Getting Started With (Gen)AI

AI algorithms implemented in healthcare facilities need to work on a single data platform, according to Cris Ross

Healthcare faces FOMO-AI—fear of missing out on the artificial intelligence revolution. However, as soon as healthcare facilities start to implement algorithms or genAI solutions, their enthusiasm is quickly dampened by data security and privacy issues or a lack of financial capabilities or know-how. Experts advise on how to approach AI in healthcare.

Based on a panel discussion on Leadership and AI at the HIMSS Europe 2024 conference and interviews with: Cris Ross (Chief Information Officer at Mayo Clinic), Brian R. Spisak, (Research Associate at Harvard University), Stéphanie Allassonnière (Associate Professor at the École Polytechnique), Ran Balicer (Chief Innovation Officer at Clalit Health Services), Jakob Nikolas Kather (Professor of Clinical Artificial Intelligence at the Technical University Dresden).

Build capacity for AI leadership

Mayo Clinic is the best benchmark for deploying AI in healthcare. The top-ranked US hospital already utilizes around 200 AI algorithms to optimize workflows and clinical processes.

According to Cris Ross, Chief Information Officer at Mayo Clinic, it’s the last call to prepare for AI. The recent advances in generative artificial intelligence (genAI) and the growing number of AI-powered medical devices mean that every hospital or clinic will have to adopt AI within a few years. “Implementing AI differs radically from implementing a classical IT system, and you need to prepare for it well, ideally starting with creating an AI strategy,” suggests Ross.

Brian R. Spisak, PhD, research associate at the National Preparedness Leadership Initiative and the Harvard T.H. Chan School of Public Health at Harvard University, suggests an approach of “Leadership First, Tech Last.” “Leaders should start by thoroughly understanding their team members’ needs—engage with them, gather insights, and grasp the details of their workflows. At the same time, they must analyze organizational processes to spot strengths, weaknesses, and areas needing improvement,” according to Spisak. Dr. Spisak promotes the term “Computational Leadership,” which entails leaders leveraging:

Established social science to precisely define and continually enhance key factors like patient satisfaction, team engagement, and burnout reduction.
High-quality data for transparent and well-informed decision-making.
Practical experience in crafting context-specific, relevant solutions that align with their team and organization.
Advanced technology to streamline processes, drive innovation, and operate at scale.

Stéphanie Allassonnière, Associate Professor at the École Polytechnique (France), mentions another success factor: AI literacy. “You need technical expertise, medical know-how, and the regulatory viewpoint to enable better data collection, processing, solution development, and transfer. And one more thing: you have to anticipate the development of AI since the speed of innovation in this field is very different from what we have encountered so far,” according to Allassonnière.

“Before implementing AI, build capacity. Train IT teams; otherwise, they will stop every AI project due to a lack of expertise. Train healthcare professionals; otherwise, they will reject every new AI tool due to a lack of understanding.”

Think twice before you use AI

Ran Balicer, PhD, Chief Innovation Officer at Clalit Health Services (Israel), suggests asking two questions: “Why” and “What’s next?” “Why” to see if a particular process cannot be improved by reaching for alternative ways. “What’s next?” because the outcomes of an AI algorithm’s calculations must be followed by concrete actions. Sometimes, it means reimagining the entire patient journey. If you cannot change the procedures that the implementation of the algorithm forces, there is no point in introducing AI.

An example is predictive analytics. If an AI algorithm calculates the probability of deterioration in the intensive care unit (ICU) or the risk of developing sepsis, this information cannot be ignored. New procedures must be designed to verify AI calculations, care pathways, and management. Otherwise, instead of improving patient outcomes, you will suddenly have only inputs (knowledge) without outputs (improvements).

“Spend time figuring out what major organizational needs can be addressed with AI, and whether AI is indeed the method of choice to address them as compared with simpler existing solutions,” according to Balicer.

Integrate AI into EHR

Cris Ross warns against introducing solutions that are not integrated with the current digital ecosystem, preferably with electronic health records (EHRs). Doctors are overburdened by using too much IT. Every single click can negatively impact workflows when repeated dozens or hundreds of times a day. While AI developers declare that their systems make healthcare professionals’ work easier, at the end of the day, someone has to operate these systems.

“AI model building is so 2010. Today, it’s about implementation,” according to Ross. The golden rule should be to embed AI in the EHR. “One system instead of switching between applications, windows, computers, and devices.” Following Mayo’s experience, AI-generated alerts sometimes disturb rather than help, causing “alert fatigue.” Suggestion-based systems work much better.

All AI algorithms need to work on a single data platform. Medical facilities should think about a chat-based interface so doctors or nurses can talk with such an AI system instead of clicking around for data. “Do not increase the burden on doctors with technological innovations, no matter how innovative they are,” Cris Ross stresses, warning against smartphone AI apps, among other things.

