........................................................................................................................................10
3.2 Assessing the performance of AI ............................................................................................. 11
3.2.1 Why measure performance? .....................................................................................................................11
3.2.2 How to be transparent, accountable, and build trust. ..............................................................................11
3.3 Performance metrics for classification models ........................................................................ 14
4 Ethical considerations ....................................................................................................... 16
4.1 Creating AI that is trustworthy ................................................................................................ 16
4.2 The ethical considerations of applying AI in healthcare ............................................................ 17
4.2.1 Ethical Issues ..............................................................................................................................................17
4.2.2 Mitigation strategy for the ethical deployment of AI in healthcare ..........................................................18
5 ISO regulations and compliance......................................................................................... 21
5.1 General AI regulation .............................................................................................................. 21
5.2 Data protection and privacy in general in healthcare ............................................................... 21
5.2.1 Human-centred AI ......................................................................................................................................22
5.2.2 Clinical validation and reproducibility of the algorithms ...........................................................................22
5.2.3 Interoperability standards..........................................................................................................................23
5.2.4 Cybersecurity measures .............................................................................................................................23
5.2.5 Accessibility and inclusivity ........................................................................................................................23
5.2.6 Considerations for the future.....................................................................................................................23
6 Common barriers to implementation................................................................................. 24
6.1 Algorithmic Bias...................................................................................................................... 24
6.2 Model Drift............................................................................................................................. 25
6.3 Overfitting .............................................................................................................................. 25
6.3.1 Additional barriers to keep in mind: ..........................................................................................................26
6.4 Selecting the right tool by establishing the pros and cons ........................................................ 26
6.4.1 Weighing the pros and cons between models...........................................................................................27
7 Implementation checklist .................................................................................................. 33
8 Training and development of competence in AI ................................................................. 34
8.1 What do pharmacists need to know about AI?......................................................................... 34
8.2 Knowledge and skills for AI ..................................................................................................... 35
9 References ........................................................................................................................ 36
Annex - Glossary of terms used in this toolkit .......................................................................... 39
p4 | An artificial intelligence toolkit for pharmacy

Acknowledgements
FIP expresses gratitude and appreciation to all the authors who co-developed the toolkit. The authors are listed below:
• Whitley Yi, PharmD, BCPS, FIP TAG Artificial Intelligence Working Group Chair, FIP Technology Advisory Group
Member, Director of Pharmacy and Member Services, Well, Adjunct Lecturer, University of Colorado Skaggs
School of Pharmacy and Pharmaceutical Sciences, USA
• Brent Sin Hidge, BPharm(NMU), FIP TAG Artificial Intelligence Working Group; Pharmaceutical Society of
South Africa National Executive Committee member; South African Association of Hospital and Institutional
Pharmacists Western Cape Chairman; Hospital Pharmacist, Netcare Blaauwberg Hospital, South Africa
• Bruno Macedo, FIP TAG Artificial Intelligence Working Group, CEO and Founder of MedFacts, Program
Manager, Calouste Gulbenkian Foundation, Portugal
• Claudia Rijcken, FIP TAG Artificial Intelligence Working Group, FIP Technology Advisory Group Member, CSO
and Founder of Pharmi, Lecturer at Utrecht University School of Pharmacy, The Netherlands
• Florencia Ojeda, Systems Engineer, MS in Artificial Intelligence, FIP TAG Artificial Intelligence Working Group
Member, Uruguay
• Joanna Klopotowska, FIP TAG Artificial Intelligence Working Group, Assistant Professor and Principal
Investigator, Amsterdam UMC Academic Medical Center, The Netherlands
• Mariana Guia, PharmD; FIP TAG Artificial Intelligence Working Group; Health Solutions Specialist, Portuguese
National Association of Pharmacies, Portugal
• Markus Manner, FIP TAG Artificial Intelligence Working Group, Development Manager, Suomen
Apteekkariliitto, Association of Finnish Pharmacies, Finland
• Martin Kondža, MPharm, PhD, FIP TAG Artificial Intelligence Working Group, Assistant Professor, University
of Mostar, Bosnia and Herzegovina / Pharmaceutical Chamber of the Federation of Bosnia and Herzegovina,
Bosnia and Herzegovina
• Mauro Tschanz, FIP TAG Artificial Intelligence Working Group, Digitalisation Expert at Swiss Pharmacy
Association (PharmaSuisse), Switzerland
• Mohd Syamir Mohamad Shukeri, B.Pharm., M. Comm. Health, R.Ph. MMPS, FIP TAG Artificial Intelligence
Working Group, Malaysian Pharmacist Society, Malaysia
• Paul Voigt, BPharm, MSc, DBiotech, ADHSML, FIP TAG Artificial Intelligence Working Group, Pharmaceutical
Society of South Africa, Inventory Operations Manager, Mediclinic Southern Africa
• Régis Vaillancourt, B.Pharm., PharmD, FFIP, FOPQ, FCSHP, FIP TAG Artificial Intelligence Working Group, Vice
president Pharmacy affairs BCE Pharma Inc, Canada
• Sangeetha Ramdave, FIP TAG Artificial Intelligence Working Group, Sessional Academic, Monash University,
Australia

FIP thanks the FIP Technology Forum chair Lars-Åke Söderlund and all FIP Technology Forum members for their support,
input, and feedback on the toolkit.
 An artificial intelligence toolkit for pharmacy| p5

1 Executive summary
The "Artificial intelligence toolkit for pharmacy" is a resource designed to assist pharmacists in integrating AI
technologies into their practice. Developed by the International Pharmaceutical Federation (FIP), this toolkit aims to
bridge the gap between the rapidly evolving field of AI and the practical needs of pharmacists.

Purpose and importance of AI in pharmacy: AI is revolutionising healthcare by enabling the analysis of vast amounts
of medical data, assisting in clinical decision-making, personalising patient treatments, predicting disease outbreaks,
and optimising operational workflows. Closely aligned with FIP Development Goal 20 (Digital Health), AI helps build a
digitally competent pharmacy workforce, improving clinical care, disease screening, health systems management, and
pharmaceutical research. Nonetheless, pharmacists face critical challenges, including data privacy, cybersecurity
threats, potential algorithm biases, and ethical concerns.

FIP's commitment and initiatives: FIP is committed to guiding and supporting pharmacists on the effective use of AI
technologies. In 2023, FIP established an AI working group to identify the needs of its constituencies around AI and
enhance inter-sectoral collaboration. This group aims to bridge knowledge gaps and promote the understanding and
use of AI technologies among pharmacists and pharmaceutical scientists.

Toolkit objectives: This toolkit provides a high-level guide for pharmacists, offering an overview of AI implementation
considerations, practical applications, and inspiring innovation. It aims to empower pharmacists to deliver safer, more
effective, and personalised patient care without undermining their critical thinking or professional judgment.

AI and FIP Development Goals: The toolkit supports the delivery of the FIP's Development Goals, and aligns with the
UN Sustainable Development Goals, focussing on education and continuous professional development (CPD), digital
literacy, and innovation in pharmacy practice. Specific objectives include developing educational resources for digital
competence, integrating AI training into professional education, and providing strategic guidance on digital health
workforce policies.

Challenges and considerations: Effective integration of AI in pharmacy requires pharmacists to understand AI's
capabilities and limitations, selecting appropriate tools tailored to specific challenges. Key considerations include
ensuring data quality, regulatory compliance, ethical frameworks, and infrastructure investment. Additionally,
addressing equitable access is vital to prevent exacerbating existing health disparities and inequities.

Conclusion: This toolkit aims to serve as a comprehensive resource, enabling pharmacists to navigate the complexities
of AI, capitalise on its potential, and clearly understand its limitations. By fostering innovation, collaboration, and a
proactive approach, FIP seeks to enhance AI's role in pharmacy practice, ultimately driving improvements in patient
care and global health outcomes.
p6 | An artificial intelligence toolkit for pharmacy

2 Preface

2.1 Purpose of the toolkit

Artificial intelligence (AI) is a branch of computer science dedicated to creating systems capable of performing tasks
that typically require human intelligence, such as learning, reasoning, problem-solving, and pattern recognition. AI
applications are increasingly utilized to analyse large volumes of medical data, assist clinical decision-making,
personalize treatment plans, predict disease outbreaks, and optimize operational workflows in healthcare settings.

Digital health is a priority for FIP, guiding pharmacists on how best to use digital advancements to support healthcare
delivery, pharmacy practice, and pharmaceutical research. AI is an emerging area within digital health that has
immense potential to improve efficiency in clinical care, disease screening and surveillance, clinical care, optimising
health systems management and facilitating advanced pharmaceutical research and development. Furthermore, AI
has the potential to promote equity in healthcare access globally, empowering patients to actively manage their
healthcare needs. However, even with the plethora of AI digital enablers entering the pharmacy sector, AI presents its
own challenges and poses risks in data privacy concerns, safety risks, encoded biases, and cybersecurity threats.

FIP acknowledges its responsibility to collaborate, guide, and educate pharmacists globally on effectively utilizing AI to
achieve optimal health outcomes. FIP is committed to fostering an environment where pharmacists can confidently
adopt these technologies. In 2023, with these goals in mind, FIP established an AI working group as a subgroup of the
FIP Technology Advisory Group, tasked with mapping AI-related needs of FIP constituencies and providing practical
resources and guidance.

This toolkit, developed by the AI Working Group, seeks to bridge the gap between the rapidly evolving field of AI and
the practical, everyday needs of pharmacists. It serves as a high-level guide for pharmacists in navigating the
complexities of AI applications by providing an overview of considerations for AI implementation. It aims to illustrate
practical applications and inspire innovation among FIP constituencies, ensuring they are well-equipped to navigate
the evolving AI landscape and leverage the benefits of AI, while understanding its limitations.

Ultimately, this toolkit aims to empower pharmacists and pharmacy teams to confidently integrate AI into their
practice in a way that complements their expertise, enabling them to deliver safer, more effective, efficient, and
personalized patient care while preserving pharmacists’ critical thinking and professional judgment.
 An artificial intelligence toolkit for pharmacy| p7

2.2 AI & the FIP Development Goals

Launched in September 2020, the FIP Development Goals seek to direct the transformation of the pharmacy
profession globally to 2030 (visit https://developmentgoals.fip.org/). Aligning with the UN Sustainable Development
Goals (SDGs), a wider global initiative, the FIP Development Goals specifically focus on enhancing pharmacy practice,
education, and pharmaceutical sciences. The ‘One FIP’ Development Goals enable the identification of commonalities
and inter-sectoral collaboration within a transformative framework for the pharmacy profession.

The FIP Development Goal on Digital Health (Development Goal 20) is structured around three elements: education
and workforce, practice, and science.

Education and workforce: ‘Enablers of digital transformation within the pharmacy workforce and effective processes
to facilitate the development of a digitally literate pharmacy workforce.’

