The global healthcare industry is undergoing a fundamental transformation as it moves from a volume-based business to a value-based business. With increasing demand from consumers for enhanced healthcare quality, healthcare providers and insurers are under pressure to deliver better outcomes.

Primary care physician and nursing shortages require overworked professionals to be even more productive and efficient. The cost dynamics of healthcare are changing, driven by people living longer, the pervasiveness of chronic illnesses and infectious diseases, and defensive medicine practices. New market entrants and new approaches to healthcare delivery are also increasing complexity and competition. Extreme pressure to do more with less is making the case for a new approach to healthcare operations that focuses on the optimization of healthcare delivery and payment processes.

Prescriptive analytics enables healthcare decision-makers optimize business outcomes by recommending the best course of action for patients or providers. They also enable comparison of multiple “what if” scenarios to assess the impact of choosing one action over another.

Consider this hypothetical example: a health insurer spots a pattern in its claims data for the previous year showing a significant portion of its diabetic patient population also suffers from retinopathy. Using predictive analytics, the insurer estimates the probability of an increase in ophthalmology claims during the next plan year. Prescriptive analytics are then used to model out the cost impact if average ophthalmology reimbursement rates increase, decrease or remain the same for the next plan year, then recommend a course of action.

IBM Decision Optimization products help deliver prescriptive analytics capabilities to drive desired business results like cost reduction and customer satisfaction. Decision-makers can evaluate millions of possibilities, balancing trade-offs and business constraints to find the best possible solution. Together, these technologies can accelerate personalized medicine, aid in dynamic fraud detection and drive behavior modification for healthier lifestyle choices.

Improving care, reducing costs

When critical healthcare decisions are made on gut feeling or using simplistic tools, the results can be less than optimal and may even endanger patient lives. In this era of big data, any successful approach to transformation must be driven by real-time data analytics and enable decisions to be evidence-based and transparent. This is where technologies such as decision optimization, which enable a more evidence-based and transparent approach to decision making, come into play.

As health care administrators look for successful strategies to meet the challenges of an aging population, increased regulation, decreased budgets and other uncertainties, decision optimization will play a critical role in weathering the ups and downs of an evolving sector and in improving outcomes.

Hospitals with geographically distributed facilities, such as Dijon University Hospital Centre (CHU Dijon) in France, face a unique challenge: intra-hospital patient transport. Getting it right can be a logistical nightmare for dispatchers; getting it wrong can spell disaster for patients and caregivers. The hospital deployed a planning and dispatching solution that applies optimization models to ever-changing hospital and transport data, helping dispatchers plan, manage and execute hundreds of daily transport requests in real time. As a result, punctuality has improved by 25 percent, patient wait times are shorter and carriers walk 33 percent less per day.

Another example is radiation therapy, where the collateral damage done to surrounding organs and structures can be significant. There are virtually thousands of possibilities for the placement, diameter, duration and intensity of the beam. A German hospital applies mathematical optimization to design radiation treatment plans, enabling clinicians to precisely target which beams turn on, when and for how long to deliver an optimal dose for each patient.

A Danish hospital’s psychiatric department requires patients to be referred within 30 days for a diagnosis, but the process could take as long as two or three months. Using decision optimization, providers can now create mathematical models to simulate decisions and see how KPIs will change. The result: an 8 to 10 percent increase in referrals per year.

On the business and planning side, hospitals can use prescriptive analytics techniques to ensure accurate staffing levels, plan support facility location and capacity requirements, manage inventory and schedule home health services.

These examples demonstrate how health care organizations worldwide improve the quality of care, cut costs and increase transparency across all functional areas.

Getting started with decision optimization

These real-world examples illustrate how IBM Decision Optimization enables healthcare organizations to evaluate what is possible and what is not, given the available resources, to make reliable decisions. By developing and deploying optimization models using mathematical and constraint programming techniques, hospitals, insurers and others in the healthcare industry can optimize decisions and create real-world applications that significantly improve outcomes.

IBM Decision Optimization complements other IBM data science solutions such as machine learning, offering healthcare providers the tools to address their most complex challenges. It offers seamless integration between diverse technologies, enabling healthcare providers and payers to improve patient care, reduce costs and increase efficiency.

You can learn about benefits of optimization and explore more client stories by visiting the Decision Optimization web page or take this interactive product tour to see Decision Optimization in action.

Bringing artificial intelligence (AI) into clinical practice Heart doctors and scientists work together to bring the benefits AI to people with diseases of the heart and blood vessels.

Mayo Clinic is a leader in the movement to bring artificial intelligence (AI) tools and technology into clinical practice to benefit people who have or are at risk of heart disease. The clinic’s AI cardiology team is applying these new approaches to early risk prediction and diagnosis of serious or complex heart problems. People who receive heart care from Mayo Clinic’s Department of Cardiovascular Medicine may benefit from access to the clinic’s leading-edge research and expertise in AI cardiology to improve patient care.

AI is intelligence exhibited by machines. It touches almost every facet of modern life, including medicine. AI is being used at Mayo Clinic to program computers. The goal is to process and respond to data quickly and consistently for better treatment outcomes.

Uses for AI include detecting heart disease, treating strokes faster and enhancing diagnostic radiology capabilities. For example, a Mayo Clinic study applied AI techniques to a new screening tool for people with a certain type of heart problem that has no obvious symptoms. The condition is called left ventricular dysfunction. The AI -assisted screening tool found people at risk of this condition 93% of the time. To put that in perspective, a mammogram is accurate 85% of the time. In addition, AI developed at Mayo Clinic is used in Apple Watch to detect a weak heart pump (low ventricular ejection fraction).

These technologies complement the knowledge of doctors. Ideally, by bringing together direct care and data analysis, AI cardiology allows doctors to spend more time with their patients and improves the shared decision-making process.

Some basics

Artificial intelligence is the ability to make computers or machines learn to solve problems that would otherwise require human effort. Advances in computing power have made it possible to analyze large amounts of data quickly with consistency and accuracy. This has enabled health care scientists to apply AI to huge, complex data sets in a way that improves decision-making, diagnosis and treatment by detecting patterns in patient data.

The basic building block of an AI system is a “neural network.” For example, a computer system is trained by ingesting and analyzing hundreds of thousands of sets of similar readings. It becomes experienced in looking at a focused problem, such as ECGs . The result is that an AI system can read a simple test, detect a heart condition and predict possible future problems.

Mayo Clinic leaders have identified several areas of opportunity for AI in health care. The clinic is well situated to advance AI because its long history of high-volume patient care has generated a massive database of historical genomes, microbiomes, ECGs , diagnostic images and other test results. That coupled with the clinic’s strong culture of close collaboration among medical doctors, engineers and scientists is driving AI into health care in meaningful ways.

You can find a deeper look at computer neural networks and deep learning (strong AI ) here.

From research to clinical practice

Cardiovascular medicine doctors and scientists at Mayo Clinic are combining AI with clinical practice for better care. Here are three examples that have moved from the research stage to use in the clinic:

Helping people who have had a stroke. In emergency rooms, when people come in with a stroke called an intracerebral hemorrhage, they get a CT scan. That scan is examined by a computer trained to analyze CT data. This method has been shown to cut the time to diagnosis and limit brain damage.

In emergency rooms, when people come in with a stroke called an intracerebral hemorrhage, they get a scan. That scan is examined by a computer trained to analyze data. This method has been shown to cut the time to diagnosis and limit brain damage. Preventing heart problems. Applying AI to ECGs has resulted in a low-cost test that can be widely used to detect the presence of a weak heart pump. A weak heart pump can lead to heart failure if left untreated. Mayo Clinic is well situated to advance this use of AI because it has a database of more than 7 million ECGs . First, all identifying patient information is removed to protect privacy. Then this data can be mined to accurately and quickly predict heart failure.

Applying to has resulted in a low-cost test that can be widely used to detect the presence of a weak heart pump. A weak heart pump can lead to heart failure if left untreated. Mayo Clinic is well situated to advance this use of because it has a database of more than 7 million . First, all identifying patient information is removed to protect privacy. Then this data can be mined to accurately and quickly predict heart failure. Detecting atrial fibrillation (AFib) sooner. AI -guided ECGs also are used to detect faulty heart rhythms before any symptoms are evident. A faulty heart rhythm also is called atrial fibrillation.

Innovation through collaboration

A team approach A doctor (right) works with an Electrophysiology Laboratory colleague to read a test result.

The collective effort of experts is driving the rapidly growing field of artificial intelligence in health care. At Mayo Clinic, several medical and surgical specialties have validated approaches to improve clinical care. These groups include cardiovascular medicine, neurology, oncology and radiology, Their advances are shared in the medical literature so that they can be adopted widely to benefit people everywhere.

These AI tools and techniques also play an important role in education. They are used by Mayo Clinic’s medical students, residents, fellows and experienced surgeons to learn new or uncommon procedures. Mayo Clinic leads by holding artificial intelligence symposiums that bring together doctors and scientists to advance this science in health care.

Research innovations in cardiovascular artificial intelligence

The Mayo Clinic cardiovascular medicine team was among the first specialties to rapidly develop and validate these new AI tools and technologies. Possible future uses still in development at Mayo Clinic include:

Predicting risk early in conditions such as embolic stroke.

Monitoring the heart and detecting arrhythmia in smart clothing projects.

Developing AI technology compatible with smartphones and high-tech stethoscopes.

Mayo Clinic physicians, scientists and engineers continually advance the study and practice of artificial intelligence that improves health care. Find more about artificial intelligence at Mayo Clinic here.

See a list of publications about cardiovascular AI by Mayo Clinic authors on PubMed, a service of the National Library of Medicine.

Research profiles

Nationally recognized expertise

Mayo Clinic is top-ranked in more specialties than any other hospital and has been recognized as an Honor Roll member according to the U.S. News & World Report’s 2024-2025 “Best Hospitals” rankings.

Mayo Clinic campuses are nationally recognized for expertise in cardiology and cardiovascular surgery:

Mayo Clinic in Rochester, Minnesota, Mayo Clinic in Phoenix/Scottsdale, Arizona, and Mayo Clinic in Jacksonville, Florida, are ranked among the Best Hospitals for heart and heart surgery by U.S. News & World Report.