“The EU AI act will not inhibit the introduction of artificial intelligence into healthcare. Just as the medical devices regulation has not stopped their development.”
—Marco Marsella, Director for Digital, EU4Health and Health Systems Modernization, European Commission

Don’t wait for better AI, and don’t get trapped by genAI

GenAI has dominated public debates. However, implementing AI should start with simple deep-learning algorithms used in predictive models. From a legal point of view, they are much easier to implement than generative AI.

“Don’t start with generative AI models. Instead, consider machine learning and deep learning solutions first, as these have a longer track record and, generally, a better safety profile,” suggests Balicer. “Generative AI can help extract information from data stored in EHRs, but the present genAI version won’t help develop predictive models.”

The greatest hopes are currently pinned on multimodal AI, which analyzes different data formats, such as medical images, notes, and recordings. It is important to remember that these models are not 100% accurate, but they can be used as supportive tools, for example, for triage.

Think about strategic implementation

AI implementation must be well-planned. Even if doctors report the need for AI, you must ask other medical workers how it will affect their work. If the application of an AI algorithm makes sense, the IT department analyzes whether the data can be integrated into the existing digital ecosystem. Then, the legal department checks how to implement it while complying with current regulations.

Before implementing artificial intelligence, it is necessary to build adaptive capacity by training IT staff and end users, including doctors, nurses, and administrative staff. The implementation will fail if the staff does not see the benefits of using AI.

Not every healthcare facility is a university hospital like Mayo. Assuming spending on IT equals 2-3% of the total budget, most medical facilities cannot afford to create their own algorithms—they will rely on vendors’ AI. This is another argument for creating an AI strategy that includes a plan to integrate new applications into the IT ecosystem.

Prioritize data

“Get your data ready”—Ran Balicer claims that “if you plan to provide patient-specific inferences, your AI-driven output can only be as good as the data it ingests. It is far from trivial to have your patient-level data integrated, clean, and ready for analysis.”

Early innovators implementing AI emphasize that no AI exists without a structured data architecture. New solutions, like command centers in a hospital, cannot be implemented without high-quality data, stored preferably in the cloud.

Jakob Nikolas Kather, Professor of Clinical Artificial Intelligence at the Technical University Dresden (Germany), recommends recognizing real-world data produced by healthcare organizations as one of the biggest assets of healthcare facilities. “This data can then be used to train foundational AI models.”

This mindset is critical to keep up with AI developments. “If you don’t prepare your organization to deploy AI systems, you will face a significant competitive disadvantage in the coming years,” Professor Kather says. Healthcare organizations must be agile to adapt quickly to a changing environment.

“We cannot fully grasp the impact of exponentially increasing AI capabilities, so we must remain open-minded and adaptable as new technologies emerge. And don’t forget good communication to prepare employees for the changing working environment.”

Skills needed for successful AI implementation in healthcare

Ran Balicer, Chief Innovation Officer at Clalit Health Services

Data Management: Proficiency in data management, including data collection, cleaning, and analysis, is fundamental to successfully implementing AI tools.
AI and Data Science Literacy: AI integration requires understanding the fundamentals of AI—machine learning, deep learning, and generative models—and knowing when to use each methodology.
Cybersecurity and Data Privacy: Understanding relevant regulations and implementing patient data privacy and robust data protection measures are essential to maintaining trust and compliance in AI-driven healthcare solutions.
Change Management: AI integration requires preparing, incentivizing, supporting, and helping individuals, teams, and organizations make organizational changes.
Collaboration and Teamwork: AI integration requires interdisciplinary cooperation between IT professionals, healthcare providers, data scientists, and administrative staff.
Empathy and Patient-Centered Care: While AI can enhance efficiency and outcomes, it is crucial to focus on patient needs and preferences and the humane aspects of care provision. In the era of AI, the roles of healthcare professionals are critical to maintaining balance and ensuring that patient needs and concerns are addressed.

Stéphanie Allassonnière, Professor of Applied Mathematics, Vice-President of Valorisation and Industrial Partnerships at Université Paris Cité; Associate Professor at École Polytechnique

Mathematical and Computer Science Expertise: Necessary for model development, implementation, and technical validation.
Medical Data Scientists: They must put the collected data in a usable format to train the models and understand the data, especially if biases and missing data are not random. This interface between the medical doctor and the mathematics researcher is crucial to proposing relevant models in line with the data.
Engineers: Able to develop these solutions into user-friendly tools.
Curiosity and Open-Mindedness: Facing new paradigms and interdisciplinary projects requires curiosity and open-mindedness from everyone involved.
Trust: A single person cannot know all the aspects of the project, and some will be completely obscure. Teams must be built on trust.