In the context of AI, this toolkit aims to support the following Development Goal mechanisms:
• Develop courses and training for a digitally literate workforce (FIP Development Goals 1 & 2).
• Embed digital health and literacy competencies within FIP advanced and specialist frameworks (FIP
Development Goals 4 & 5).
• Provide multidisciplinary learning tools to enhance digital health literacy (FIP Development Goal 8).
• Promote the use and interpretation of AI in training and education of pharmacists and pharmaceutical
scientists. Create opportunities for ongoing education and development to ensure current practice with
technology advancements (FIP Development Goal 9).
• Provide guidance on incorporating AI into digital health workforce development policies, including
employment (FIP Development Goal 13).

Practice: ‘Systems and structures in place to develop and deliver quality digital health and pharmaceutical care
services through the digital literacy and utilisation of technology and digital enablers, configuration of responsive
digital services to widen access and equity’.

In the context of AI, this toolkit aims to support the following Development Goal mechanisms:
• Provide insight into digital health enablers and AI-driven technologies that support cutting-edge service
delivery and clinical decision-making (FIP Development Goal 20).
• Discuss digital literacy and governance concerning ownership, ethics, privacy, operational implications and
information quality. Provide guidance for policies to support and enhance health data management and
accountability for patient outcomes (FIP Development Goal 20).
• Promote AI-driven digital health initiatives that enhance equitable access to pharmaceutical care (FIP
Development Goal 20).

Science: ‘Application of digital technology in healthcare delivery and development of innovative medical products.’

In the context of AI, this toolkit aims to support the following Development Goal mechanism:
● Facilitate AI integration as a data science solution, enabling pharmacists to deliver enhanced patient care
through innovative healthcare technologies (FIP Development Goal 20).
p8 | An artificial intelligence toolkit for pharmacy

3 Background
The primary problem that informed the need for this AI toolkit is the integration of AI into daily pharmacy practice in
industry, hospital and community settings, aiming to optimise pharmacists' work processes, enhance patient care,
stimulate multidisciplinary interaction, and educate pharmacists on the efficient use of AI tools.

The integration of machine learning (ML) and artificial intelligence technologies is reshaping the roles and
responsibilities of healthcare professionals, including pharmacists. AI has the potential to augment pharmacists'
capabilities across a spectrum of tasks such as medication management (in both a logistical and pharmaceutical
context), patient counselling, drug interaction checks, personalised medicine, pharmaceutical research, and much
more.

Recent advancements in generative AI have further accelerated AI adoption, prompting the need for thoughtful and
responsible deployment. While generative AI holds significant promise for pharmaceutical care; it also presents
substantial limitations and challenges that must be addressed, to ensure effective and ethical integration.

A key significant concern is the quality and integrity of data used and produced by AI systems. Since AI outputs are
only as reliable as the data they learn from, inaccuracies, biases, or inconsistencies in training datasets can lead to
incorrect conclusions or predictions. This is particularly critical in pharmaceutical care where patient safety and
efficacy of medications are concerned.

Another critical issue is ‘’AI hallucination’’, where generative AI systems produce incorrect, nonsensical, or fabricated
outputs. This poses significant challenges in critical applications such as pharmaceutical care.1 For instance, AI might
produce plausible sounding, yet false medical advice or treatment recommendations, stemming from limitations in
training data, pattern recognition errors, or inference mistakes. Addressing these hallucinations involves improving
training data quality, implementing verification steps, and ensuring human oversight to maintain accuracy and
reliability. As pharmaceutical care providers increasingly integrate AI tools, understanding and mitigating AI
hallucinations is crucial to ensure safe and effective patient care.

In addition to technical safeguards and human oversight, regulatory considerations play a central role in the successful
integration of AI. Pharmaceutical care is highly regulated, and AI applications must navigate complex regulatory
requirements concerning drug development, clinical trials, and patient privacy. Compliance with these regulations on
both the local and international level is essential to ensure patient safety and trust in the new technologies being
implemented.

Furthermore, the implementation of AI in pharmaceutical care requires substantial investment in terms of

infrastructure, training, and management. Health care providers and organisations must be prepared to invest in
robust IT systems and staff training to handle these advanced technologies effectively.

Ethical considerations also play a critical role. The use of AI must align with ethical standards concerning patient
confidentiality and safety, consent, and the avoidance of bias in treatment recommendations. Ensuring that AI
systems are transparent and their workings understandable to regulators and practitioners is crucial for their
acceptance and trustworthiness.

Finally, while AI can enhance the personalisation and efficiency of care, there's a risk of deepening existing health
disparities if not carefully managed. Ensuring equitable access to the benefits of AI technology in pharmaceutical care
is essential to avoid exacerbating health inequalities.

This toolkit will dive further into these topics and provide a list of considerations to keep in mind when exploring ways
to leverage AI tools for implementation.

3.1 The fundamentals of AI

AI models are divided into two broad classifications: predictive and generative. Predictive models learn patterns in
data and apply those patterns to new data it has never seen before, i.e., to predict potential outcomes. Predictive AI
 An artificial intelligence toolkit for pharmacy| p9

has already found several use cases within pharmacy and in healthcare in general, including tablet recognition, drug
design, cancer diagnosis, and population health analysis. Some types of predictive models include:

1. Classification and labelling

2. Risk prediction and forecasting
3. Recommendation

Generative models learn patterns in data to generate new data. The output for generative AI can range from text to
spreadsheets to images to video. ChatGPT is a popular example of generative AI. While generative AI has significant
potential and is already being leveraged across multiple industries, it also poses significant limitations and risks that
must be addressed to ensure it is used in an ethical and responsible manner.

3.1.1 Learning styles for AI

Both predictive and generative models are driven by machine learning. Machine learning models that use neural
networks are considered ‘deep learning’. By its very nature, deep learning models function as ‘’black boxes’’, which
means that the way the model comes to its conclusions is unexplainable by the AI itself or its human user. This can
pose conceptual hurdles for regulatory bodies.

The method of learning applied by a particular model or tool will influence its functionality and utility. Not all learning
styles are appropriate for a given use. These learning methods are:

1. Supervised
2. Unsupervised
3. Reinforcement

Supervised learning involves training the model on pre-labelled data. Labelled data implies that the machine is
provided with a description of the data it is presented with, for example an image of a cat labelled as “CAT”. This
forms a library of reference material the machine uses when predicting outcomes or generating new data based on
unlabelled new data inputs. Supervised learning has been used in healthcare to develop models to predict the
presence of a disease, create risk scores or forecast the prognosis of a disease.2 A predictive AI engine utilising
supervised learning has also been applied to drug discovery, resulting in a novel, broad spectrum antibiotic known as
Halicin.3

A common challenge with supervised learning in healthcare is having enough labelled data. It takes time to create
labelled datasets. It tends to be easier to label things like medical images (i.e., x-rays or MRIs). However, it is more
difficult when labelling a medical event or outcome, such as an exacerbation of a chronic condition, which requires
domain expertise and making judgement calls. Not everyone may interpret the data in the same way, which in turn
impacts the ultimate output of the model. Privacy and security are also an issue in creating and sharing labelled
datasets.

Unsupervised learning involves training a model on unlabelled data. Deep learning is then used for the machine to
identify patterns, relationships or similarities in the data. Often these patterns are unidentifiable by humans reviewing
the same data. An example would be conducting a cluster analysis that groups patients into different phenotype
groups based on similarities in many variables or data points available for each patient. However, one must be careful
if using this learning method to make a predictive model; unsupervised learning models can also be prone to
nonsensical or biased outputs as they have not been trained on what an acceptable output would be.

Supervised and unsupervised learning are very powerful in identifying patterns, including text, chemical structures,
DNA sequences, images, and video and using them in making predictions. However, it is less powerful in dynamic
environments, where the sequence of a series of events or decisions is important to the outcome. For such purposes,
reinforcement learning may be more applicable.
p10 | An artificial intelligence toolkit for pharmacy

Reinforcement learning “rewards” the model for making correct decisions. The model learns to predict the next action
or set of actions that increase the likelihood of a delayed reward.4 The reward process needs to be programmed into
the model, so that after every decision the model makes, it receives either a positive reward or a negative reward.
This lets the model know if decisions are taken in the right direction. Two significant experiments were conducted
using an AI model known as AlphaZero to play the games of Go and chess.5 Using reinforcement learning, the model
learned what series of game moves were most likely to result in the delayed reward of winning the game. The model
developed novel strategies that had never occurred to humans in the history of the game of chess. At the time of
writing, no human has beaten AlphaZero. The concept of reinforcement learning can be applied to the medical field
where multiple decisions are made for a future distant reward, such as survival to hospital discharge or making
medication adjustments to achieve a target A1C.

3.1.2 Types of AI tools

AI is a broad technology category; however, all AI tools or applications can be broken down into three major
components: the input, the algorithm/model, and the output as outlined in Figure 1.

Figure 1: AI model components

Understanding each of these components is important to understanding how the technology can be leveraged in each
situation. Table 1 outlines the different types of AI tools available based on the input data the model uses.

Table 1: Types of AI tools based on input

Model input AI tool description

Vision These are models that use images as their input data. This includes diagnostic tools for examining medical
images such as chest X-rays or CT scans. An example includes a pill identification model able to identify
unlabelled pills and tablets with 85% accuracy.6

Text Models that use text as the input means that they can interact with natural speech or text. This can be in
the form of a chatbot (similar to tools like ChatGPT), or it could be a document classification model.

Voice These types of models take raw voice audio as the input. Examples include dictation programmes, smart
speakers or devices, or smart assistants on mobile devices. There are also AI-enabled clinical applications
that use voice markers to diagnose or to measure a patient’s stress levels.

Tabular Tabular data includes any type of data that could be organised or stored in a spreadsheet. Models that use
this type of data include risk predictor algorithms.

Multi-modal A multi-modal model indicates that the model can accept more than one type of input. For example, a
chatbot may respond to text, but also to an uploaded image.
 An artificial intelligence toolkit for pharmacy| p11

3.2 Assessing the performance of AI

Assessing the performance of machine learning models is crucial for validating the technical robustness and accuracy
of these models but also for ensuring transparency, accountability, and building trust among users.

Performance metrics serve as the foundational tools that guide developers and stakeholders in understanding how
well a model performs in various scenarios, identifying strengths and weaknesses, and making informed decisions
about model improvements and applications.

The following section summarises common methodologies and tools used to evaluate AI models and ensure
interpretability for transparency and understanding.

3.2.1 Why measure performance?

Measuring the performance of AI models is not merely a technical necessity but a foundational practice for
establishing transparency, trust, and accountability. Transparency involves providing clear and accessible information
about how models function and their performance, ensuring that users can understand and trust the AI's insights to
support decision making. Accountability ensures that developers and organisations are responsible for their models'
performance, the ethical implications, and the potential impact on patients. Model cards serve as an important tool to
help achieve this.

3.2.2 How to be transparent, accountable, and build trust

Model cards serve as comprehensive and structured guides with essential details about a machine learning model. 7
They provide a standardised framework for information about the model’s characteristics, performance, limitations,
and intended use cases, and allow a comparative analysis between different models. Google Research developed the
Model Card Toolkit (MCT) to increase transparency in machine learning. The MCT helps developers create model
cards, which provide essential information about a model's origins, intended use, and ethical considerations. By using
the MCT, developers can easily document their models and ensure they are used responsibly.8 Looking at a model
card should help one determine if a model is appropriate for a specific task or use. A typical model card can include
various components, as described in Table 2.