Mayo Clinic Children’s in Rochester is ranked the No. 1 hospital in Minnesota, and the five-state region of Iowa, Minnesota, North Dakota, South Dakota and Wisconsin, according to U.S. News & World Report’s 2024–2025 “Best Children’s Hospitals” rankings.

Conclusion: The value of this study is the identified processual phases and activities specific and typical to AI implementations to be carried out by an adopting healthcare organization when AI systems are deployed. The study advances previous research by outlining the types of necessary comprehensive assessments and legal preparations located in the implementation planning phase. It also extends prior understanding of what the staff’s training should focus on throughout different phases of implementation. Finally, the overall processual, phased structure is discussed in order to incorporate activities that lead to a successful deployment of AI systems in healthcare.

Results: The data were deductively mapped onto the steps of QIF using direct qualitative content analysis. All the phases and steps in QIF are relevant to AI implementation in healthcare, but there are specificities in the context of AI that require incorporation of additional activities and phases. To effectively support the AI implementations, the process frameworks should include a dedicated phase to implementation with specific activities that occur after planning, ensuring a smooth transition from AI’s design to deployment, and a phase focused on governance and sustainability, aimed at maintaining the AI’s long-term impact. The component of continuous engagement of diverse stakeholders should be incorporated throughout the lifecycle of the AI implementation.

Methods: This study employed a qualitative research design and included three components: (1) a review of 30 scientific articles describing differences empirical cases of real-world AI implementation in healthcare, (2) analysis of qualitative interviews with healthcare representatives possessing first-hand experience in planning, running, and sustaining AI implementation projects, (3) analysis of qualitative interviews with members of the research group´s network and purposively sampled for their AI literacy and academic, technical or managerial leadership roles.

Introduction: Absence of structured guidelines to navigate the complexities of implementing AI-based applications in healthcare is recognized by clinicians, healthcare leaders, and policy makers. AI implementation presents challenges beyond the technology development which necessitates standardized approaches to implementation. This study aims to explore the activities typical to implementation of AI-based systems to develop an AI implementation process framework intended to guide healthcare professionals. The Quality Implementation Framework (QIF) was considered as an initial reference framework.

1 Introduction

Artificial intelligence (AI) applications are increasingly being developed and tested in healthcare settings, enabled by enhanced analytical capabilities and advancements in computational technologies. These advancements allow for the processing of complex datasets and support from sophisticated AI-based applications for improvements in healthcare processes and health outcomes (1, 2).

However, previous research has highlighted the complexity of implementing AI in healthcare (3). The deployment of AI in healthcare is inherently more complex compared to sectors like manufacturing, retail, and finance, where standardized processes, lower regulatory constraints, and reduced risks to human life enable more straightforward implementation (4). Additionally, industries outside healthcare benefit from relatively easier access to large, high-quality datasets for training AI models, further facilitating adoption.

Healthcare, on the other hand, involves intricate interactions between human providers, patients, and technology, where the risks are exceptionally high, and any compromise in patient safety or quality of care must be rigorously avoided. The integration of AI introduces additional challenges and uncertainties related to clinicians’ liability in using the AI model outcomes in decision-making or neglecting them, autonomy of the patient and clinician, explainability, ethics, and the need for oversight and quality control (5–9). Compared to other sectors, healthcare faces structural barriers to technological adoption, encompassing technological, individual, social, and organizational domains (6, 10–12). For AI specifically, these barriers encompass challenges related to data, methodologies, technology, regulations and policies, human factors, environmental conditions, and organizational structures (6–9). For example, data quality is often compromised for the sake of faster workflows creating challenges in obtaining high-quality data for AI model training. Consequently, clinicians and patients may lack trust in the AI models for decision-making, a situation further burdened by unclear boundaries of liability. Also, organizational structures and different data storage practices lead to difficulties in combining the data.

Clinicians have recognized a lack of structured guidelines to navigate these complexities, ensuring that AI-based applications are implemented in a more predictable manner that enhances, rather than hinders, clinical workflows and patient outcomes (13, 14). Healthcare leaders emphasize the importance of considering implementation early in the development process, moving beyond the technology development and recognizing the significant challenges associated with practical implementation (8). Policymakers are also understanding the need for comprehensive guidelines to ensure that AI implementations comply with stringent regulatory (15) and ethical standards (16).

However, research indicates that the implementation of AI-based applications in healthcare has often lacked standardized and structured methodologies (2, 17). Despite growing attention to the challenges associated with implementation, there remains a significant need for empirical, experience-based research to identify strategies, processes, and methods that can effectively address these barriers (18, 19). Such research has the potential to inform the development of guidelines to support the integration of AI in healthcare.

Several frameworks have emerged for research and development purposes specifically addressing AI, such as the research-based evaluation framework for AI-based decision-support systems (20), the reporting guidelines for clinical trial reports for interventions involving AI (21), the provisional AI implementation framework (19), International consensus guideline for trustworthy and deployable AI in healthcare (22), the adoption of AI in the Healthcare Industry Model (23), and the organizational governance framework for AI adoption (24). Several studies have addressed elements crucial to implementation such as managerial, cultural, individual, and technological challenges, with some insights into the implementation process (2, 9, 17). However, studies have not been sufficiently detailed to provide sufficient processual support that could guide AI implementations to facilitate their integration and use in practice.

For a more standardized approach to AI implementation, implementation science emphasizes the importance of comprehensive planning from the outset (25). Such planning involves creating conditions that are favorable for implementation and for carrying out the process in a structured, well-planned, and orderly manner. Process frameworks are used in implementation science to provide guidance for implementation by outlining key considerations and activities that need to be undertaken before, during and after the implementation process (26).

The Quality Implementation Framework (QIF) (27) is a widely recognized process framework within implementation science, known for its adaptability and applicability across diverse contexts, including healthcare (28–31). Aiming at increasing quality of implementations, QIF’s creators Meyers et al. Have built the QIF on 25 theories, models, and frameworks across multiple research and practice areas. The resulting QIF framework contained 14 activities clustered into a 4-phase temporal sequence (27). Its structured and comprehensive approach offers detailed guidance on key activities and considerations necessary for successful implementation, making it a promising candidate for structuring the implementation of AI-based applications in healthcare. Unlike many other implementation process frameworks, which often provide more generalized guidance on translating research into action, the QIF stands out for its practical specificity and focus on actionable steps (28). The structure of the QIF includes four phases with detailed activities in each. The first phase recommends considering the host setting and examining how the innovation and the context fit each other. The second phase is focused on structuring the implementation process and assigning personnel. The third phase guides towards a necessity of support and feedback mechanisms for the implementation process. The final fourth phase is dedicated to learning and reflection to create capabilities leading to improved future implementations in the organization.

This study constitutes an integral part of a broader project (18) aimed to develop, test and evaluate a framework to guide the implementation of AI-based applications in healthcare. This paper focuses on the framework development part (32), with the specific aim to explore the activities typical to implementation of AI-based systems for developing an AI implementation process framework intended to guide healthcare professionals, leaders, and decision makers who plan, lead, or are involved in the implementation processes. To achieve this aim, the QIF was considered as an initial reference framework.

2 Materials and method

2.1 Design

This study employed a qualitative research design (33), utilizing two complementary methods for data collection and analysis: a review of scientific articles and analysis of qualitative interviews (Figure 1). The literature review on multiple examples of real-world AI implementations in healthcare practice provided a broad overview of process steps and lessons that comprised AI implementation within different types of AI systems, clinical areas and geographies as well as the lifecycle of AI systems—from development to deployment and adoption. The qualitative interviews had a two-fold goal: first, to collect direct individual reflections on experiences in the AI implementation process, and second, to reflect on what structure and content of the QIF may contribute to inform a future process framework specifically adapted for the implementation of AI in healthcare. The integration of these methods and diverse data sources enabled data triangulation, which increased the quality and credibility of research findings (34) and provided a more rounded understanding of how AI implementations are and should be conducted in practice. The knowledge generated from this study will inform future methodological developments to support AI implementations. The interviews performed in this study did not include sensitive personal information of the participants, and therefore, according to the Swedish legislation on research ethics, an ethical approval was not required.

Figure 1

Figure 1. Study design.

2.2 Sample and data collection

2.2.1 Literature review

The literature review started with identifying the scoping reviews and systematic literature reviews (in PubMed) on empirical studies of AI implementation in healthcare. A set of references to the empirical studies was extracted from these reviews resulting in 38 articles published between 2011 and 2023, covering various countries of origin [the articles’ search strategy is described in (9)]. Earlier articles were excluded from the scope of the present study due to the rapid pace of technological advancement and the aim to focus on contemporary AI solutions. The selected articles were reviewed by title, abstract, and full text when needed to identify ones that contained implementation process data and resulted in 30 articles included. The remaining 8 articles had no focus on the implementation process and were excluded from this review.

Most of the articles included represented AI systems developed, tested, and gone through real-world implementations within healthcare organizations discussing process steps and challenges during the different stages of AI system’s lifecycle, from development to adoption. Of the total of 30 articles, 14 articles (47%) examined AI applications in treatment, 13 articles (43%) explored AI applications in diagnosis, and 3 articles (10%) investigated AI applications in prevention (Supplementary Material S1). For simplicity, the categorization of the articles was made mutually exclusive, with each article classified under the most predominant aspect as determined by the author’s understanding. However, some articles may overlap and fit into multiple categories. The data extracted from these included articles were process steps, activities, involved stakeholders, and lessons learned.

2.2.2 Interview study I

Seven interviews were conducted between September–November 2023 with experts with the aim of collecting their perceptions of the critical activities in the AI implementation process and appropriateness of QIF for use in real world practice during AI implementations in healthcare (Table 1). The participants were explained the study purpose and the procedure, and that the project does not handle any sensitive personal data. They were also informed about the possibility to access their anonymized data and provide corrections, if needed. Then, the participants provided verbal consent to participate and to express personal experiences and perceptions. The interview questions related to how can a detailed guide support real-world implementations of AI in healthcare, and their impressions of the QIF in connection with the implementation of AI, its potential benefits and disadvantages, and the level of detail (the interview guide is provided in Supplementary Appendix S2). Participants were members of the research group´s network and purposively sampled for their AI literacy and academic, technical or managerial leadership roles. The data was collected via video communications (averaging 30 min, with a range of 10–50 min) and e-mail conversations, that were sometimes repeated to ask for more detail or to clarify written feedback.

Table 1

Table 1. Participants in the interview study I.