Table 2: Components of a model card

Component Detail

Description Brief description and unique identification of the model, including time of production and
model version, if applicable.

Purpose, users, and context Clear articulation of the model’s intended purpose, answering questions such as: for what
purpose the model was developed; for which purpose the model is not suitable to be used;
the potential users that can benefit; the context in which it should or should not be applied;
and potential limitations of generalisation for different purposes, users and contexts.

How to use Understandable guidelines written in simple language to ensure users can effectively deploy
the model. If possible, flowcharts and schemes should be provided to ease the user's
comprehension.

Performance metrics Metrics vary according to model type and purpose (e.g., generation, prediction, classification,
etc.). It is important that performance metrics are accompanied by an interpretation of what
they mean within the context of use. Whenever possible, examples should be provided to
facilitate the user’s understanding, particularly related to the uncertainty around the model’s
results. When applicable, the model’s performance across different population subgroups
should be listed to provide insight into potential biases that may exist.
p12 | An artificial intelligence toolkit for pharmacy

Training data Transparency of the training data is essential. This includes the source, quality, any
preprocessing steps applied and any specific groups or subgroups that may be over or
underrepresented. Understanding the dataset helps users gauge the model’s applicability and
potential biases.

Ethics Ethical considerations are paramount in AI. This includes addressing bias, fairness, and
privacy. Ensuring that models do not perpetuate biases, and that user data is handled
responsibly is critical for maintaining ethical standards.

Uncertainty Every model has limitations and uncertainties. Documenting these helps manage user
expectations and informs them of potential risks and areas where the model might
underperform.

Author, code, license, and Clearly stating the authorship, availability of the code, licensing terms, and ownership rights
ownership enhances transparency and encourages community contributions and improvements.

Contacts and resources Providing contact information and additional resources supports users in understanding and
effectively using the model. It also opens channels for feedback and improvement.

To better understand the specific metrics used to evaluate how well a model performs, Table 3 provides an overview
of the metrics which might be seen on a model card or provided by the model developers.

Table 3: Model metrics, their use, and examples

Metrics for measuring

Model type Description Example
performance

PREDICTIVE AI

Classification Accuracy Evaluates models that Example: A model classifies patient records
Precision label or categorise inputs into categories like prescription refill
Recall into classes. Metrics assess requests or appointment scheduling
F1 Score the correctness of enquiries.
Sensitivity predictions.
Specificity Applying performance metrics: High
ROC-AUC accuracy and precision indicate correct
Confusion Matrix categorisation, minimising errors. High
recall ensures important categories are not
For definitions of these missed. A high Area Under the Receiver
metrics, see next section. Operating Characteristic Curve (ROC-AUC)
combined with Confusion Matrix (see Table
4) suggests effective distinction between
categories.

Regression Mean Squared Error (MSE) Evaluates models Example: A dose prediction model for
R² predicting continuous personalised medicine.
Mean Absolute Error (MAE) values. Metrics measure
Root Mean Squared Error the deviation between Applying performance metrics: Low MSE
(RMSE) predicted and actual and RMSE indicate close alignment
values. between predicted and actual dosages,
reducing underdosing or overdosing risks. A
high R² suggests the model effectively
 An artificial intelligence toolkit for pharmacy| p13

captures the relationship between patient

characteristics and required dosage.

Time series Mean Absolute Percentage Evaluates accuracy and Example: A forecasting model predicts
Error (MAPE) error in future value future medication demand.
Root Mean Squared Error predictions based on past
(RMSE) data. Applying performance metrics: Low MAPE
and RMSE indicate accurate forecasts,
helping maintain optimal inventory levels.

GENERATIVE AI

Generation Large Language Models Measures similarity Example: A model generates medication
(LLM): between generated and counselling scripts tailored to patients.
reference content.
Perplexity (PPL) 9 Applying performance metrics: A high BLEU
score indicates that the scripts closely
Bilingual Evaluation match high-quality references prepared by
Understudy (BLEU)10, 11 clinical pharmacists, ensuring patients
receive accurate and comprehensible
General Language advice.
Understanding
Evaluation (GLUE)12

Recall-Oriented
Understudy for Gisting
Evaluation (ROUGE)13
p14 | An artificial intelligence toolkit for pharmacy

3.3 Performance metrics for classification models

It is beyond the scope of this toolkit to provide a comprehensive view of all performance metrics, so this section goes
into greater detail around some of the most common model performance metrics one might come across. Precision,
recall, F1 score, specificity, and sensitivity are frequently used metrics to assess a model's effectiveness for prediction
through classification.14 The ROC-AUC curve and Confusion Matrix are useful tools to understand a model's capacity to
distinguish between different classes on prediction. The following content explains the theoretical fundamentals of
these metrics.

● Precision, also known as positive predictive value, measures the accuracy of positive predictions made by the
model. It is particularly useful in scenarios where the cost of false positives is high. Precision answers the
question: "Of all the instances predicted as positive, how many are actually positive?"

The formula for precision is given by:

𝑇𝑃
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 =
𝑇𝑃 + 𝐹𝑃

𝑇𝑃 (True Positives) are the instances correctly predicted as positive.

𝐹𝑃 (False Positives) are the instances incorrectly predicted as positive.

● Recall, also known as sensitivity or true positive rate, measures the model's ability to identify all relevant
instances. It is crucial in situations where missing a positive instance is costly. Recall answers the question:
"Of all the actual positive instances, how many did the model correctly identify?"

The formula for recall is:

𝑇𝑃
𝑆𝑒𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦 (𝑅𝑒𝑐𝑎𝑙𝑙) =
𝑇𝑃 + 𝐹𝑁

𝑇𝑃 (True Positives) are the instances correctly predicted as positive.

𝐹𝑁 (False Negatives) are the instances incorrectly predicted as negative.

● F1 Score is the mean of precision and recall, providing a balance between the two metrics. It is particularly
useful when there is an uneven class distribution (i.e., one of the classes is rare) or when both false positives
and false negatives need to be minimised.

The formula for the F1 score is:

𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 𝑥 𝑅𝑒𝑐𝑎𝑙𝑙
𝐹1 𝑆𝑐𝑜𝑟𝑒 = 2 𝑥
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 + 𝑅𝑒𝑐𝑎𝑙𝑙

● Specificity, also known as true negative rate, measures the model's ability to identify negative instances
correctly. It is important in situations where the cost of false negatives is high. Specificity answers the
question: "Of all the actual negative instances, how many did the model correctly identify?"

The formula for specificity is:

𝑇𝑁
𝑆𝑝𝑒𝑐𝑖𝑓𝑖𝑐𝑡𝑦 =
𝑇𝑁 + 𝐹𝑃
 An artificial intelligence toolkit for pharmacy| p15

● ROC AUC Curve is a graphical representation of a classifier's performance. The Receiver Operating
Characteristic (ROC) curve plots the true positive rate (sensitivity) against the false positive rate (1-specificity)
at various threshold settings. The Area Under the Curve (AUC) provides an aggregate measure of the model's
performance across all thresholds. A model with an AUC of 1 is perfect, while an AUC of 0.5 suggests no
discriminative ability.

• Confusion Matrix is a table used to describe the performance of a classification model on a set of test data for
which the true values are known. It allows visualisation of the performance of an algorithm. The matrix is N x
N, where N is the number of classes. It provides insights into the types of errors being made by the classifier.

The confusion matrix in Table 4 has the following structure:

Table 4: Confusion matrix

Predicted positive Predicted negative

Actual positive True positive (TP) False negative (FN)

Actual negative False positive (FP) True negative (TN)

Assessing the performance of AI models through structured and transparent methods is crucial for fostering trust and
accountability. By prioritising transparency, ethical considerations, and involving end users in the assessment process,
AI developers can create models that are not only technically sound but also trustworthy and aligned with user needs.
p16 | An artificial intelligence toolkit for pharmacy

4 Ethical considerations

4.1 Creating AI that is trustworthy

Recent advancements in machine learning (ML) and artificial intelligence (AI) have transformed various industries,
including healthcare. While these technologies offer tremendous potential benefits, they also pose ethical dilemmas
that must be addressed to ensure responsible and equitable use. The “Ethics guidelines for trustworthy AI" by the
High-Level Expert Group on Artificial Intelligence (HLEG on AI), commissioned by the EU in 2019, provides
comprehensive guidelines for ensuring ethical and trustworthy development and deployment of AI.15 According to the
Guidelines, trustworthy AI should be:

1. Lawful – respecting all applicable laws and regulations;

2. Ethical – respecting ethical principles and values; and

3. Robust – from a technical perspective whilst also taking into account its social environment.

The guidelines outline seven key requirements that AI systems should meet to be considered trustworthy (see Table
5).

Table 5: Seven key requirements for trustworthy AI

Key Requirement Details

1. Human agency and oversight AI systems should support human autonomy and decision-making,
with humans ultimately accountable for AI outcomes.

2. Technical robustness and safety AI systems must be secure, reliable, and resilient throughout their
lifecycle to avoid unintended harm

3. Privacy and data governance AI systems should respect privacy, ensuring protection of personal
data and compliance with data protection regulations.

4. Transparency The processes and decisions made by AI systems should be

explainable, understandable, and accessible to users.

5. Diversity, non-discrimination, and fairness AI systems should be inclusive and fair, avoiding biases and
discrimination based on various attributes.

6. Societal and environmental well-being AI development and deployment should benefit society and the
environment, promoting sustainability and societal well-being.

7. Accountability Stakeholders involved in AI systems (developers, deployers, etc.)

should be accountable for their decisions and actions related to AI.
 An artificial intelligence toolkit for pharmacy| p17

4.2 The ethical considerations of applying AI in healthcare

The ethical implications of using ML and AI in healthcare are complex and multifaceted. While these technologies
offer transformative potential—such as improving patient safety—they also introduce ethical challenges such as bias
in patient care when influencing clinical decision-making.16

One example of a major ethical concern is “inconclusive evidence”, which highlights that algorithms are probabilistic
and never infallible.17 For example, a smartwatch might incorrectly diagnose an irregular heartbeat due to
inaccuracies in heart rate measurements. Such errors can arise when algorithms are calibrated to general population
norms—such as specific skin tones—rather than individual patient characteristics. This demonstrates how biases in AI
development can lead to disparities in patient care.

Additionally, responsible AI deployment in healthcare must address patient safety, privacy, transparency, fairness,
informed consent, and workforce implications, particularly within pharmacy practice.18 Ensuring ethical AI integration
requires careful consideration of the four foundational principles 19 of medical ethics:

• Autonomy – a patient’s right to self-determination;

• Beneficence – the responsibility to ‘do good’;
• Non-maleficence – the responsibility to ‘do no harm’; and
• Justice – to treat all equally and equitably.

By proactively addressing these ethical challenges, we can maximise the benefits of AI in healthcare while upholding
ethical standards and ensuring patient welfare.

The challenge of defining ethical AI: Despite widespread acknowledgment that AI should adhere to ethical principles,
there is ongoing debate regarding the practical definition of "ethical AI." This includes uncertainties around specific
ethical requirements, technical standards, and best practices for implementation. The World Health Organization
(WHO) has outlined key ethical principles for AI in healthcare, including: protecting autonomy; promoting human well-
being; ensuring transparency, explainability and intelligibility; fostering responsibility and accountability; ensuring
inclusiveness and equity; and, promoting responsive development.20 Similarly, a review of the global landscape of AI
ethics guidelines identified five converging ethical principles: transparency, justice and fairness, non-maleficence,
responsibility, and privacy.21 Taking these ethical principles into account, the following section elaborates on key
ethical issues that must be considered for the responsible implementation of AI in healthcare.

4.2.1 Ethical Issues

4.2.1.1 Privacy and security
● The use of ML and AI involves handling large volumes of sensitive patient data, raising concerns about privacy
breaches and data security.

● Methods for anonymisation and encryption of data must be robust to protect patient confidentiality.

● Researchers and end-users need to consider the confidentiality and privacy risks associated with the data
used.

● Privacy and security measures should address both training data fed into the machine and the outputs
resulting from the machine learning findings.

4.2.1.2 Transparency
● Machine learning models can be complex and difficult to interpret. Researchers and AI developers must
consider the transparency and explainability of their models.22

● ML and AI algorithms often operate as "black boxes," making it challenging to understand their decision-
making processes.
p18 | An artificial intelligence toolkit for pharmacy

● Ensuring that machine learning is transparent allows for better reproducibility and understanding of the
underlying processes.

● Ensuring transparency in algorithmic outputs and establishing accountability for algorithmic errors or biases
are critical ethical considerations.22

4.2.1.3 Accountability
● Throughout the machine learning process, accountability is paramount. Models should be used only for their
intended purposes, and stakeholders must be aware of their responsibilities.

● Researchers, AI developers, and end-users such as clinicians, should maintain accountability to prevent
unintended consequences or misuse of machine learning outcomes.

4.2.1.4 Bias and fairness

● ML models trained on biased datasets can perpetuate or amplify existing biases in healthcare, leading to
inequitable treatment.

● Bias can also be introduced through the model architecture itself and decisions made during the training
process.23

● Ethical ML practices require strategies to identify and mitigate bias in data and algorithms to ensure fairness
and equity in healthcare delivery.

4.2.1.5 Informed consent

● Patients may not fully understand the implications of AI-driven diagnostics or treatment recommendations.24

● Ensuring informed consent for AI-enabled interventions involves educating patients about AI's role and
limitations in healthcare decision-making.

4.2.1.6 Impact on healthcare professionals

● ML and AI technologies may alter the roles and responsibilities of healthcare professionals, potentially
leading to workforce displacement or deskilling.

● Ethical guidelines should address the ethical implications of technology-driven changes in professional
practices and job roles.

4.2.2 Mitigation strategy for the ethical deployment of AI in healthcare

To effectively address the ethical challenges associated with AI deployment within healthcare, stakeholders from both
the healthcare and technology sectors must collaborate to develop comprehensive guidelines and regulations. Table 6
outlines key mitigation strategies and actions for the ethical integration of AI, focusing on enhancing ethical decision-
making and actively engaging pharmacists and other stakeholders.

Table 6: Strategies for the ethical deployment of AI

Strategy Actions

Establish clear ethical ● Develop and disseminate ethical guidelines governing the development and deployment
guidelines of AI technologies in healthcare, addressing key issues such as bias, transparency, patient
privacy, and environmental considerations.
 An artificial intelligence toolkit for pharmacy| p19

Integrate ethics training ● Incorporate ethics training into educational curricula for healthcare professionals
into educational specialising in digital health, including AI.24
programmes
● Provide ongoing professional development opportunities focused on the ethical use of AI
in both clinical practice and science.

Implement robust ● Develop mechanisms for regular auditing and validation of AI algorithms to detect and
auditing and validation address biases and ensure fairness.
mechanisms
● Engage third-party auditors to provide independent assessments of AI systems.

Engage patients and ● Actively involve patients and communities in discussions about the role and impact of AI
communities on healthcare decision-making processes.

● Ensure transparency with patients about how their data is used and obtain explicit
consent.

Deploy an enhanced Comprehensive ● Involve a diverse range of stakeholders, including AI and ML

ethical decision-making stakeholder engagement developers, healthcare decision-makers, research ethics
framework committees, regulators, and others committed to promoting
equitable and responsible use of clinical ML technologies.

Integration of ethical ● Align the ethical design of AI and ML tools with established
norms norms within the healthcare environment, encompassing
principles of biomedical ethics, clinical research ethics, and
social justice.

Holistic ethical guidance ● Emphasise ethical considerations at each stage of AI and ML

across the AI and ML integration in healthcare, reflecting on the implications of
lifecycle these principles for fostering fairness in ML applications.

Empower pharmacists Assessing model ● Train pharmacists to interpret AI recommendations and

as end users transparency understand the underlying decision-making processes.25

● Ensure AI systems have comprehensive documentation

detailing their design, training data, and decision rationale
(see section on Assessing Model Performance).

Identifying and ● Regularly monitor AI outputs for signs of bias and report any
mitigating bias inconsistencies to relevant authorities or developers.

● Utilise feedback mechanisms to report biases and collaborate

with developers to improve model fairness.

Ensuring fairness and ● Advocate for equitable AI use across all patient
equity demographics, ensuring AI does not disproportionately
impact vulnerable populations.23

● Verify AI recommendations, particularly in complex cases, to

ensure they enhance patient care quality.

Ongoing education and ● Engage in continuous professional development to stay

training updated on AI advancements and ethical considerations.

● Reflect on the ethical implications of AI in everyday practice

and integrate these considerations into patient care.

● Create a culture of accountability among end-users, who have

a responsibility to prevent unintentional harms caused by the
models they use.
p20 | An artificial intelligence toolkit for pharmacy

Patient and community ● Communicate openly with patients about the role of AI in
engagement their care, explaining the rationale behind AI-driven
recommendations.

● Foster community dialogue addressing public concerns and

enhancing trust in AI technologies.

Proactive risk ● Develop strategies to anticipate and address potential ethical

management dilemmas, such as AI errors, hallucinations or unintended
consequences.

● Collaborate with other healthcare professionals to share best

practices and improve AI deployment in pharmacy.

By implementing the mitigation strategies in Table 6, pharmacists can play a crucial role in ensuring the ethical and
responsible use of AI in healthcare. Empowered with the right tools and knowledge, pharmacists can help mitigate risks
and maximise the benefits of AI for patient care, contributing to a more transparent, fair, and patient-centred approach,
fostering trust and accountability in AI applications.
 An artificial intelligence toolkit for pharmacy| p21

5 ISO regulations and compliance

When integrating AI technology into pharmacy practice, it's essential for pharmacists to navigate several critical areas
effectively to ensure compliance on important areas like patient data confidentiality, quality and algorithm
reproducibility, and ICT security. AI systems used in pharmacy practice must adhere to specific healthcare and IT-
related regulations which vary by country or region.

5.1 General AI regulation

AI regulations vary significantly across different global regions, reflecting each area's unique approach to balancing
innovation with oversight. Across the regions, there is a clear trend towards establishing frameworks that ensure AI is
used ethically and responsibly without stifling innovation. While Europe emphasises robust regulatory frameworks,
the Americas are characterised by a more decentralised and sector-specific approach. Asia-Pacific shows a mixed
strategy. Table 7 further describes the regulatory landscape across each region.

Table 7: Regional AI regulation

Region Regulation landscape

Europe The European Union (EU) has been proactive with its comprehensive "AI Act," which classifies AI applications
based on risk levels and sets stringent requirements for high-risk uses, including biometric surveillance and
subliminal manipulation. This legislation emphasises transparency, accountability, and fairness in AI use,
aiming to align with European values of human oversight and privacy. The UK’s strategy focuses on principles
of safety and transparency while emphasising avoidance of overly restrictive regulations to foster AI
development and promote innovation.26

Americas In the USA, there is not yet unified AI legislation; instead, the country employs various guidelines and
frameworks to manage AI applications, often focusing on sector-specific measures. Canada has introduced
the AI and Data Act (AIDA), promoting safety and responsible AI practices, and Brazil is actively developing
comprehensive AI legislation to regulate high-risk AI systems and ensure transparency and accountability in AI
deployments.

Asia-Pacific China has taken significant steps in AI regulation with specific laws for algorithmic recommendations and
deep synthesis technologies, aiming to ensure content integrity and fairness while maintaining control over AI
innovations. Japan relies more on non-binding guidelines and sector-specific rules rather than comprehensive
national legislation, promoting a flexible approach to support AI innovation. India and other Asian countries
are currently formulating its approach, signaling a shift towards more definitive regulations soon.

Africa Various countries are at different stages of developing national strategies and policies to manage and harness
the benefits of AI technology. Legislation and guidelines are still heavily reliant on the various data protection
laws. Nations like South Africa, Mauritius, Egypt, and Kenya are pioneering efforts to draft and implement AI
guidelines that address both the opportunities and ethical challenges posed by AI. These initiatives often
consider adapting global standards, such as the EU's AI Act, to local contexts, promoting innovation while
safeguarding ethical values and societal norms. The African Union is also playing a role, suggesting a move
towards a more unified continental strategy. However, challenges such as limited technological infrastructure
and the need for local AI expertise remain significant hurdles. These regional differences highlight the
importance of international collaboration and dialogue to harmonise AI governance, ensuring that global
standards can accommodate local needs while promoting safe and beneficial AI development globally.

5.2 Data protection and privacy in general in healthcare

Protecting the confidentiality, integrity, and availability of patient data is paramount. AI systems process vast amounts
of sensitive data, and ensuring this data is protected against unauthorised access, breaches, and leaks is a
fundamental ethical and legal requirement. Effective data protection measures help maintain patient trust, a
p22 | An artificial intelligence toolkit for pharmacy

cornerstone of healthcare provision. Failure to protect data can lead to financial losses, erosion of patient trust, and
severe regulatory penalties, including criminal prosecution in certain jurisdictions.

In the USA, HIPAA (Health Insurance Portability and Accountability Act) is a primary regulatory framework that
pharmacies must comply with to protect patient information.27 Pharmacies must ensure that all systems handling
patient data, including AI technologies, meet HIPAA standards to avoid penalties such as those enforced on entities
like CVS Pharmacy and Rite Aid for improper disposal of personal health information (PHI). Pharmacies in the USA
must adopt robust encryption for stored data and secure transmission protocols to meet HIPAA requirements.

In Europe, similar protections are provided under the General Data Protection Regulation (GDPR). 28 Under GDPR,
pharmacists must ensure the confidentiality of patient data with strict access controls, and monitor for unauthorised
access. Pharmacy owners should conduct regular reviews of data retention policies and ensure that staff understand
the importance of patient confidentiality.