2.2.3 Interview study II

Given the focus on understanding the process of real-world AI implementation in practice, and aiming to deep-dive into how the sustainable change of practice was achieved, persons possessing first-hand experience in planning, running, and sustaining such projects were purposively recruited (35) from four cases of AI implementation that had AI systems procured and in use in practice. To achieve a rounded view of the process and variability between clinical areas and the types of AI, these participants represented positions within different cases carried out in large healthcare organizations from different geographical regions in Sweden (Table 2). Two of the projects were AI-based decision-support systems in different clinical areas, one project was an AI-based patient monitoring system, and one project was an AI-based system facilitating clinical administration. These cases were selected using convenience sampling (36) due to a scarcity of available cases where AI systems have been integrated into routine practice in healthcare. The interviews were conducted in May–September 2024 with 11 professionals in 9 interviews (Table 2). The participants were explained the study purpose and the procedure, and that the project does not handle any sensitive personal data. They were also informed about the possibility to access their anonymized data and provide corrections, if needed. Then, the participants provided verbal consent to participate and to express personal experiences and perceptions. The interview used open-ended questions with a focus to discern implementation process activities, their sequence, involved stakeholders, and lessons learned. The interviewees were asked questions like “Why and how did the implementation process start, who initiated the project, what was the trigger?”, followed by “How was this activity performed, who were involved?”, then directing the conversation to “What activities happened next?” until the sequence of activities was identified. Each interview lasted between 1.5 and 3 h, totaling 15 h.

Table 2

Table 2. Cases and participants in the interview study II.

2.3 Data analysis

The data on processual activities, decisions, involved stakeholders, challenges and lessons learned from every empirical study selected during the literature review were listed for mapping an individual process of every case. The studies provided varying level of detail, often with an extensive focus on a few activities depending on the purpose of the study or on the issues the authors had aimed to highlight. Then the data were deductively mapped onto the steps of QIF using direct qualitative content analysis (31) aiming to compare the processes described in the studies and the QIF. Data not fitting onto the QIF but reflecting realities of the AI implementation process were registered separately.

The interview transcripts from the interview study I detailing the perceptions of experts about QIF were reviewed using direct qualitative content analysis (33). Insights were labelled using the numbering of the QIF steps. The insights not fitting onto the QIF but reflecting realities of the AI implementation process were registered separately. This allowed to determine areas of alignment and gaps between QIF and AI implementation practice and to identify potential modifications or extensions of the QIF.

To extract the AI implementation process from the case interviews (interview study II), the interview transcripts were reviewed aiming to familiarize with the context, to identify start and end points, key actions, decisions, and involved parties and their responsibilities. Then the data were listed for mapping an individual process of every case. Then the data were deductively mapped onto the steps of QIF using direct qualitative content analysis (33). This allowed for comparisons between the processes of the case and the QIF. Data not fitting onto the QIF but reflecting realities of the AI implementation process were registered separately.

Three pairs of co-authors evaluated the data independently, discussed and reached a consensus through a number of iterations (37). Data not fitting onto the QIF were included in two different ways and later processed during the data analysis. First, data (condensed quotes) within the same area as respective QIF steps were documented in a separate column next to each step of QIF (Supplementary Material S3). Second, the data on activities of the practical AI implementation process that QIF does not cover were documented, analysed and summarized separately in the paragraph “Limitations of QIF for AI implementation” below.

3 Results

The data analysis showed that all the phases and steps in QIF are relevant in the context of AI implementation in healthcare practice. The processual activities extracted from the literature studies and the interviews representing the AI implementation cases have reflected the QIF process while providing additional details that should be further incorporated to enhance AI specifics. The expert interviews have also validated that QIF can provide a sound conceptual basis for describing and guiding the AI implementation process, although the specificities of the AI context would require some refinement in phases and steps and adding more details to effectively support it. Further, every step of QIF is discussed in the context of AI based on the data outlined in Supplementary Material S3, ending the section with the limitations of QIF for AI implementation. Definitions of the Cases 1–4 as well as Experts 1–7 referred to below can be found in Tables 1, 2.

3.1 QIF in the context of AI implementation in healthcare

3.1.1 QIF step 1—conducting a needs and resource assessment

According to QIF, this step recommends a thorough understanding of the problem, its root causes, and the intended beneficiaries of the innovation. The data underscores the relevance of step 1 in AI implementation, where practitioners utilize data analysis to assess the magnitude of the problem and systematically review published research for evidence supporting AI’s applicability to the problem (13, 38–42) Furthermore, the data analysis identified additional activities related to the needs assessment process relevant to AI implementation:

• Preparing a case demonstrating a clear need: what should be solved and why (Case 2). Such information is useful for managerial approvals, in communication activities and in change management.

3.1.2 QIF step 2—conducting a fit assessment

According to QIF, this step recommends assessing how the innovation fits with the setting by considering identified needs, the organization’s mission, values, strategy and priorities, as well as cultural preferences. The data confirms the necessity of step 2 for AI implementation, where practitioners investigate whether an internal development or purchasing of the AI system is aligned with the identified needs, the innovation strategy, institutional priorities, and identified needs as well as identifying what the added value would be (39, 40, 43–45). The data analysis identified additional activities related to the fit assessment process specific to AI implementation:

• Identifying if the AI system has a suitable regulatory certification (Expert 3, Case 3)

• Investigating relevance of the data that the commercial AI model was built on, possibly testing performance on own data (Expert 3, Expert 6)

• Investigating what are the conditions for retraining the model and for monitoring model’s performance (Expert 6)

• Investigating ethical aspects of the AI model: bias, participation, integrity, demographics, etc. (Expert 3)

• Investigating impact on data security (Case 3)

• Analyzing potential risks for the users and the patients (Expert 2, Case 4)

• Assessing legal aspects of the solution and the collaboration with the vendors (Case 3, Expert 4).

• Assessing compatibility of the AI system with organization’s processes (46).

3.1.3 QIF step 3—conducting a capacity readiness assessment

According to QIF, this step recommends assessing if the organization has adequate resources, skills, and motivation to implement the innovation. The data indicates that step 3 is also relevant to AI implementation, where the practitioners assess the required resources and consider potential changes in staff’s employment and roles (13, 39, 45), (Case 2, Case 4). The data analysis identified additional activities related to the capacity and readiness assessment relevant to AI implementation:

• Analyzing benefits vs. costs (43)

• Investigating the opportunity cost: what might be lost due to the introduction of AI, e.g., deskilling staff. (Expert 3)

• Analyzing possibilities for relocation or requalification of staff (Case 4)

• Investigating sufficiency of the technical environment in an organization (e.g., computing power) (Expert 6)

3.1.4 QIF step 4—possibility for adaptation

According to QIF, this step recommends assessing whether the innovation should be modified to fit the setting and the target group, and how the changes would be documented and monitored during the implementation. The data indicates that step 4 is relevant to AI implementation, with practitioners focusing on identifying organizational constraints and determining if any changes or additional developments of the AI system, IT infrastructure, or workflows are necessary (13, 39, 43–49). The data analysis identified that conducting a local pilot is an important activity dedicated to understanding the local relevance of the AI system and necessary adaptations. The pilot should be set up considering the technical infrastructure, legal side, contracts, and workflows. Before the pilot, training should take place, and evaluation should be planned, including the collection of user feedback based on the pilot (Case 1, Case 3, Case 4).

3.1.5 QIF step 5—obtaining explicit and implicit buy-in and approvals/permissions

According to QIF, this step includes achieving the buy-in by leadership, staff, communities, addressing organizational resistance, and recruiting innovation champions. The data confirms that step 5 is relevant to AI implementation, highlighting practitioners’ efforts to demonstrate the added value of the system, to organize support from staff and leadership, and appoint champions (13, 39–41, 44–45, 48–55). The data analysis identified additional activities related to obtaining the buy-in specific to the AI implementation process:

• Involving clinicians in designing a user-friendly AI system and its user interface, in system’s training and in contextualizing data representation (41)

• Addressing clinicians’ legal liability questions and preparing the documents detailing clinicians’ responsibilities in the context of the AI system (40, 44), (Expert 3), (Case 1, Case 2)

• Addressing algorithm’s explainability, availability, quality and safety (45, 48)

• Using a thoughtful framing about AI in communication with the stakeholders (13)

3.1.6 QIF step 6—building organizational capacity

According to QIF, this step includes investigating which aspects of the organization’s infrastructure, skills, and motivation require enhancement to accommodate the innovation. Also, QIF specifically points out that these enhancements do not directly assist with the implementation but instead create a supportive environment for success. The data analysis underscores the relevance of step 6 to AI implementation, particularly highlighting practitioners’ efforts to build relationships within the organization that enable better workflows related to the AI system (42). Although this step recommends investigating the capacity of the infrastructure, the analysis indicates that this activity should be planned earlier, for example, within the fit assessment in Step 2 in QIF.

3.1.7 QIF step 7—staff recruitment/maintenance

According to QIF, this step involves recruiting staff responsible for the implementation process and for supporting the frontline staff. It also includes an assessment of whether the roles of staff might change. The data indicates that step 7 is relevant to AI implementation, with project leaders and teams being formed or newly recruited to support the practitioners in this process (13, 56). The data analysis identified additional aspects specific to AI implementation:

• Managing AI implementation might require new recruitment not only for the project’s execution but also for later managing routine tasks involving the AI system (40, 42).

• Assessing potential changes in roles recommended in this step of QIF would be conducted earlier in the implementation process of AI, possibly during the organization’s capacity assessment (Step 3 in QIF). Without such understanding, it is hard to determine whether and how the introduction of the AI system would add value, and it complicates striving for the buy-in defined in earlier steps (Step 5) and the ethical handling of staff.

3.1.8 QIF step 8—effective pre-innovation staff training

According to QIF, this step includes teaching staff about the innovation and its values to enable them to apply the innovation. The data highlights the importance of step 8 to AI implementation, revealing that training is not limited only to the pre-innovation phase, but continuous throughout the next phases (13, 39, 42, 48, 49, 56–58), with each phase focusing on the following specific objectives:

• Pre-innovation training should focus on skills and competence building in the areas that enable and support the use of AI (50). Depending on the context, it could include, for example, cross-unit collaboration and communication, skills of training other people, empathy, attentive listening, and similar interpersonal competencies (50). In the pre-innovation phase, general technical AI competences should be developed, equipping the staff with conceptual, ethical, and legal knowledge about AI.

• During the implementation, training should concern the vision for change and the urgency for improvement, and the scientific and clinical basis for the specific AI system (42, 48, 49). Next, it is crucial to develop skills in the practical use of the specific AI system and the new workflows, highlighting the strengths and weaknesses of the AI model for trust enhancement (13).

• Post-implementation training intends to ensure sustainability of the new way of working and the AI system’s use (39). Apart from on-boarding of the new staff that joined after the implementation, the post-implementation training could include case-based reviews for training purposes (49).

• Preparing quality training materials and formalized flowcharts for the users to follow as well as conducting the training through different phases, can be costly and this cost category should be considered before the start of the implementation (13), (Case 1, Case 3, Case 4).

3.1.9 QIF step 9—creating implementation teams

In QIF, this step includes determining who in the organization would carry the responsibility for the implementation process and outcomes as well as who would be the supporting team. The data indicates that step 9 is relevant to AI implementation, with implementation project leaders and teams being formed or newly recruited (13, 40, 56), (Expert 7, Case 1, Case 2, Case 4).

3.1.10 QIF step 10—developing an implementation plan

According to QIF, this step includes setting up tasks and timelines and foreseeing challenges that can hinder the implementation. The data confirms the necessity of step 10 for AI implementation, emphasizing that practitioners worked on setting clear goals, planning test phases and establishing a structured meeting schedule (13, 44). Additional actions specific to developing a plan for AI implementation were:

• Assigning an organizational owner of the AI system which would be responsible for the system’s rollout, maintenance and changes in the future, as well as the related budget (Case 2)

• Creating a communication plan (Case 4)

3.1.11 QIF step 11—technical assistance/coaching/supervision governance

This step in QIF is meant to address the practical problems arising during the implementation. The data indicates that step 11 is relevant to AI implementation, emphasizing the importance of creating robust support mechanisms. Practitioners concentrated on ensuring the availability of specialists, establishing IT support and chat systems and equipping staff with essential resources, including appropriate equipment and training materials (13, 52), (Case 1, Case 4). The data analysis did not identify any further activities specific to AI implementation.

3.1.12 QIF step 12—process evaluation

This step in QIF concerns the evaluation of the implementation process; how it unfolds over time in light of the innovation being implemented, and how different individuals performed. The data has shown that step 12 is relevant to AI implementation, and underscores its relevance for practitioners to evaluate not only the AI implementation itself but also key factors such as barriers and facilitators to adoption, unintended social consequences, impact on clinical roles, responsibilities, and trust, and perceptions of evidence (13, 48). The data analysis identified additional evaluation aspects relevant to AI implementation:

• Assessing clinical and operational impact to demonstrate safety, efficacy, and added value (13, 39, 44, 48, 60), (Case 1)

• Assessing and monitoring users’ engagement and AI system’s usage (59), (Case 4)

• Evaluating effects on processes and staff caused through the new practice when using AI system (39, 60)

• Gathering and analyzing user feedback (13, 61)

3.1.13 QIF step 13—supportive feedback mechanism

According to QIF, this step recommends creating a process and channels through which feedback on the implementation process could be communicated with those involved in the innovation. The data has shown that step 13 is relevant to AI implementation, focusing on gathering feedback and further requests for adaptations, identifying shortcomings or risks, and addressing the need for additional training. This feedback and insights were collected post-deployment of AI through governance activities through analyzing data and using different channels such as e-mail, web-based survey, or meetings with staff (13, 41, 45, 50, 57, 58, 62, 63), (Case 1). The data analysis did not identify any further activities specific to AI implementation.

3.1.14 QIF step 14—learning from experience

According to QIF, this step recommends analyzing and reflecting upon lessons learned from the implementation process and sharing them with other interested parties. The data analysis underscores the relevance of step 14 to AI implementation, building on the premise that AI system’s implementation is never finished and needs continuous monitoring, maintenance, and development to fit the business, particularly highlighting importance of such reflection and knowledge transfer in the situations where consultants were recruited for assisting in the implementation (Expert 1).

3.2 Limitations of QIF for AI implementation

The analysis revealed critical limitations in QIF when applied to AI implementation in healthcare, necessitating emphasis on two additional phases and important activities. These include a dedicated phase for the practical implementation activities that occur after planning, ensuring a smooth transition from design to deployment, and a phase focused on governance and sustainability, aimed at maintaining the AI’s long-term impact. Additionally, the activity of continuous engagement of diverse stakeholders throughout the lifecycle of the AI implementation project is essential for its success. Incorporating these components into a dedicated framework is critical to support effective and sustainable deployment of AI in healthcare.

3.2.1 Practical implementation activities after planning

The data underscores the importance of incorporating a phase specifically dedicated to the actual implementation activities that occur after planning. The data analysis revealed that during the AI-specific implementation phase, the actual changes in the local process workflows and protocols should be established. Their descriptions and all the related documentation should be updated by incorporating AI details, and those changes should be incorporated into the training materials (44, 46, 49–51), (Case 2, Case 4). Accordingly, existing roles and competences might require adaptations in the job descriptions which should also be necessary to include in the training (39, 41, 42). On the technological side, actual changes in the IT infrastructure of the organization should be implemented establishing the secure data management, building test and production environments (46, 51), (Case 2). The data analysis also underscores the importance of change management activities, such as continuous communication with staff, AI vendors and developers and conducting training of AI system’s users addressing the system’s technical specifics, the new workflows and roles (Case 2, Case 4). Finally, AI system’s usage monitoring and compliance procedures should be established to follow up on the actual utilization of the system (13, 41, 45).

3.2.2 Governance activities for ensuring sustainability of Ai system

The data highlights the need to incorporate a dedicated phase focused on governance and sustainability to ensure the ongoing maintenance and success of AI innovations. The data analysis revealed that the effective means for promoting sustainability of the AI system implemented is through creating a governance body or committee (13, 41, 45) that could include representatives from leadership, frontline staff, information technology and innovation specialists. Activities of the governance committee include pilot study, reorganizing the workflows, monitoring AI system’s performance and effectiveness and a possible “model drift”, addressing concerns of staff collected through the feedback mechanisms, promoting and tracking usage of the system, developing reporting, monitoring for a proper insertion of patient records, providing post-implementation training, prioritizing and approving changes, and planning for further financing (13, 41, 45, 50), (Expert 6).

3.2.3 Continuous engagement of diverse stakeholders

The data analysis revealed that engagement of different stakeholders spans throughout the lifecycle of the AI implementation project. The engagement should start at the ideation and problem formulation phase (46) and should be sustained throughout the implementation cycle, which is crucial for managing change (13, 40, 45, 46, 50, 55, 56, 62). Obtaining stakeholders’ confirmation that the problem is relevant allows to consider a global view and to formulate clear use case at the onset of the implementation since different stakeholders might need different output from AI or have different interpretations; it also promotes their readiness for change (40, 45, 46). Throughout the implementation process, engaging the stakeholders can help in several ways. It includes obtaining scientific, theoretical and practical knowledge of the health area which is useful for technology professionals (50). Next, the stakeholders can help in defining the parameters of the targeted population, data features to train the model on, and to think through how the model could be deployed (41, 46). Later, they can report on the user experience and whether the AI model aligns with their needs (50). Moreover, the stakeholders can play a role in designing the evaluation of the AI model (41) and can participate in creating the training materials (13). Finally, early engagement assists in change management and training since clinicians develop intuition and gain experience that advances their ability to train others (13).

4 Discussion

This study aimed to explore the activities typical to implementation of AI-based systems for developing an AI implementation process framework intended to guide healthcare professionals. To achieve this aim, QIF was considered as an initial reference framework (26).

Previous research has indicated that a lack of stepped guidance and AI specificity in the presently available implementation frameworks can lead to patient safety risks, dissatisfaction of clinical staff, extra costs and delays in the AI system’s deployment as demonstrated by previous research (13, 17, 39, 40). The key findings of this study are the identified processual phases and activities specific and typical to AI implementation projects to be carried out by an adopting healthcare organization when AI systems are deployed in routine practice. By reflecting upon the QIF, additional phases of implementation and governance were discerned, and different AI-specific activities were detected in connection to the QIF steps.

A significant finding of the study is that successful AI implementation necessitates comprehensive pre-implementation assessments. Some types of identified assessments align with the ones suggested by QIF, such as the in-depth analysis of the problem and stakeholder needs, assessing compatibility of the AI system with an organization’s strategy, processes, IT infrastructure, and potential impact on patients, clinical roles, and the organizational culture. These aspects in need of assessment were also recommended by previous research discussing the innovation process or determinants in the context of digital health and health innovation (64–66). The present study identified additional types of recommended assessments in the context of AI and located them as necessary activities in the planning phase: investigating explainability of AI algorithm (45, 48), relevance of an AI model to the local data (Expert 3) and conditions for re-training the model (Expert 6), level of data security and protection (Case 3), risk-consequence analysis (Expert 2, Expert 3, Case 4), cost-benefit analysis (43), considering the opportunity cost (Expert 3), ethics and patient safety considerations (45, 48), (Expert 3). Furthermore, the present study identified several preparatory actions in the legal area not described by the previous research. In addition to the traditional activities like negotiating different contracts and agreements between the organization and the vendor and verifying whether the AI system is suitably certified (59), this study added activities like setting up agreements regarding the data flow and protection and defining the policy explaining clinicians’ liability (40, 43, 44), (Expert 3).

The present study identified a number of important activities that could be added to the actual implementation phase. These implementation activities include creating new or updating and standardizing local procedures and clinical protocols integrating the AI system, changing or creating new job descriptions, adapting IT architecture and setting up secure data management, conducting training and continuously communicating with users (44, 46, 49–51, 57), (Case 2, Case 4). Previous research either did not emphasize them entirely (65) or has reflected upon elements of it through different determinants, barriers and facilitators (64). This study was not able to place these activities within the process view of QIF which indicates that a specialized framework for AI implementation process should integrate them in the future.