The African union has adopted the Convention on Cyber Security and Personal Data Protection to facilitate alignment
of legislation across member states. This convention is not binding but serves as a guideline. Across the continent, 36
out of 55 countries have enacted data protection laws. South Africa has enacted the Protection of Personal
Information Act (POPIA) which applies to any processing of personal information. It imposes criminal consequences
for failing to adequately protect personal data. The act stipulates rights and responsibilities of data subjects and users,
security measures required and how information can be sent out of the country, which has implications for any
offshore data processing. Similarly, Egypt, Kenya, Nigeria, and Zimbabwe have implemented their respective Data
Protection Acts (DPA).

5.2.1 Human-centred AI
Pharmacists must maintain a human-centred approach, ensuring that AI supports but does not replace the personal
interaction and care that is critical in healthcare. Equity in healthcare access is another vital consideration;
pharmacists must be vigilant that AI tools do not inadvertently exacerbate disparities by favouring certain populations
over others. Continuous education on AI developments and ethical guidelines is essential, empowering pharmacists to
make informed decisions and advocate for the responsible use of technology. By balancing innovation with ethical
vigilance, pharmacists can harness AI to enhance patient care while upholding the profession's core values of trust,
confidentiality, and equity.

5.2.2 Clinical validation and reproducibility of the algorithms

Regulation around AI applications varies by country. Autonomous AI applications are those that operate without a
human to verify or check the output. Some countries require robust clinical validation for autonomous AI, meaning
that the AI application must be tested and proven to be safe and effective for its intended use. Regulation for these
types of applications are often included within medical device regulations. Applications that are not autonomous may
not require such strict regulation.

In Europe, the Medical Device Regulation (MDR) is crucial for pharmacists integrating AI technology into healthcare, as
it provides a comprehensive framework ensuring the safety, efficacy, and quality of medical devices, including AI-
driven tools.29 Under MDR, AI technologies classified as medical devices must meet stringent requirements for risk
management, clinical evaluation, and post-market surveillance. For pharmacists, adherence to MDR is essential to
ensure that AI tools used in patient care are reliable, secure, and comply with European standards. This regulation
also emphasises transparency and traceability, ensuring that AI systems are not only effective but also explainable and
accountable, ultimately protecting patient safety and enhancing trust in AI-driven healthcare solutions.

In the USA, AI technology in healthcare is primarily regulated by the Food and Drug Administration (FDA) under its
Digital Health and Software as a Medical Device (SaMD) frameworks.29, 30 The FDA evaluates AI-driven tools based on
their intended use, potential risk to patients, and whether they meet the criteria for a medical device. For AI
technologies classified as medical devices, the FDA requires a rigorous review process, including premarket approval
or clearance, to ensure safety, effectiveness, and quality. The agency also emphasises post-market surveillance and
the need for continuous learning and adaptation of AI systems, ensuring they maintain performance standards over
time. This regulatory approach aims to balance innovation with patient safety, providing a robust framework for the
integration of AI into healthcare. However, AI applications that are used only as clinical decision support are often not
required to be tested and proven safe and effective before use, because there is a human safety net. These AI
applications must still be responsibly deployed in a way that promotes safety and minimises risk, understanding that
 An artificial intelligence toolkit for pharmacy| p23

the burden of responsibility falls on the clinical end-user to ensure efficacy and safety. It is important to be able to
distinguish between AI applications that require validation and those that do not.

In Asia, the regulation of AI technology in healthcare varies across different countries, but several leading nations such
as Japan, China, and South Korea are developing comprehensive frameworks to govern the use of AI in medical
contexts. In Japan, the Pharmaceuticals and Medical Devices Agency (PMDA) oversees AI technologies that qualify as
medical devices, focusing on safety, efficacy, and post-market monitoring, similar to the FDA's approach. China has
implemented strict regulations through the National Medical Products Administration (NMPA), requiring rigorous
testing and approval for AI medical devices, with a strong emphasis on data security and patient privacy. South Korea,
under the Ministry of Food and Drug Safety (MFDS), also regulates AI in healthcare, mandating approval processes and
ensuring continuous monitoring of AI tools in the market. Overall, while regulatory frameworks in Asia are still
evolving, there is a clear trend toward aligning with international standards, ensuring that AI in healthcare is safe,
effective, and secure across the region.

5.2.3 Interoperability standards

Ensuring that AI systems in pharmacy practice adhere to interoperability standards like Health Level 7 (HL7) or Fast
Health Interoperability Resource (FHIR) is crucial for seamless data exchange across different healthcare systems.31
This compatibility helps in maintaining a unified and efficient healthcare information environment.

5.2.4 Cybersecurity measures

Pharmacies must implement strong cybersecurity measures to protect against threats. This includes employing
encryption, secure access controls, and regular security audits to safeguard patient information. Strategies like two-
factor authentication and robust password policies are recommended to enhance security. ISO27001 and ISO27799 to
protect health data are essential certifications to consider.

5.2.5 Accessibility and inclusivity

AI tools should be accessible to all users, including those with disabilities, and designed to prevent healthcare
disparities. This requires careful design and implementation practices to ensure that AI systems are usable by a
diverse patient population and do not inadvertently exclude any groups.

AI can be designed to emphasise inclusivity by ensuring that algorithms are trained on diverse and representative
datasets, which helps to mitigate biases and improves the system's performance across different demographic groups.
For example, in healthcare, AI can be developed to consider varying genetic backgrounds, socioeconomic factors, and
gender differences, ensuring that diagnostic tools and treatment recommendations are effective for all populations.
Additionally, AI interfaces can be designed to be accessible to individuals with disabilities, incorporating features like
voice recognition for those with limited mobility or visual impairments. Multilingual support and culturally sensitive
content are other ways AI can be tailored to serve a broader range of users, ensuring that technology benefits
everyone, regardless of their background or abilities.

By addressing these areas, pharmacists can effectively use AI to complement their expertise, enabling safer, more
effective, and personalised patient care. Each region may have specific regulatory requirements, and staying informed
through continuous education and monitoring of regulatory updates is essential for maintaining compliance and
ensuring the effective use of AI in pharmacy practice.

5.2.6 Considerations for the future

As pharmaceutical care providers increasingly implement AI technology, future trends in AI regulation to keep in mind
include the potential for stricter data privacy laws, particularly concerning patient data, and the establishment of
clearer guidelines on AI accountability and transparency. Regulators may require AI systems in healthcare to undergo
rigorous testing and validation processes to ensure safety and efficacy, similar to the approval process for new drugs.
Additionally, there could be a push for more robust frameworks that ensure AI-driven decisions remain explainable
and interpretable to healthcare providers and patients, to maintain trust and uphold ethical standards in patient care.
p24 | An artificial intelligence toolkit for pharmacy

6 Common barriers to implementation

While an AI model's accuracy and performance are crucial, its ultimate clinical efficacy is determined by its practical
application within the deployment environment. Even the most advanced AI models can fall short of their potential if
they are not effectively integrated into practice. The deployment of AI models presents various barriers and
challenges that must be addressed to ensure their successful implementation and impact.

6.1 Algorithmic bias

Bias in AI occurs when the algorithms produce systematic and unfair discrimination against certain individuals or
groups. This can stem from biased training data, flawed model design, data and concept drift, or improper
deployment. In the context of healthcare and pharmaceuticals, the consequences of such biases can be particularly
severe, affecting patient outcomes and treatment efficacy. 32, 33 Fortunately, there are strategies that can be deployed
to help address bias, as described in Table 8. While these strategies may not eliminate bias, they can help mitigate it.

Table 8: Key strategies to address algorithmic bias

Strategy Detail

Diverse data collection Ensuring that training data is representative of diverse populations can help mitigate bias.
This involves collecting data from various demographic, geographic, and socio-economic
backgrounds.

Algorithmic fairness Implementing fairness-aware algorithms that actively check for and correct bias during
model training can prevent biased outcomes. Given the various models of fairness, each
addresses different aspects of bias, such as demographic parity

For further learning, the Alan Turing Institute offers a course 34 on assessing and mitigating
bias and discrimination in AI, which provides practical techniques and comprehensive
insights into algorithmic fairness.

Continuous monitoring Regular audits of AI systems to detect and address bias are essential. This involves setting up
feedback loops where the system's outputs are continually assessed for fairness. There are
multiple definitions for fairness. Castelnovo et al.35 define fairness through the following
concepts:

● Individual fairness: Similar individuals should receive similar outcomes. This

includes concepts like Fairness Through Awareness (FTA) and Fairness Through
Unawareness (FTU), which deal with treating individuals similarly based on their
features without considering sensitive attributes like gender.

● Group fairness: This focuses on treating groups equally and includes:

○ Independence (demographic parity): Decisions should be independent of
sensitive attributes.
○ Separation (equality of odds): Ensuring equal error rates across groups.
○ Sufficiency (calibration): Ensuring predictions are equally accurate across
groups.

Transparency and explainability Making AI systems transparent and their decision-making processes explainable can help
identify and rectify bias. Stakeholders must understand how and why AI systems make
certain decisions. Although deep learning algorithms are intrinsically unexplainable, some
tools can be used to help predict how the model makes its decisions.

Add your own guardrails The case of generative AI is a different issue. Data that did not exist is generated in the
model’s outputs. Since these models are not trained to be factual, a major concern with
these models is hallucinations, where the model provides a response that is false or
inaccurate. Furthermore, these models are non-deterministic: when given a single prompt,
 An artificial intelligence toolkit for pharmacy| p25

there are many correct (and incorrect) answers. Prompt engineering plays a crucial role in
this case.

6.2 Model drift

Model drift, when the performance of an AI model degrades over time due to changes in underlying data patterns, is a
critical issue.36 If not appropriately managed, it can lead to the model slowly becoming more inaccurate over time.
Table 9 lists strategies for minimising the impact of model drift.

Table 9: Strategies to manage model drift

Strategy Detail

Regular retraining Periodically retraining models with new data ensures they remain accurate and relevant.

Performance monitoring Setting up continuous monitoring systems to track model performance in real-time allows for
early detection of drift. Metrics such as accuracy, precision, recall, and F1-score can be
monitored.

Version control Maintaining version control for models helps track changes and reverts to previous versions if
necessary. This also aids in understanding the evolution of the model's performance.

Anomaly detection: Implementing anomaly detection systems can alert stakeholders to unexpected changes in
model behavior, prompting timely interventions.

6.3 Overfitting
Overfitting occurs when a model learns the noise in the training data instead of the signal, leading to poor
generalisation on unseen data.37 This results in models performing very well during the training phase but performing
poorly when they are tested on any data outside the training dataset. To prevent overfitting, the strategies in Table 10
can be implemented.

Table 10: Strategies to reduce overfitting

Strategy Detail

Cross-validation There are techniques and best practices that can be used called cross-validation, which is a
technique to assess a model's accuracy by training and testing it on different subsets of data
to ensure it performs well on various data sets. An example of this is k-fold cross-validation.

Regularisation techniques These are techniques that penalise models for being too complex, promoting simplicity and
generalisation. Examples of regularisation methods include L1 (Lasso) and L2 (Ridge).

Pruning and early stopping Pruning decision trees (a type of machine learning model) and stopping the training process
early when performance on validation data starts to degrade are effective methods to combat
overfitting.