This study underscores the importance of incorporating creation of the organizational structures and processes for the post-implementation governance period (13, 41, 45) as an integral part of the AI implementation framework. In the “follow-up” stages of AI-implementation, these processes are even more important, because not only the continued human adherence guarantees the sustained use, but also the monitoring of the model’s performance and patient safety. Notably, this crucial phase dedicated to governance of maintenance and sustainability of the AI systems is absent in existing implementation frameworks such as (26) and (64). While specialized frameworks and different national guidelines focusing on governance of AI are beginning to emerge (67), a process-oriented perspective remains lacking. This perspective should address several key questions: When and how should governance work commence? What practical organizational structures and processes can facilitate effective governance? What is the role of leadership in this context? How can various stakeholders (e.g., healthcare organizations and vendors) collaborate to ensure the sustained use of AI systems? Ultimately, how can we guarantee the enduring effectiveness and sustainability of AI governance? Addressing these critical questions is paramount to ensuring that governance, maintenance, and sustainability of AI systems remain a top priority for healthcare practitioners during the implementation process.

Further, other significant finding is the importance of prioritizing staff training throughout all phases of AI implementation, pre-, during, and post-implementation, to ensure success. These findings extend previous research that was limited to the overall content of the training for AI implementation (22, 67). This study emphasizes that pre-innovation training might be insufficient in the context of AI. It requires a continuous, iterative approach starting pre- and throughout the lifecycle of the AI system and should have a focus on the related clinical context and processes surrounding the AI system. In addition, specific competences and skills of staff should be developed prior to engaging in AI implementation (50). The study also revealed insights into the intensity of resources required for conducting effective training that enables staff to engage in using the AI system in their practice sustainably. The cost category related to continuous training and preparation of the training materials should be considered when evaluating the organization’s capacity for conducting the AI implementation and in planning the costs (41), (Case 2).

Lastly, another crucial finding is that engagement of the stakeholders should span throughout all the phases of the AI implementation (13, 40, 45, 46, 50, 55, 56, 62) which corresponds to previous research highlighting the importance of stakeholder input and involvement (22, 64, 65). However, earlier implementation frameworks have not discerned the practical activities of stakeholder involvement through the processual perspective of different phases of implementation. During the preparatory phase, the stakeholders’ role is to provide the subject knowledge, to confirm the relevance of the problem and the use case, to help define the target population and the parameters of the training data, and to envision the new workflows integrating the AI system (40, 41, 45, 46). The stakeholders should also be part of the AI system’s piloting and evaluation activities and provide feedback (41). During the implementation phase, the stakeholders are involved in training as trainers or trainees and assist in communication and dissemination activities (13). During the governance and maintenance phase, the stakeholders can have roles in the AI system’s governance committee, provide continuous feedback and ideas for improvement (13, 41, 45).

While this study offers valuable insights into the process of AI systems’ implementation, several limitations exist. First, the dataset used for analysis does not include scientific publications published since 2024 or cases representing generative AI applications. While the present study provides a solid foundation for understanding the implementation process for AI systems, newer types of AI models might present additional important implementation practices not covered in this article. Further, the data collected through the interview studies with the representatives from AI implementation cases and the experts represent the Swedish context and might not depict some practices manifesting in other healthcare contexts. Moreover, the cases were selected using convenience sampling which might provide limited generalizability of the results and overrepresentation of certain types of AI systems. Also, access to interviewees in Case 1 was limited to the implementation leader. This limitation was partially addressed by conducting two interviews with the same person to obtain more details on the implementation process.

Future research should discern processual differences based on the typology of AI (for example, AI systems dedicated to improving administrative processes of healthcare organizations compared to diagnostic AI systems). Another potential research avenue could relate to exploring potential differences in implementation when AI models are developed internally at healthcare organizations and procured from external vendors. How can stakeholders be educated to ensure their proper engagement throughout all phases. The future research could also address practicalities behind different activities in the implementation process and how they influence the success of the implementation and its sustainability. Systematizing such knowledge would create a better understanding of the interconnections between the activities and would help understand their relative weight for success of the implementation.

5 Conclusion

The study aimed to explore the activities typical to implementation of AI-based systems for developing an AI implementation process framework intended to guide healthcare professionals. To achieve this aim, the Quality Implementation Framework was considered as an initial reference framework. The analysis revealed its gaps when applied to AI implementation in healthcare, necessitating the inclusion of additional phases and important components. These components include a dedicated phase for the practical implementation activities that occur after planning, ensuring a smooth transition from design to deployment, and a phase focused on governance and sustainability, aimed at maintaining the AI’s long-term impact. Additionally, the component of continuous engagement of diverse stakeholders throughout the lifecycle of the AI implementation project is essential for its success. The identified processual peculiarities that the AI carries allow for a more informed practitioner action and more specific, AI-tailored development of the implementation methodologies.

Data availability statement

The datasets presented in this study can be found in online repositories. The names of the repository/repositories and accession number(s) can be found in the article/Supplementary Material.

Author contributions

MN: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Visualization, Writing – original draft, Writing – review & editing. JN: Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration, Resources, Supervision, Visualization, Writing – original draft, Writing – review & editing. PN: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Writing – review & editing. FG: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Writing – review & editing. MN: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Writing – review & editing. IL: Conceptualization, Methodology, Writing – original draft, Writing – review & editing. PS: Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration, Resources, Supervision, Visualization, Writing – review & editing.

Funding

The author(s) declare that financial support was received for the research and/or publication of this article. This project has received funding from the Swedish Innovation Agency (2019-04526) and the Knowledge Foundation (20200208 01H).

Acknowledgments

The authors are grateful to the interview participants for their insights and dedication in exploring the specificities of the AI implementation process.

Conflict of interest

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Generative AI statement

The author(s) declare that no Generative AI was used in the creation of this manuscript.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fdgth.2025.1550459/full#supplementary-material

References

1. Bajwa J, Munir U, Nori A, Williams B. Artificial intelligence in healthcare: transforming the practice of medicine. Future Healthc J. (2021) 8(2):e188–94. doi: 10.7861/fhj.2021-0095 PubMed Abstract | Crossref Full Text | Google Scholar

2. Sharma M, Savage C, Nair M, Larsson I, Svedberg P, Nygren JM. Artificial intelligence applications in health care practice: scoping review. J Med Internet Res. (2022) 24(10):e40238. doi: 10.2196/40238 PubMed Abstract | Crossref Full Text | Google Scholar

3. Li YH, Li YL, Wei MY, Li GY. Innovation and challenges of artificial intelligence technology in personalized healthcare. Sci Rep. (2024) 14:18994. doi: 10.1038/s41598-024-70073-7 PubMed Abstract | Crossref Full Text | Google Scholar

4. Liu CF, Huang CC, Wang JJ, Kuo KM, Chen CJ. The critical factors affecting the deployment and scaling of healthcare AI: viewpoint from an experienced medical center. Healthcare. (2021) 9(6):685. doi: 10.3390/healthcare9060685 PubMed Abstract | Crossref Full Text | Google Scholar

5. Chomutare T, Tejedor M, Svenning TO, Marco-Ruiz L, Tayefi M, Lind K, et al. Artificial intelligence implementation in healthcare: a theory-based scoping review of barriers and facilitators. IJERPH. (2022) 19(23):16359. doi: 10.3390/ijerph192316359 PubMed Abstract | Crossref Full Text | Google Scholar

6. Tachkov K, Zemplenyi A, Kamusheva M, Dimitrova M, Siirtola P, Pontén J, et al. Barriers to use artificial intelligence methodologies in health technology assessment in central and east European countries. Front Public Health. (2022) 10:921226. doi: 10.3389/fpubh.2022.921226 PubMed Abstract | Crossref Full Text | Google Scholar

7. Bérubé M, Giannelia M, Vial T. Barriers to the implementation of AI in organizations: findings from a delphi study. Proceedings of the 54th Hawaii International Conference on System Sciences (2021). doi: 10.24251/HICSS.2021.805 Crossref Full Text | Google Scholar

8. Petersson L, Larsson I, Nygren JM, Nilsen P, Neher M, Reed JE, et al. Challenges to implementing artificial intelligence in healthcare: a qualitative interview study with healthcare leaders in Sweden. BMC Health Serv Res. (2022) 22:850. doi: 10.1186/s12913-022-08215-8 PubMed Abstract | Crossref Full Text | Google Scholar

9. Nair M, Svedberg P, Larsson I, Nygren JM. A comprehensive overview of barriers and strategies for AI implementation in healthcare: mixed-method design. PLoS One. (2024) 19(8):e0305949. doi: 10.1371/journal.pone.0305949 PubMed Abstract | Crossref Full Text | Google Scholar

11. He J, Baxter SL, Xu J, Xu J, Zhou X, Zhang K. The practical implementation of artificial intelligence technologies in medicine. Nat Med. (2019) 25(1):30–6. doi: 10.1038/s41591-018-0307-0 PubMed Abstract | Crossref Full Text | Google Scholar

12. Singh RP, Hom GL, Abramoff MD, Campbell JP, Chiang MF. Current challenges and barriers to real-world artificial intelligence adoption for the healthcare system, provider, and the patient. Trans Vis Sci Tech. (2020) 9(2):45. doi: 10.1167/tvst.9.2.45 PubMed Abstract | Crossref Full Text | Google Scholar

13. Sendak MP, Ratliff W, Sarro D, Alderton E, Futoma J, Gao M, et al. Real-world integration of a sepsis deep learning technology into routine clinical care: implementation study. JMIR Med Inform. (2020) 8(7):e15182. doi: 10.2196/15182 PubMed Abstract | Crossref Full Text | Google Scholar

14. Morales HM, Guedes M, Silva JS, Massuda A. COVID-19 in Brazil-preliminary analysis of response supported by artificial intelligence in municipalities. Front Digit Health. (2021) 17(3):648585. doi: 10.3389/fdgth.2021.648585 Crossref Full Text | Google Scholar

15. Palaniappan K, Lin EYT, Vogel S. Global regulatory frameworks for the use of artificial intelligence (AI) in the healthcare services sector. Healthcare. (2024) 12(5):562. doi: 10.3390/healthcare12050562 PubMed Abstract | Crossref Full Text | Google Scholar

16. OECD. The OECD AI principles (2019). Available at: https://www.oecd.org/en/topics/ai-principles.html (Accessed April 28, 2025). Google Scholar

17. Gama F, Tyskbo D, Nygren J, Barlow J, Reed J, Svedberg P. Implementation frameworks for artificial intelligence translation into health care practice: scoping review. J Med Internet Res. (2022) 24(1):e32215. doi: 10.2196/32215 PubMed Abstract | Crossref Full Text | Google Scholar

18. Svedberg P, Reed J, Nilsen P, Barlow J, Macrae C, Nygren J. Toward successful implementation of artificial intelligence in health care practice: protocol for a research program. JMIR Res Protoc. (2022) 11(3):e34920. doi: 10.2196/34920 PubMed Abstract | Crossref Full Text | Google Scholar

19. van der Vegt AH, Scott IA, Dermawan K, Schnetler RJ, Kalke VR, Lane PJ. Implementation frameworks for end-to-end clinical AI: derivation of the SALIENT framework. JAMIA. (2023) 30(9):1503–15. doi: 10.1093/jamia/ocad088 PubMed Abstract | Crossref Full Text | Google Scholar

20. Ji M, Genchev GZ, Huang H, Xu T, Lu H, Yu G. Evaluation framework for successful artificial intelligence-enabled clinical decision support systems: mixed methods study. J Med Internet Res. (2021) 23(6):e25929. doi: 10.2196/25929 PubMed Abstract | Crossref Full Text | Google Scholar

21. Liu X, Rivera SC, Moher D, Calvert MJ, Denniston AK. Reporting guidelines for clinical trial reports for interventions involving artificial intelligence: the CONSORT-AI extension. Lancet Digit Health. (2022) 2(10):e537–48. doi: 10.1016/S2589-7500(20)30218-1 PubMed Abstract | Crossref Full Text | Google Scholar

22. Lekadir K, Feragen A, Fofanah AJ, Frangi AF, Buyx A, Emelie A, et al. FUTURE-AI: International consensus guideline for trustworthy and deployable artificial intelligence in healthcare. arXiv [preprint] (2023). Available at: https://arxiv.org/abs/2309.12325 (Accessed December 19, 2024). Google Scholar

23. Roppelt JS, Kanbach DK, Kraus S. Artificial intelligence in healthcare institutions: a systematic literature review on influencing factors. Technol Soc. (2024) 76:102443. doi: 10.1016/j.techsoc.2023.102443 Crossref Full Text | Google Scholar

24. Kim JY, Boag W, Gulamali F, Hasan A, Hogg HD, Lifson M, et al. Organizational governance of emerging technologies: aI adoption in healthcare. Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency (2023). p. 1396–417. doi: 10.1145/3593013.3594089 Crossref Full Text | Google Scholar

25. Nilsen P, Svedberg P, Nygren J, Frideros M, Johansson J, Schueller S. Accelerating the impact of artificial intelligence in mental healthcare through implementation science. Implement Res Pract. (2022) 3:26334895221112033. doi: 10.1177/26334895221112033 PubMed Abstract | Crossref Full Text | Google Scholar

27. Meyers DC, Durlak JA, Wandersman A. The quality implementation framework: a synthesis of critical steps in the implementation process. AJCP. (2012) 50:462–80. doi: 10.1007/s10464-012-9522-x PubMed Abstract | Crossref Full Text | Google Scholar

28. Nilsen P, Moore JE. Process models. In: Nilsen P, editor. Implementation Science. London: Routledge (2024). p. 41–52. Google Scholar

29. Eboreime EA, Eyles J, Nxumalo N, Eboreime OL, Ramaswamy R. Implementation process and quality of a primary health care system improvement initiative in a decentralized context: a retrospective appraisal using the quality implementation framework. Int J Health Plann Manage. (2019) 34(1):e369–86. doi: 10.1002/hpm.2655 PubMed Abstract | Crossref Full Text | Google Scholar

30. Nor M, Taib NA, Saad NA, Zaini M, Ahmad HS, Ahmad Z, et al. Development of electronic medical records for clinical and research purposes: the breast cancer module using an implementation framework in a middle income country—Malaysia. BMC Bioinform. (2019) 19(13):402. doi: 10.1186/s12859-018-2406-9 Crossref Full Text | Google Scholar

31. McCargar SI, Olsen J, Steelman RJ, Huang JH, Palmer EA, Burch GH, et al. Implementation of a standardized oral screening tool by paediatric cardiologists. Cardiol Young. (2020) 30(12):1815–20. doi: 10.1017/S1047951120002826 PubMed Abstract | Crossref Full Text | Google Scholar

32. Nilsen P, Svedberg P, Neher M, Nair M, Larsson I, Petersson L, et al. A framework to guide implementation of AI in health care: protocol for a cocreation research project. JMIR Res Protoc. (2023) 12(1):e50216. doi: 10.2196/50216 PubMed Abstract | Crossref Full Text | Google Scholar

34. Creswell JW, Clark VLP. Designing and Conducting Mixed Methods Research. Thousand Oaks, CA: Sage publications (2017). Google Scholar

35. Palinkas LA, Horwitz SM, Green CA, Wisdom JP, Duan N, Hoagwood K. Purposeful sampling for qualitative data collection and analysis in mixed method implementation research. Adm Policy Ment Health. (2015) 42(5):533–44. doi: 10.1007/s10488-013-0528-y PubMed Abstract | Crossref Full Text | Google Scholar

36. Elfil M, Negida A. Sampling methods in clinical research; an educational review. Emergency. (2017) 5(1):e52.28286859 PubMed Abstract | Google Scholar

37. Kononowech J, Hagedorn H, Hall C, Helfrich CD, Lambert-Kerzner AC, Miller SC, et al. Mapping the organizational readiness to change assessment to the consolidated framework for implementation research. Implement Sci Commun. (2021) 2:19. doi: 10.1186/s43058-021-00121-0 PubMed Abstract | Crossref Full Text | Google Scholar

38. Lai L, Wittbold K, Dadabhoy FZ, Sato R, Landman AB, Schwamm LH, et al. Digital triage: novel strategies for population health management in response to the COVID-19 pandemic. Healthc. (2020) 8(4):10049. doi: 10.1016/j.hjdsi.2020.100493 Crossref Full Text | Google Scholar

39. Scheinker D, Brandeau ML. Implementing analytics projects in a hospital: successes, failures, and opportunities. INFORMS J. Appl. Anal. (2020) 50:176–89. doi: 10.1287/inte.2020.1036 Crossref Full Text | Google Scholar

40. Reis L, Maier C, Mattke J, Creutzenberg M, Weitzel T. Addressing user resistance would have prevented a healthcare AI project failure. MIS Q Exec. (2020) 19(4):279–96. doi: 10.17705/2msqe.00038 Crossref Full Text | Google Scholar

41. Sendak M, Elish MC, Gao M, Futoma J, Ratliff W, Nichols M, et al. The human body is a black box” supporting clinical decision-making with deep learning. Proceedings of: The 2020 Conference on Fairness, Accountability, and Transparency (2020). p. 99–109. doi: 10.1145/3351095.3372827 Crossref Full Text | Google Scholar

42. Sandhu S, Lin AL, Brajer N, Sperling J, Ratliff W, Bedoya AD, et al. Integrating a machine learning system into clinical workflows: qualitative study. JMIR. (2020) 22(11):e22421. doi: 10.2196/22421 PubMed Abstract | Crossref Full Text | Google Scholar

43. Wijnhoven F. Organizational learning for intelligence amplification adoption: lessons from a clinical decision support system adoption project. Inf Syst Front. (2021) 24(3):731–44. doi: 10.1007/s10796-021-10206-9 Crossref Full Text | Google Scholar

44. Strohm L, Hehakaya C, Ranschaert ER, Boon WP, Moors EH. Implementation of artificial intelligence (AI) applications in radiology: hindering and facilitating factors. Eur Radiol. (2020) 30(10):5525–32. doi: 10.1007/s00330-020-06946-y PubMed Abstract | Crossref Full Text | Google Scholar

45. Baxter SL, Bass JS, Sitapati AM. Barriers to implementing an artificial intelligence model for unplanned readmissions. ACI Open. (2020) 4(2):e108–13. doi: 10.1055/s-0040-1716748 PubMed Abstract | Crossref Full Text | Google Scholar

46. Ng R, Tan KB. Implementing an individual-centric discharge process across Singapore public hospitals. Int J Environ Res Public Health. (2021) 18(16):8700. doi: 10.3390/ijerph18168700 PubMed Abstract | Crossref Full Text | Google Scholar

47. Davis MA, Rao B, Cedeno P, Saha A, Zohrabian VM. Machine learning and improved quality metrics in acute intracranial hemorrhage by non-contrast computed tomography. Curr Probl Diagn Radiol. (2020) 51:556–61. doi: 10.1067/j.cpradiol.2020.10.007 PubMed Abstract | Crossref Full Text | Google Scholar

48. Romero-Brufau S, Wyatt KD, Boyum P, Mickelson M, Moore M, Cognetta-Rieke C. Implementation of artificial intelligence based clinical decision support to reduce hospital readmissions at a regional hospital. Appl Clin Inform. (2020) 11:570–7. doi: 10.1055/s-0040-1715827 PubMed Abstract | Crossref Full Text | Google Scholar

50. Gonçalves LS, Amaro ML, Romero AD, Schamne FK, Fressatto JL, Bezerra CW. Implementation of an artificial intelligence algorithm for sepsis detection. Rev Bras Enferm. (2020) 73(3):e20180421. doi: 10.1590/0034-7167-2018-0421 PubMed Abstract | Crossref Full Text | Google Scholar

51. Petitgand C, Motulsky A, Denis JL, Régis C. Investigating the barriers to physician adoption of an artificial intelligence- based decision support system in emergency care: an interpretative qualitative study. Stud Health Technol Inform. (2020) 270:1001–5. doi: 10.3233/SHTI200312 PubMed Abstract | Crossref Full Text | Google Scholar

52. Saverino A, Baiardi P, Galata G, Pedemonte G, Vassallo C, Pistarini C. The challenge of reorganizing rehabilitation services at the time of COVID-19 pandemic: a new digital and artificial intelligence platform to support team work in planning and delivering safe and high quality care. Front Neurol. (2021) 12:643251. doi: 10.3389/fneur.2021.643251 PubMed Abstract | Crossref Full Text | Google Scholar

53. Herman B, Siichokchatchawan W, Nantasenamat C, Pongpanich S. Artificial intelligence in overcoming rifampicin resistant-screening challenges in Indonesia: a qualitative study on the user experience of CUHAS-ROBUST. J Health Res. (2021) 36(6):1018–27. doi: 10.1108/JHR-11-2020-0535 Crossref Full Text | Google Scholar