Ensemble methods: Ensemble methods involve using multiple techniques at once to improve generalisation and
reduce the likelihood of overfitting.
p26 | An artificial intelligence toolkit for pharmacy

6.3.1 Additional barriers to keep in mind

In addition to technical challenges, such as minimising bias, managing drift, and reducing overfitting, there are
significant administrative and human-related barriers to integrating AI into practice. These include navigating complex
clinical workflows and adequately training pharmacists on new tools and processes. Successful implementation also
depends on engaging end users and fostering their willingness to adopt the new technology. Table 11 outlines important
barriers that must be addressed to support successful AI integration.

Table 11: Additional challenges to implementation

Barriers Detail

Integrating with complex Integrating AI into complex pharmaceutical and clinical workflows involves several
workflows considerations. This is a major challenge to AI adoption and implementation.38

Interoperability Ensuring AI systems can seamlessly integrate with existing IT infrastructure and software
systems is crucial. This involves using standard protocols and APIs (application programming
interfaces) for communication. Also, having necessary resources (both human and economic)
is crucial.

User training and support Providing comprehensive training and support for end users is essential to facilitate smooth
adoption. This includes creating user-friendly interfaces and offering ongoing technical
support.

Workflow customisation AI solutions should be tailored to fit the specific needs and processes of the pharmacy,
pharmaceutical company, organisation, or individuals who are using it. Customisation ensures
that AI tools complement existing workflows rather than disrupting them.

Scalability AI systems should be designed to scale with the organisation's growth. This involves planning
for increased data volumes, additional functionalities, and expanded user bases.

Collaboration and Promoting collaboration between AI experts, pharmacists, and other stakeholders is vital.
communication Clear communication channels help ensure that AI solutions address real-world challenges
effectively. Both technical expertise and industry know-how are needed to integrate both
worlds. Adding AI to a company should be supported by a whole culture pushing past barriers
and complications.

High initial costs The development and deployment of AI solutions require substantial investment in
infrastructure, skilled personnel, and technology, which can be a deterrent for many
organisations.32, 38

Change management Integrating AI into existing workflows requires a cultural shift within organisations. Resistance
to change among staff and stakeholders can hinder the adoption process.32, 38

6.4 Selecting the right tool by establishing the pros and cons
Selecting the right tool is not always easy and often more than one AI solution could be effective. It may not always be
the functional capabilities of the tool that drive the decision, but other factors that come into play, such as the
technical environment in which it will be implemented, the cost of creating, building, or maintaining the model, the
level of transparency needed, regulations, or data access, among others.

When selecting the right tool, one must take the capabilities and limitations of the tool into account, but also the
implementation design. For example, building a generative AI tool to provide health information to patients, carries a
higher risk if the tool is patient-facing. However, if the model provides information to a clinician who first reviews it
 An artificial intelligence toolkit for pharmacy| p27

before it is provided to a patient, the risk is much lower. For the second scenario, the options of what tools can be
used are much broader because it may not be necessary to have all the guardrails in place right away that would be
needed for the first scenario.

Likewise, building a clinical decision support tool for drug dosing in a critical care environment involves different
considerations. Creating a model that recommends what dose a patient should receive each day is very different to
creating a model that alerts a clinician that the patient’s condition is changing (or is predicted to change soon) and
prompts the clinician to evaluate what dose adjustment is needed. Both scenarios would require different training
data to build the model because the output of each model would be different. For the first scenario, the output of the
model is a recommended dose (e.g., 5 mg). For the second scenario, the model output predicts whether or not a dose
adjustment will be required, which involves less risk because a clinician is still making the decision, but it is augmented
with AI. For the first scenario, a reinforcement learning model may be more appropriate, while a supervised learning
model may be more appropriate for the latter.

6.4.1 Weighing the pros and cons between models

When using deep learning AI tools, there are many benefits, but also general limitations to keep in mind, such as data
dependency and limited model interpretability (see Table 12). Diving deeper, it is essential to weigh the strengths and
weaknesses of different AI models based on their type, learning style, and input data. Figure 2 outlines commonly
used AI models, while Table 13 provides a high-level summary of the pros and cons to consider for each model type,
depending on how the models are implemented and the ultimate goal of deployment.

Table 12: General pros and cons of deep learning models

Pros Cons

Adaptability: One can improve the performance of the Data dependency: Requires large amounts of data to train
model, as it is exposed to more data. effectively and accurately.

Versatility: Applicable to a wide range of problems, from Opacity: Complex models like deep neural networks can be "black
image recognition to natural language processing boxes," making it hard to understand how decisions are made
 p28 | An artificial intelligence toolkit for pharmacy

Figure 2: Deep learning models

 An artificial intelligence toolkit for pharmacy| p29

Table 13: High-level summary of pros and cons to consider for different model types

Model Examples (not all-inclusive) Pros Cons Considerations

Generative Chatbot ● Does not require input data to ● Not trained to be accurate. ● Usually, it is not practical to develop in-
natural language be structured Prone to hallucinations. house. Often must use an existing
processing tools ● A computer foundation model.
(large language programme that ● Versatile ● Contextual limitations:
models) processes natural text Struggle with understanding ● Some models are proprietary while others
and simulates a ● A single model can be used for context, sarcasm, and are open source
conversation with a many different purposes nuanced language
human39 ● Performance varies across different
● Many products/resources do ● Bias: Can inadvertently learn models. Some are more accurate at
Content creation not require coding and are easy and propagate biases present certain skills than others (medical
to use in training data knowledge, mathematics, reasoning, drug
● Generates original information, etc.).
text based on a ● For general uses, many models ● Can struggle with domain-
prompt are ready to use “out of the specific tasks ● It is important to review how a model
box” (e.g. does not require performs against various benchmarks,
Document classification additional training or fine ● Model does not include any how it was trained, its training data cutoff,
tuning) information (e.g., current and sources/risk for bias. However, not all
● Labels a document events, new drugs, guideline models share this data.
based on the topic or ● Models can be connected to an updates) that occurred after
other feature external data source its training data knowledge ● Requires using techniques such as
cutoff date Retrieval Augmented Generation (RAG)
Summarisation for the model to provide real-time
information or answer questions about a
● Generates a summary specific document or internal knowledge
of a document or base.
prompt
 An artificial intelligence toolkit for pharmacy| p30

Model Examples (not all-inclusive) Pros Cons Considerations

Non-generative Classification: ● Developers have more control ● It takes longer to build; ● Can be built in-house
natural language over the model output requires AI expertise
processing tools ● Labelling the topic of ● Using pre-trained weights can help
a document ● Can be built with less data ● Not generalisable and improve accuracy
specific. Models created for
Data extraction: ● Does not require as much one use are not able to be ● Depending on the underlying model and
computing power (able to be used for other purposes use of neural networks or not, these
● Extracting a diagnosis run easier on individual devices) without retraining. models may be more comprehended than
or medication list other black box models that use neural
from a clinical note ● Effective for narrow use cases ● Often requires labeled data networks
for training, which can be
Predictive: resource intensive to build

● Using hospital clinical

notes to estimate the
risk of a patient
needing end-of-life
care

Multi-modal Chatbots: ● Can interact with models ● The same limitations for ● Not all models are truly multi-modal.
generative AI through multiple channels generative AI apply to multi- Some appear to be multi-modal by
models ● A computer modal models combining several different models (e.g.,
programme that ● Models can be designed to text-generator, image-generator,
simulates a allow for follow-up questions transcription model, etc.) together.
conversation with a and back-and-forth dialogue
human by processing (e.g., after asking a model to ● Considerations are similar to those for
text, images, audio look at an image, the user can generative natural language processing
and/or video ask follow-up questions based tools above
on the model’s
AI assistants: description/response)

● Multiple models
coordinating with
each other and/or
other tools to
complete a task
 An artificial intelligence toolkit for pharmacy| p31

Model Examples (not all-inclusive) Pros Cons Considerations

Supervised Diagnostic: ● Many AI solutions have ● Models are black boxes and ● If using a deep learning model for medical
learning model - demonstrated sufficient are not explainable imaging from a third-party vendor, check
image ● Diagnosing or accuracy, safety, and reliability if the model has been approved by the
recognition characterising disease as determined by regulatory ● Can often only be used with applicable drug or device regulatory body
features from medical bodies, such as the Food and images of the same size, type (e.g., FDA, USA; Health Sciences Authority,
imaging 40-42 Drug Administration (FDA)43 and resolution it was trained Singapore; South African Health Products
on Regulatory Authority (SAHPRA), South
Africa; INFARMED, Portugal; Medicines
● May be tied to a specific and Healthcare Products Regulatory
imaging device - if switching Agency (MHRA), United Kingdom of Great
devices, would require a new Britain and Northern Ireland)44
model

Supervised Risk predictions/forecasting: ● Many existing tools for ● Models are black boxes and ● Not ideal for “recommendation” systems -
learning model - implementation are not explainable it does not truly predict the best action,
tabular data ● Creating a risk score but predicts what actions were taken
for the likelihood of a ● Can achieve high accuracy and ● May not always perform historically (which may not always have
patient being re- performance better than traditional been correct)
admitted to hospital statistical methods (e.g.,
after discharge45 ● A significant amount of logistic regression) ● It can be difficult and resource intensive
healthcare data already exists to scale. Every new use case typically
● Predicting drug in tabular format ● Requires large amounts of requires a new model and training data
shortages46 data to see performance
● Can be fine-tuned based on gains over simpler statistical
local data methods

● Can be used to map existing ● Perpetuates pre-existing

data to a new metric or risk errors and bias in data
score (e.g., mapping pedometer
data to a walking stability score) ● Not generalisable
 An artificial intelligence toolkit for pharmacy| p32

Model Examples (not all-inclusive) Pros Cons Considerations

Reinforcement Recommendation systems: ● Can learn without explicit ● These models can often have ● Well-aligned to a healthcare environment
learning model supervision, making them unintended consequences if where the “reward” may be a long-term
● Optimises a sequence suitable for tasks where the reward function is not outcome, rather than a short-term
of actions/decisions annotated data is scarce or well defined benefit. For example, multiple treatment
to achieve a long- unavailable decisions are made during a long
term goal ● Black box hospitalisation stay with the goal of
● Designed to maximise a reward increasing the chance of the patient
that might be far in the future ● The model may make surviving to discharge (long-term goal)
(e.g., what series of actions unpredictable decisions to
should be taken in a game to “explore” other action ● These models should be integrated with
win) pathways, which could be knowledge-based systems to create safety
unsafe in a healthcare guardrails
● Excels in problems that require environment
a sequence of actions to ● Rigorous validation is needed to ensure
achieve a goal ● It is difficult to implement the right “reward” was chosen and that
guardrails in the model to the model isn’t accidentally optimising for
prevent it from making the wrong one
unsafe recommendations
 An artificial intelligence toolkit for pharmacy| p33

7 Implementation checklist
As a summary to the information in the AI toolkit, the following checklist serves as a list of questions to help guide
discussions when pursuing implementation of an AI tool or solution. It is not meant to serve as a comprehensive
guide, but rather it should be used as a starting point. Each work environment or setting will have unique
requirements that must be taken into consideration.