54. Wen A, Fu S, Moon S, Wazir E, Rosenbaum M, Kaggal A, et al. Desiderata for delivering NLP to accelerate healthcare AI advancement and a mayo clinic NLP-as-a-service implementation. NPJ Digit Med. (2019) 2:130. doi: 10.1038/s41746-019-0208-8 PubMed Abstract | Crossref Full Text | Google Scholar

55. Murphree DH, Wilson PM, Asai SW, Quest DJ, Lin Y, Mukherjee P, et al. Improving the delivery of palliative care through predictive modeling and healthcare informatics. JAMIA. (2021) 28(6):1065–73. doi: 10.1093/jamia/ocaa211 PubMed Abstract | Crossref Full Text | Google Scholar

56. Schlicher J, Metsker M, Shah H, Demirkan H. From NASA to healthcare: real-time data analytics (mission control) is reshaping healthcare services. Perspect Health Inf Manag. (2021) 18(4):1. Available at: https://pubmed.ncbi.nlm.nih.gov/34975356/ Google Scholar

57. Xu H, Li P, Yang Z, Liu X, Wang Z, Yan W, et al. Construction and application of a medical-grade wireless monitoring system for physiological signals at general wards. J Med Syst. (2020) 44:1–15. doi: 10.1007/s10916-020-01653-z Crossref Full Text | Google Scholar

58. McCoy A, Das R. Reducing patient mortality, length of stay and readmissions through machine learning-based sepsis prediction in the emergency department, intensive care unit and hospital floor units. BMJ Open Qual. (2017) 6:e000158. doi: 10.1136/bmjoq-2017-000158 PubMed Abstract | Crossref Full Text | Google Scholar

59. Lacey G, Zhou J, Li X, Craven C, Gush C. The impact of automatic video auditing with real-time feedback on the quality and quantity of handwash events in a hospital setting. Am J Infect Control. (2020) 48(2):162–6. doi: 10.1016/j.ajic.2019.06.015 PubMed Abstract | Crossref Full Text | Google Scholar

60. Cruz NP, Canales L, Muñoz JG, Pérez B, Arnott I. Improving adherence to clinical pathways through natural language processing on electronic medical records. In: Ohno-Machado L, Séroussi B, editors. MEDINFO: Health and Wellbeing e-Networks for All. Amsterdam, the Netherlands: IOS Press (2019). p. 561–5. doi: 10.3233/SHTI190285 Crossref Full Text | Google Scholar

61. Joerin A, Rauws M, Ackerman ML. Psychological artificial intelligence service, tess: delivering on-demand support to patients and their caregivers: technical report. Cureus. (2019) 11:e3972. doi: 10.7759/cureus.3972 PubMed Abstract | Crossref Full Text | Google Scholar

62. Chonde DB, Pourvaziri A, Williams J, McGowan J, Moskos M, Alvarez C, et al. Radtranslate: an artificial intelligence-powered intervention for urgent imaging to enhance care equity for patients with limited English proficiency during the COVID-19 pandemic. J Am Coll Radiol. (2021) 18(7):1000–8. doi: 10.1016/j.jacr.2021.01.013 PubMed Abstract | Crossref Full Text | Google Scholar

63. Sun TQ. Adopting artificial intelligence in public healthcare: the effect of social power and learning algorithms. Int J Environ Res Public Health. (2021) 18(23):12682. doi: 10.3390/ijerph182312682 PubMed Abstract | Crossref Full Text | Google Scholar

64. Greenhalgh T, Wherton J, Papoutsi C, Lynch J, Hughes G, A’Court C, et al. Beyond adoption: a new framework for theorizing and evaluating nonadoption, abandonment, and challenges to the scale-up, spread, and sustainability of health and care technologies. J Med Internet Res. (2017) 19(11):e367. doi: 10.2196/jmir.8775 PubMed Abstract | Crossref Full Text | Google Scholar

66. Lee E, Sjöqvist B, Hagiwara M, Söderholm HM, Candefjord S. Development of verified innovation process for healthcare solutions (VIPHS): a stepwise model for digital health. Stud Health Technol Inform. (2023) 302:736–40. doi: 10.3233/SHTI230250 PubMed Abstract | Crossref Full Text | Google Scholar

Understanding foundational and emerging AI medical terminology is crucial to navigating the ever-changing landscape of artificial intelligence in healthcare.

AI has the potential to significantly bolster industry-wide efforts to wrangle large amounts of digital health data and generate actionable insights from it — so much so that health systems are prioritizing AI initiatives despite implementation challenges. Additionally, industry leaders recommend that healthcare organizations stay on top of AI governance, transparency and collaboration.

This primer explores some of the most common terms and concepts stakeholders must understand to harness healthcare AI successfully.

2. Algorithmic bias Algorithmic bias, also known as AI bias or machine learning bias, results from an algorithm being trained on an incomplete or poor-quality data set. When the training data is skewed or contains cognitive biases, those biases are transferred to the algorithm and its results. For example, in healthcare, if an AI-enabled clinical decision-making algorithm is trained using only data from male patients or patients of only one race, it only produces results representative of those populations. Nondiverse clinical data sets can negatively affect health equity efforts. The root of a biased algorithm often lies in the data it’s built upon, so it’s essential to identify cognitive, sample, automation and prejudice biases during data collection and preparation. Maintaining unbiased algorithms requires due diligence at every machine learning development cycle stage, not just during data collection and preparation. Stewards of responsible AI should also consider the presence of bias during the development phase and after deployment. Strategies to reduce algorithmic bias include identifying potential sources of bias, following an established framework to address bias and documenting how data is selected.

3. Artificial intelligence AI systems learn from large amounts of data and use those lessons to make predictions. AI simulates human intelligence and promises to perform tasks efficiently and effectively. AI tools are driven by algorithms, which act as instructions a computer follows to perform a computation or solve a problem. Generally, there are four types of AI categorized based on functionality: reactive, limited memory machines, theory of mind and self-aware. AI tools use predefined logic or rules-based learning to understand patterns in data using machine learning or neural networks to simulate the human brain and generate insights through deep learning. AI models can use these learning approaches to engage in computer vision, a process for deriving information from images and videos. These models can also use natural language processing (NLP) to derive insights from text and generative AI (GenAI) to create content. AI models can be classified as either explainable — meaning users have some insight into the “how” and “why” of an AI tool’s decision-making — or black box, a phenomenon in which the tool’s decision-making process is hidden from users. Currently, all AI models are considered narrow or weak AI. These are tools designed to perform specific tasks within certain parameters. Artificial general intelligence, or strong AI, is a theoretical system under which an AI model could be applied to any task. Much of the conversation around AI in healthcare is centered on currently realized AI — i.e., tools that exist for practical applications today or in the very near future. To understand health AI, one must have a basic understanding of data analytics in healthcare. Data analytics aims to extract useful information and insights from various data points or sources. In healthcare, information for analytics is typically collected from sources such as electronic health records (EHRs), claims data and peer-reviewed clinical research. Analytics efforts often aim to help health systems meet a key strategic goal, such as improving patient outcomes, enhancing chronic disease management, advancing precision medicine or guiding population health management. However, these initiatives require analyzing vast amounts of data, which is often time- and resource-intensive. AI presents a promising way to streamline the healthcare analytics process.

4. Black box AI A black box AI system is one in which the system’s decision-making process remains hidden from users. It’s known for its lack of transparency and is the opposite of explainable, or white box, AI. Black box AI’s inner workings are a mystery to the user and its developers, as it provides an output without explaining how it reached that conclusion. For example, OpenAI’s ChatGPT is a black box AI system. Black box AI systems are typically trained on real-world data and refine their output as they receive more data. Common concerns include AI bias, lack of transparency and security challenges. Due to their nature, it’s difficult to determine, validate and understand the source of the information a black box AI system provides. In healthcare, understanding potential sources of AI bias is crucial to ensuring that an AI tool is equitable and accurate. As such, researchers have pushed for additional transparency and auditing frameworks for healthcare AI tools to increase their explainability.

5. Cognitive computing Cognitive computing typically refers to systems that simulate human reasoning and thought processes to augment human cognition. Cognitive computing tools can help aid decision-making and assist humans in solving complex problems by parsing through vast amounts of data and combining information from various sources to suggest solutions. Cognitive computing systems must be able to learn and adapt as inputs change, interact organically with users, remember previous interactions to help define problems, and understand contextual elements to deliver the best possible answer based on available information. To achieve this, these tools use self-learning frameworks, machine learning, deep learning, NLP, speech and object recognition, sentiment analysis and robotics to provide users with real-time analyses. Cognitive computing’s focus on supplementing human decision-making power makes it promising for various healthcare use cases, including patient record summarization and acting as a medical assistant to clinicians.

6. Deep learning Deep learning is a subset of machine learning that analyzes data to mimic how humans process information. Deep learning algorithms use artificial neural networks (ANNs) to imitate the brain’s neural pathways. ANNs use a layered algorithmic architecture, allowing insights to be derived from how data is filtered through each layer and how those layers interact. This enables deep learning tools to extract more complex patterns from data than their simpler AI- and machine learning-based counterparts. ANN inputs and outputs remain independent of one another. Like machine learning models, deep learning algorithms can be supervised, unsupervised or somewhere in between. These are the four main types of deep learning used in healthcare: Deep neural networks. DNNs have a greater depth of layers. The deeper the DNN, the more data translation and analysis tasks can be performed to refine the model’s output.

DNNs have a greater depth of layers. The deeper the DNN, the more data translation and analysis tasks can be performed to refine the model’s output. Convolutional neural networks. CNNs are specifically applicable to visual data. With a CNN, users can evaluate and extract features from images to enhance image classification.

CNNs are specifically applicable to visual data. With a CNN, users can evaluate and extract features from images to enhance image classification. Recurrent neural networks. RNNs generate insights using temporal or sequential data. These networks use information from previous layers’ inputs to influence later inputs and outputs. RNNs are commonly used to address challenges related to NLP, language translation, image recognition and speech captioning. In healthcare, RNNs have the potential to bolster applications such as clinical trial cohort selection.

RNNs generate insights using temporal or sequential data. These networks use information from previous layers’ inputs to influence later inputs and outputs. RNNs are commonly used to address challenges related to NLP, language translation, image recognition and speech captioning. In healthcare, RNNs have the potential to bolster applications such as clinical trial cohort selection. Generative adversarial networks. GANs use multiple neural networks to create synthetic data instead of real-world data. Like other types of GenAI, GANs are popular for voice, video and image generation. GANs can generate synthetic medical images to train diagnostic and predictive analytics-based tools. Recently, deep learning technology has shown promise in improving the diagnostic pathway for brain tumors.