Checklist

Defining the use case

What problem is the AI tool meant to solve?

Can the problem be solved with a non-AI-enabled solution or tool?

Have all the appropriate stakeholders been engaged, including both decision-makers and end users?

How will this tool fit into the existing workflow or how will the workflow be adjusted based on the tool?

Model selection

Will the AI model be developed in-house or will a third-party vendor be used?

If building the model in-house, does sufficient training data exist?

If using a third-party foundation model, will it require fine-tuning (additional training) with local or domain-
specific data?

If using a third-party AI-powered product or software, does it need to be evaluated and approved as safe and
effective by a regulatory body (e.g., FDA cleared software as a medical device)? If so, has it been?

If using an existing model, how well does the model perform (see Table 3)? How does it compare with other
models or existing benchmarks?

Compliance

Will the model have access to or utilise protected health information?

Will using the model require data to be shared outside of the organisation? For example, does the model require
using an Application Programming Interface (API) or is the data shared with a cloud server external to the
organisation? If so, what limitations does this pose on what data can be included in the input of the model?

What compliance regulations must be followed (see section on Compliance and ISO regulation)?

Can the model be deployed locally?

Vendor selection

If using a third-party vendor, do they provide a model card (as described in Table 2) or details about their training
data and model performance (see metrics in Table 3)?

How often does the vendor audit model performance or retrain their model?

Does the vendor provide updated performance metrics after any model updates are deployed?

Safety

Have the potential failure modes of the model been outlined? What will the mitigation strategies be?

How will the model be audited on a continual basis?

How often does the model need to be re-validated based on potential drift?

Based on the model’s training data, is the model less accurate for specific subpopulations? How will this be
mitigated?
 An artificial intelligence toolkit for pharmacy| p34

8 Training and development of competence in AI

8.1 What do pharmacists need to know about AI?
For pharmacists and pharmacy teams in both hospital and community settings, the following competencies are
essential to harness the full potential of these technologies:

• Understanding of generative AI capabilities and limitations: Pharmacists must comprehend what generative
AI tools can and cannot do, including their scope, reliability, and the contexts in which they operate
optimally. This knowledge ensures appropriate reliance on AI for decision-making, avoiding over trust that
could lead to errors.

• Data literacy: The ability to interpret and evaluate data output from AI tools is crucial. Pharmacists need to
understand how to read, analyse, and make informed decisions based on the data generated by AI, which is
essential for accurate patient care and medication management.

• Ethical and legal considerations of AI use: Understanding the ethical implications and legal boundaries of
using AI in pharmacy practice is paramount. This includes patient privacy concerns, data security, and the
ethical use of AI in decision-making processes to ensure patient safety and compliance with regulations.

• Critical thinking and decision-making: While AI can provide recommendations, the ultimate decision-making
responsibility lies with the pharmacist. The ability to critically assess AI-generated advice, considering the
unique contexts and needs of each patient, is essential for effective pharmacy practice.

• Communication skills: Pharmacists must effectively communicate AI-generated information to patients and
healthcare teams. This includes translating complex AI data into understandable advice and ensuring that AI-
supported decisions are transparent and justifiable.

• Continuous learning and adaptability: The field of AI is rapidly evolving; therefore, pharmacists need to
commit to ongoing education and adaptation to new technologies. This continuous learning ensures that
pharmacy practice remains at the cutting edge, using the most current AI tools to improve patient care.

• Collaborative skills for interdisciplinary teams: Working with interdisciplinary teams, including IT
professionals, data scientists, and other healthcare providers, is crucial for implementing and optimising AI
tools in pharmacy practice. Effective collaboration ensures that AI implementations are well-coordinated and
meet the diverse needs of healthcare delivery.

• Patient-centred care: Pharmacists must ensure that AI tools are used in a way that prioritises patient needs
and outcomes. This involves using AI to personalise medication management and support, enhancing the
quality of care delivered to patients.

• Innovation and creativity: Finally, as in all ecosystems that work with AI, pharmacists should cultivate an
innovative mindset, seeking creative ways to apply AI in pharmacy practice. This includes developing new
workflows, patient care strategies, and management practices enhanced by AI, driving forward the pharmacy
field.

Each of these competencies is critical for pharmacists and pharmacy teams to effectively integrate generative AI tools
into their practice. Together, they enable the delivery of high-quality, efficient, and personalised patient care,
ensuring that pharmacy professionals remain at the forefront of healthcare innovation.
 An artificial intelligence toolkit for pharmacy| p35

8.2 Knowledge and skills for AI

Integrating generative AI tools into pharmacy practice represents a transformative shift towards more efficient and
effective healthcare delivery. In pharmacy practice, AI systems are increasingly used to optimise medication
management, enhance prescription accuracy, and streamline inventory control. For instance, AI-powered prescription
verification systems can automatically detect potential errors in medication orders, reducing the burden of manual
verification tasks for pharmacists. While this improves efficiency and patient safety, it also affects the traditional role
of pharmacists in overseeing medication dispensing processes and providing clinical care. Pharmacists need to adapt
to working alongside AI systems, requiring new skills in managing and interpreting AI-generated data.

Pharmacists eager to learn more about AI and its applications in healthcare and pharmacy practice have a variety of
learning environments at their disposal. Online courses and webinars offered by institutions such as Coursera, edX,
Udemy, and universities provide foundational knowledge on AI, its principles, and healthcare applications and are
easily accessible.

In Asia, Coursera China offers tailored courses in partnership with top Chinese universities, focusing on AI in
healthcare with content available in Mandarin. NUS-ISS (National University of Singapore, Institute of Systems
Science) provides specialised programmes on AI in healthcare, catering to professionals in Singapore and surrounding
regions. K-MOOC (Korean Massive Open Online Course), supported by the Ministry of Education in South Korea, offers
AI-related courses with a focus on applications in healthcare and pharmacy, available in Korean. Additionally, NTU’s
Nanyang Technological University in Singapore offers executive education and online learning modules on AI
applications in healthcare, targeting healthcare professionals across Asia. These platforms provide pharmacists with
region-specific insights and skills to effectively integrate AI into their practice.

Professional organisations, like the American Society of Health-System Pharmacists (ASHP) and the International
Pharmaceutical Federation (FIP), offer specialised training sessions, workshops, and conferences focused on AI in
pharmacy practice. Academic journals and publications often feature articles on AI research and case studies, keeping
pharmacists updated on the latest advancements and practical applications.

Collaboration with multidisciplinary teams, including data scientists and healthcare professionals through seminars
and hands-on projects, can provide practical insights and real-world experience. Additionally, participation in AI-
focused forums and discussion groups, both online and within professional networks, fosters an environment of
continuous learning and knowledge sharing.

These diverse educational resources can help equip pharmacists with the skills and understanding needed to
effectively integrate AI into their practice, to enhance patient care and optimise healthcare delivery.
 An artificial intelligence toolkit for pharmacy| p36

9 References
1. Templin T, Perez MW, Sylvia S, Leek J, Sinnott-Armstrong N. Addressing 6 challenges in generative AI for digital
health: A scoping review. PLOS Digital Health. 2024;3(5):e0000503. Available at:
https://pubmed.ncbi.nlm.nih.gov/38781686.
2. Jiang T, Gradus JL, Rosellini AJ. Supervised Machine Learning: A Brief Primer. Behav Ther. 2020;51(5):675-87.
Available at: https://pubmed.ncbi.nlm.nih.gov/32800297.
3. Stokes JM, Yang K, Swanson K, Jin W, Cubillos-Ruiz A, Donghia NM, et al. A Deep Learning Approach to Antibiotic
Discovery. Cell. 2020;180(4):688-702.e13. Available at: https://pubmed.ncbi.nlm.nih.gov/32084340.
4. Arulkumaran K, Deisenroth MP, Brundage M, Bharath AA. Deep Reinforcement Learning: A Brief Survey. IEEE
Signal Processing Magazine. 2017;34(6):26-38. Available at: https://ieeexplore.ieee.org/document/8103164
5. Silver D, Hubert T, Schrittwieser J, Antonoglou I, Lai M, Guez A, et al. Mastering chess and shogi by self-play with
a general reinforcement learning algorithm. arXiv preprint arXiv:171201815. 2017. Available at:
https://arxiv.org/abs/1712.01815.
6. Heo J, Kang Y, Lee S, Jeong D-H, Kim K-M. An Accurate Deep Learning–Based System for Automatic Pill
Identification: Model Development and Validation. J Med Internet Res. 2023;25:e41043. Available at:
https://pubmed.ncbi.nlm.nih.gov/36637893.
7. Mitchell M, Wu S, Zaldivar A, Barnes P, Vasserman L, Hutchinson B, et al., editors. Model Cards for Model
Reporting2019: ACM. Available at: https://dl.acm.org/doi/10.1145/3287560.3287596.
8. Fang H, Miao H, Shukla K, Nanas D, Xu C, Greer C, et al. Introducing the model card toolkit for easier model
transparency reporting. Google AI Blog. 2020. Available at: https://research.google/blog/introducing-the-model-
card-toolkit-for-easier-model-transparency-reporting.
9. Jelinek F, Mercer RL, Bahl LR, Baker JK. Perplexity—a measure of the difficulty of speech recognition tasks. The
Journal of the Acoustical Society of America. 1977;62(S1):S63-S. Available at:
https://pubs.aip.org/asa/jasa/article/62/S1/S63/642598/Perplexity-a-measure-of-the-difficulty-of-speech.
10. Lin C-Y, Och FJ, editors. Orange: a method for evaluating automatic evaluation metrics for machine translation.
COLING 2004: Proceedings of the 20th International Conference on Computational Linguistics; 2004. Available at:
https://aclanthology.org/C04-1072.
11. Papineni K, Roukos S, Ward T, Zhu W-J, editors. Bleu: a method for automatic evaluation of machine translation.
Proceedings of the 40th annual meeting of the Association for Computational Linguistics; 2002. Available at:
https://aclanthology.org/P02-1040.
12. Wang A. Glue: A multi-task benchmark and analysis platform for natural language understanding. arXiv preprint
arXiv:180407461. 2018. Available at: https://arxiv.org/abs/1804.07461.
13. Lin C-Y, editor ROUGE; A package for automatic evaluation of summaries. Text summarization branches out; 2004;
Barcelona, Spain: Association for Computational Linguistics. Available at: https://aclanthology.org/W04-1013.
14. Hossin M, Sulaiman MN. A review on evaluation metrics for data classification evaluations. International journal
of data mining & knowledge management process. 2015;5(2):1. Available at:
https://www.aircconline.com/ijdkp/V5N2/5215ijdkp01.pdf.
15. High-Level Expert Group on Artificial Intelligence. Ethics guidelines for trustworthy AI2019. Available at:
https://digital-strategy.ec.europa.eu/en/library/ethics-guidelines-trustworthy-ai.
16. Harrer S. Attention is not all you need: the complicated case of ethically using large language models in healthcare
and medicine. eBioMedicine. 2023;90:104512. Available at: https://pubmed.ncbi.nlm.nih.gov/36924620.
17. Morley J, Machado CC, Burr C, Cowls J, Joshi I, Taddeo M, et al. The ethics of AI in health care: a mapping review.
Social Science & Medicine. 2020;260:113172. Available at: https://pubmed.ncbi.nlm.nih.gov/32702587.
18. Mccradden M, Odusi O, Joshi S, Akrout I, Ndlovu K, Glocker B, et al. What's fair is… fair? Presenting JustEFAB, an
ethical framework for operationalizing medical ethics and social justice in the integration of clinical machine
learning: JustEFAB. Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency;
 An artificial intelligence toolkit for pharmacy| p37

Chicago, IL, USA: Association for Computing Machinery; 2023. p. 1505–19. Available at:
https://dl.acm.org/doi/10.1145/3593013.3594096.
19. Beauchamp TL, Childress JF. Principles of biomedical ethics. 7th ed. New York: Oxford University Press; 2013.
20. WHO Guidance. Ethics and governance of artificial intelligence for health. World Health Organization. 2021.
Available at: https://www.who.int/publications/i/item/9789240029200.
21. Jobin A, Ienca M, Vayena E. The global landscape of AI ethics guidelines. Nature machine intelligence.
2019;1(9):389-99. Available at: https://www.nature.com/articles/s42256-019-0088-2.
22. Abby Sen A, Joy J, Jennex M. Illuminating Ethical Dilemmas in AI-Driven Regulation of Healthcare Marketing on
Social Media. 2025. Available at: https://hdl.handle.net/10125/109472.
23. Leben D. Ethical issues for AI in medicine. Digital Health. 2025:309-19.
24. Hasan HE, Jaber D, Khabour OF, Alzoubi KH. Ethical considerations and concerns in the implementation of AI in
pharmacy practice: a cross-sectional study. BMC Medical Ethics. 2024;25(1):55. Available at:
https://bmcmedethics.biomedcentral.com/articles/10.1186/s12910-024-01062-8.
25. Hefti E, Xie Y, Engelen K. Machine learning model better identifies patients for pharmacist intervention to reduce
hospitalization risk in a large outpatient population. Journal of Medical Artificial Intelligence. 2025;8. Available at:
https://jmai.amegroups.org/article/view/9455/html.
26. Act EAI. The EU Artificial Intelligence Act. Retrieved May; 2024. Available at: https://artificialintelligenceact.eu.
27. Health Insurance Portability and Accountability Act of 1996, 104 (1996). Available at:
https://aspe.hhs.gov/reports/health-insurance-portability-accountability-act-1996.
28. Regulation (EU) 2016/679 of 27 April 2016 on the protection of natural persons with regard to the processing of
personal data and on the free movement of such data, and repealing Directive 95/46/EC (General Data Protection
Regulation), (2023). Available at: https://eur-lex.europa.eu/eli/reg/2016/679/oj/eng.
29. Baird P, Cobbaert K. Software as a medical device. A comparison of the EU’s approach with the US's approach.
Available at: https://pages.bsigroup.com/l/35972/2023-12-11/3t7658k.
30. Hermon R. Software as a Medical Device (SaMD): Useful or useless term? 2021. Available at:
https://researchnow.flinders.edu.au/en/publications/software-as-a-medical-device-samd-useful-or-useless-
term.
31. HL-7. HL7 2023 [Available at: https://www.hl7.org/.
32. Meskó B, Topol EJ. The imperative for regulatory oversight of large language models (or generative AI) in
healthcare. npj Digital Medicine. 2023;6(1):120. Available at: https://www.nature.com/articles/s41746-023-
00873-0.
33. Levi R, Gorenstein D. AI in medicine needs to be carefully deployed to counter bias–and not entrench it. NPR
Health Shots. 2023. Available at: https://www.npr.org/sections/health-shots/2023/06/06/1180314219/artificial-
intelligence-racial-bias-health-care.
34. Kohsiyama A, Lomas E, Polle R, Almeida D, Kazim E, Inness C. Assessing and Mitigating Bias and Discrimination in
AI: The Alan Turing Institute; 2024. Available at: https://www.turing.ac.uk/courses/assessing-and-mitigating-bias-
and-discrimination-ai.
35. Castelnovo A, Crupi R, Greco G, Regoli D, Penco IG, Cosentini AC. A clarification of the nuances in the fairness
metrics landscape. Scientific Reports. 2022;12(1):4209. Available at: https://www.nature.com/articles/s41598-
022-07939-1.
36. Holdsworth J, Belcic I, Stryker C. What is model drift? : IBM; 2024. Available at:
https://www.ibm.com/think/topics/model-drift.
37. AWS. What is Overfitting? : AWS Amazon; 2024. Available at: https://aws.amazon.com/what-is/overfitting/.
38. Ahmed MI, Spooner B, Isherwood J, Lane M, Orrock E, Dennison A. A Systematic Review of the Barriers to the
Implementation of Artificial Intelligence in Healthcare. Cureus. 2023;15(10):e46454. Available at:
https://pmc.ncbi.nlm.nih.gov/articles/PMC10623210.
 An artificial intelligence toolkit for pharmacy| p38

39. Adamopoulou E, Moussiades L. An Overview of Chatbot Technology. Artificial Intelligence Applications and
Innovations; Neos Marmaras, Greece: Springer, Cham; 2020. p. 373-83. Available at:
https://link.springer.com/chapter/10.1007/978-3-030-49186-4_31.
40. Kumar Y, Koul A, Singla R, Ijaz MF. Artificial intelligence in disease diagnosis: a systematic literature review,
synthesizing framework and future research agenda. J Ambient Intell Humaniz Comput. 2023;14(7):8459-86.
Available at: https://link.springer.com/article/10.1007/s12652-021-03612-z.
41. Mello-Thoms C, Mello CAB. Clinical applications of artificial intelligence in radiology. Br J Radiol.
2023;96(1150):20221031. Available at: https://pubmed.ncbi.nlm.nih.gov/37099398.
42. Shen J, Zhang CJP, Jiang B, Chen J, Song J, Liu Z, et al. Artificial Intelligence Versus Clinicians in Disease Diagnosis:
Systematic Review. JMIR Med Inform. 2019;7(3):e10010. Available at:
https://medinform.jmir.org/2019/3/e10010.
43. Benjamens S, Dhunnoo P, Meskó B. The state of artificial intelligence-based FDA-approved medical devices and
algorithms: an online database. npj Digital Medicine. 2020;3(1):118. Available at:
https://www.nature.com/articles/s41746-020-00324-0.
44. World Health Organization. Regulatory considerations on artificial intelligence for health: World Health
Organization; 2023. Available at: https://www.who.int/publications/i/item/9789240078871.
45. Davis S, Zhang J, Lee I, Rezaei M, Greiner R, McAlister FA, et al. Effective hospital readmission prediction models
using machine-learned features. BMC Health Services Research. 2022;22(1):1415. Available at:
https://bmchealthservres.biomedcentral.com/articles/10.1186/s12913-022-08748-y.
46. Pall R, Gauthier Y, Auer S, Mowaswes W. Predicting drug shortages using pharmacy data and machine learning.
Health Care Management Science. 2023;26(3):395-411. Available at:
https://ideas.repec.org/a/kap/hcarem/v26y2023i3d10.1007_s10729-022-09627-y.html.
 An artificial intelligence toolkit for pharmacy| p39

Annex - Glossary of terms used in this toolkit

Term Definition

Algorithm A sequential procedure or set of rules, designed to solve a particular problem or perform a specific task. It
is a clear, finite sequence of instructions that takes an input, processes it, and produces an output.

Application A set of guidelines and protocols that enable different software applications to communicate with one
Program Interface another. It outlines the methods and data formats that applications use to request and exchange
(API) information, allowing for smooth interaction. APIs are frequently used to integrate systems, services, or
applications, helping with tasks such as data retrieval, invoking services, and sharing functionalities.

Artificial A branch of computer science dedicated to creating systems capable of performing tasks that typically
Intelligence (AI) require human intelligence, such as learning, reasoning, problem-solving, and pattern recognition.

Artificial general A type of AI that has the human-like ability to understand, learn, and apply knowledge across a broad
intelligence (AGI) range of tasks, similar to human cognitive abilities. AGI can adapt to new challenges, reason through
problems, and transfer knowledge from one domain to another.

Artificial The use of advanced artificial intelligence techniques to analyse and interpret data related to
pharmacology pharmacology. It aims to enhance the understanding of drug interactions, drug discovery, and the effects
intelligence (API) of substances on biological systems.

Big data A vast collection of data that are too big for traditional data management systems to process. AI uses big
data to enhance model efficiencies in learning and analysis.

Data science An interdisciplinary field focused on extracting insights and knowledge from data through a variety of
techniques, such as statistical analysis, data mining, machine learning, and data visualisation.

Deep learning Involves the use of large, multi-layer artificial neural networks that process data using continuous (real
number) representations. It offers improved generalisation from small datasets and scales more efficiently
with large datasets and computing resources.

Deployment The process of integrating an AI model into a production environment to make predictions and analyses
based on data.

Generative AI Generative AI in pharmacy refers to the application of advanced AI techniques that can create new data,
(pharmacy) solutions, or insights based on existing information. These technologies leverage machine learning models,
particularly generative models, to enhance various aspects of pharmaceutical practice and research.

Large language LLMs are a specific type of generative AI model focused on producing human-like text. While generative AI
model (LLM) refers to a broad range of AI techniques and models designed to create new content—whether text,
images, audio, or video—LLMs specialise in text generation.

Machine learning A branch of AI that focuses on how computer agents can enhance their perception, knowledge, reasoning,
(ML) or actions through experience or data. Machine learning integrates concepts from fields such as computer
science, statistics, psychology, neuroscience, economics, and control theory.

Natural language A branch of artificial intelligence dedicated to developing systems that automatically create human-like
generation (NLG) text from structured data. NLG allows computers to generate written content, including reports,
summaries, or responses, in natural language. This process transforms data, facts, or insights into coherent
and meaningful sentences or paragraphs, improving how humans understand and interact with
information. NLG is widely applied in areas such as chatbots, automated reporting, content creation, and
personalised communication.
 An artificial intelligence toolkit for pharmacy| p40

Natural language A field of artificial intelligence focused on enabling computers to understand, interpret, and generate
processing (NLP) human language. It involves the development of algorithms and models that allow machines to process
and analyse text or speech data in a way that is meaningful and useful.

Neural network Computer models inspired by the human brain's structure. These interconnected artificial neurons,
organised in layers, learn from data to make predictions in machine learning, underpinning deep learning.

Prompt The input that a user puts into to an AI model to receive a specific response.

Retrieval An advanced technique in natural language processing that combines information retrieval with text
augmented generation.
generation (RAG)

Robotic process A rival technology that uses software robots to automate repetitive tasks based on fixed rules and inputs,
automation (RPA) often performing tasks more efficiently than humans.
International

Pharmaceutical

Federation

Fédération

Internationale

Pharmaceutique

Andries Bickerweg 5

2517 JP The Hague

The Netherlands

T +31 (0)70 302 19 70

F +31 (0)70 302 19 99

[email protected]

www.fip.org