7. Generative AI GenAI tools use a prompt involving text, images, videos or other machine-readable inputs to generate new content. GenAI models are trained on vast data sets to create realistic responses to users’ prompts. GenAI tools typically rely on other AI approaches, like NLP and machine learning, to generate content that reflects the characteristics of the model’s training data. There are multiple types of GenAI, including large language models, GANs, RNNs, variational autoencoders, autoregressive models and transformer models. Since ChatGPT’s release in November 2022, GenAI has garnered significant attention from stakeholders across industries, including healthcare. The technology has demonstrated considerable potential for automating specific administrative tasks. EHR vendors are using GenAI to streamline clinical workflows, health systems are pursuing the technology to optimize revenue cycle management, and payers are investigating how GenAI can improve member experience. On the clinical side, researchers are assessing how GenAI could improve healthcare-associated infection surveillance programs. Despite the excitement around the technology, healthcare stakeholders should be aware that GenAI can exhibit bias like other advanced analytics tools. Additionally, GenAI models can hallucinate by perceiving patterns that are imperceptible to humans or nonexistent, leading the tools to generate nonsensical, inaccurate or false outputs.

8. Machine learning Machine learning is a subset of AI in which algorithms learn from patterns in data without being explicitly trained. Machine learning tools are often used to make predictions about potential future outcomes. Unlike rules-based AI, machine learning techniques can use increased exposure to large, novel data sets to learn and improve their performance. The following machine learning categories benefit healthcare applications in different ways: Supervised learning. This method uses algorithms trained on labeled data — data inputs associated with corresponding outputs — to identify specific patterns. This helps the tool make accurate predictions when presented with new data.

This method uses algorithms trained on labeled data — data inputs associated with corresponding outputs — to identify specific patterns. This helps the tool make accurate predictions when presented with new data. Unsupervised learning. This type of machine learning uses unlabeled data to train algorithms to discover and flag unknown patterns and relationships among data points.

This type of machine learning uses unlabeled data to train algorithms to discover and flag unknown patterns and relationships among data points. Semi-supervised learning. This method relies on a mix of supervised and unsupervised learning approaches during training.

This method relies on a mix of supervised and unsupervised learning approaches during training. Reinforcement learning. This type of machine learning algorithm relies on a feedback loop for algorithm training. It uses labeled data inputs to take various actions, such as making a prediction, and generates an output. If the algorithm’s action and output align with the programmer’s goals, its behavior is reinforced with a reward. In this way, algorithms developed using reinforcement techniques generate data, interact with their environment and learn a series of actions to achieve a desired result. These approaches to pattern recognition make machine learning particularly useful in healthcare applications such as medical imaging and clinical decision support.

9. Natural language processing NLP is a branch of AI concerned with how computers process, understand and manipulate human language in verbal and written forms. Using machine learning and text mining techniques, NLP is often used to convert unstructured language into a structured format for analysis, translate from one language to another, summarize information or answer a user’s queries. There are also two subsets of NLP: Natural language understanding. NLU addresses computer reading comprehension, focusing heavily on determining the meaning of a piece of text. NLU tools use a sentence’s grammatical structure and intended meaning to help establish a structure for how the computer should understand the relationship between words and phrases. As a result, the tool can accurately capture the nuances of human language.

NLU addresses computer reading comprehension, focusing heavily on determining the meaning of a piece of text. NLU tools use a sentence’s grammatical structure and intended meaning to help establish a structure for how the computer should understand the relationship between words and phrases. As a result, the tool can accurately capture the nuances of human language. Natural language generation. NLG helps computers write human-like responses. These tools combine NLP analysis with rules from the output language, such as syntax, lexicons, semantics and morphology, to choose how to phrase a response when prompted. NLG drives GenAI technologies like OpenAI’s ChatGPT. In healthcare, NLP can sift through unstructured data, such as EHRs, to support various use cases. To date, the approach has supported the development of a patient-facing chatbot, helped detect bias in opioid misuse classifiers and flagged contributing factors to patient safety events.

Healthcare’s interest in artificial intelligence shows no obvious signs of slowing down. Major industry events have produced a steady stream of AI products and partnerships, particularly for popular use cases such as clinical documentation, process automation and data aggregation.

A vast majority of healthcare systems use AI, according to the “AI Adoption and Healthcare Report 2024” released late last year by the Healthcare Information and Management Systems Society in partnership with Medscape. The HIMSS survey of IT and medical professionals found that 86% of hospitals and health systems use AI and that 43% had been using the technology for at least one year. That’s a significant increase from a 2022 American Hospital Association survey indicating that just 19% of hospitals were using AI.

Amid excitement about the technology’s future are concerns about implementing AI in healthcare. “There’s a high bar to meet basic administrative of business process needs,” said Dr. Michael E. Matheny, a practicing internist and professor at Vanderbilt University. “It’s important for all of us to consider the use of AI in a careful, measured way to respect the need to support patients and communities.”

Healthcare leaders might not know what successful AI implementation looks like, said Dr. Saurabha Bhatnagar, a Harvard University professor, in a Harvard Medical School Trends in Medicine article. They might assume it’s a matter of acquiring off-the-shelf software, he noted, or they could inadvertently focus on pilots that come with significant upfront costs.

To get over the hump, hospitals and health systems need to think carefully about what AI systems they want to implement, how AI models will be validated and monitored, and who will “own” the AI use case. While every organization’s needs differ, there are 10 steps to implementing AI in healthcare.

1. Map out AI governance The ideal AI governance structure will bring together expertise from IT, data science, the C-suite and bioethics, Matheny conjectured. The governance team will solicit proposals from business owners ranging from nursing leaders to subspecialty chairs to operational administrators. Proposals can come in different forms, said Dr. Brian Anderson, CEO of the Coalition for Health AI (CHAI). Some might propose the use of specific AI tools to solve a problem, while others identify inefficiencies or other challenges AI could solve.

2. Define goals and set expectations The governance team’s first goal in evaluating proposals is ensuring AI is appropriate. In certain cases, rules-based if-then logic could suffice. If AI is the answer, Anderson said, the next step is determining what type of AI is most appropriate, such as generative AI for content creation, deep neural networks for pattern recognition or traditional AI for data analytics. At this juncture, it’s also the time to address the age-old build vs. buy question. Health systems with in-house data science expertise might be able to fine-tune an existing foundational model, Anderson explained, but if expertise is lacking, or if the solution set in question is complex, organizations likely will need to buy AI technologies.

3. Go to the market For health systems looking to buy, Anderson recommended a formal, open request for proposal. He pointed to CHAI’s registry for AI models as a starting point for evaluating vendors and making informed decisions. Business units, he added, need to be part of vendor evaluation and pay particular attention to how well an AI product aligns with the business unit’s unique workflows. Procurement is a critical time for building vendor relationships, Anderson said, to make governance and performance monitoring easier down the line. Follow these 10 best practices when deploying AI in hospitals and health organizations.

4. Ensure data privacy There’s controversy when it comes to using private patient and business data for AI models, Matheny noted. Algorithms need sufficient training data, but anonymizing data lowers its utility. Existing foundational models are widely available, but they share data with their commercial owners. To address this issue, Anderson recommended deploying foundational models locally so organizations can fine-tune models with their own data sets behind their firewalls.

5. Explore use cases with easy wins The most common use cases for AI in healthcare, according to the HIMSS report, are augmenting administrative tasks such as transcribing notes and meetings, reviewing electronic health records (EHRs) and medical literature, and analyzing imaging studies. In the Harvard Trends in Medicine article, Bhatnagar indicated AI tools can be especially helpful for organizations transitioning to value-based care models. With so many variables to monitor, AI can automatically aggregate data and aid in analysis, enabling health system leaders to devote more time to improving outcomes.

6. Solicit stakeholder feedback Both the AI governance group and business owners should convene regularly and discuss how AI tools are being used, Matheny said. Feedback on performance, technical accuracy and clinical efficacy are critical for determining an AI project’s value. This ongoing review should become part of the organization’s culture and structure, Matheny added, “especially as the breadth and depth of AI increases, the need to evaluate it isn’t going to diminish.”

7. Follow ethical standards In summarizing ethical guidelines for AI from the U.S. and Europe, a Harvard Business Review analysis suggested several principles for implementing trustworthy AI in healthcare. In simple terms, AI systems should be safe, algorithms should be unbiased and account for the needs of vulnerable subpopulations, and patients should be informed when AI is being used and have the option to opt out. But bias can be tricky, Anderson noted. Some models are built with “justified bias,” such as a breast cancer risk model for middle-aged black women, who, research has shown, are more likely to be diagnosed with more aggressive forms of cancer and at an earlier age. “You need to continually monitor the performance of these models against their intended use and also against other populations for unjustified bias,” Anderson advised.

8. Validate and monitor models Validating the performance of AI models and the tools that use them should include feedback from clinical and technical stakeholders. Organizations need to ensure products were implemented properly, the model is ingesting the right data and the model’s outputs are meeting expectations, Matheny said. External validation of models could be necessary, he added, if in-house expertise is lacking. At the same time, increasingly sophisticated models could be better suited for internal testing because changes in the way a model’s algorithms are weighed might not translate well to a different institution’s test environment.

9. Provide training and support To help clinical users determine whether they should use AI tools, Anderson proposed two questions: “Is this tool appropriate for the patient I see in front of me? How do I know?” For administrative and operational users, the questions should focus on the process rather than the patient. Organizations should encourage users to think critically and ask questions, Anderson suggested. Transparency is key. Users should have access to models and be aware of how the models were trained.

10. Understand the limitations of AI tools When healthcare professionals encounter mature technologies with rigorous validation, they’re unlikely to question an output. Matheny pointed to blood test results that have well-established and accepted benchmarks. The same can’t be said for many AI tools, so end users need to be reminded that AI is fallible. “It may be wrong sometimes, and you may need to override it,” Matheny said. The consequences of being wrong also need to be considered: An AI system booking two appointments at 11:30 is far different from an AI system prematurely shutting off a patient’s ventilator. AI applications are tangible and wide-ranging in healthcare.