The challenges of scaling up clinical AI

November 16, 2026
The challenges of scaling up clinical AI
AI
News

Artificial intelligence offers an ever-increasing range of possibilities for supporting healthcare professionals in diagnostics, treatment, and administrative processes. Yet the transition from promising AI applications to routine use in hospitals remains complex. It is not only the technology, but above all the availability of reliable data, the design of workflows, regulations, and digital skills that determine whether AI actually delivers added value. This became clear during a meeting organized by the AI company Kaiko, where healthcare administrators, doctors, developers, and legal experts exchanged experiences regarding the implementation of clinical AI.

The meeting was part of the Zurich AI Festival. The central question was how hospitals can integrate AI into daily patient care without individual applications leading to the fragmentation of systems, data, and responsibilities. The experiences of organizations such as the Swiss Insel Gruppe and the Dutch Antoni van Leeuwenhoek Hospital show that successful implementation requires a combination of technological infrastructure, clinical engagement, and clear decision-making.

Digital foundation is a priority

According to Michael Stickel, Chief Medical Information Officer (CMIO) of the Insel Gruppe in Bern, hospitals must start with a cohesive digital infrastructure. He advocates integrating AI into the electronic health record (EHR) as much as possible, rather than introducing separate systems for each application. An integrated approach can simplify the management of data, access rights, oversight, and contracts. Stickel pointed out that approval of a single AI platform by the cantonal privacy authority can take more than two years. With separate applications, similar procedures may be required again.

At the same time, he acknowledges that integration into the EPD is not feasible for every application. Hospitals must therefore balance the need for a centralized infrastructure against the flexibility to use specialized solutions.

A second challenge is selecting suitable applications. According to Stickel, hospitals can choose from hundreds of potential AI applications, but often lack a clear method for setting priorities. At the Insel Gruppe, proposals are therefore evaluated based on six criteria: clinical added value, improvement of work processes, patient safety, direct use by patients, legal risks, and financial implications. Before an application is rolled out more widely, it undergoes a data protection and risk assessment. Subsequently, a few trained healthcare professionals test the technology in practice.

AI makes tumor assessment more accurate

A concrete example of clinical AI was presented by Kevin Groot Lipman, AI Implementation Lead at the Antoni van Leeuwenhoek Hospital and Kaiko. He discussed the assessment of treatment outcomes in patients with pleural mesothelioma, a form of cancer in which tumors develop around the lungs. Due to the irregular shape of these tumors, it is difficult to determine whether a treatment is sufficiently effective using traditional measurements. The standard assessment includes examining changes in tumor diameter. A 20 percent increase can serve as an important threshold for assessing disease progression.

According to Groot Lipman, however, the choice of measurement location can lead to different outcomes. In a comparison he cited, two medical experts reached the same assessment in only 53 percent of cases. His research team therefore developed an AI method that not only measures individual diameters but also determines the total tumor volume. The model analyzes CT images and marks the pixels that belong to the tumor. According to Groot Lipman, the AI can perform this segmentation within a minute, after which the volume measurement is available to the clinician in about fifteen minutes.

The application illustrates how AI can supplement existing diagnostic measurement methods with more detailed information. Whether this approach also leads to better treatment decisions and patient outcomes must be determined separately.

Healthcare professionals must remain involved

The meeting also addressed the organization of AI within hospitals. A key point of consideration is who determines which applications are implemented and who is responsible for their use. According to the participating experts, healthcare professionals must play a central role in these decisions. This includes not only physicians but also nurses, medical coders, and other staff who work with the same patient data.

A joint assessment can prevent departments from independently purchasing applications that are not sufficiently compatible with one another. In addition, training is a key requirement. Stickel considers insufficient AI skills to be one of the biggest bottlenecks in implementation. Within his organization, employees therefore only gain access to an AI application after they have completed the corresponding training. This approach underscores that digital transformation requires not only investments in software but also the development of knowledge and skills within healthcare organizations.

Regulation and responsibility

The legal aspects of clinical AI pose an additional challenge. During a panel discussion, it was emphasized that doctors remain responsible for their medical decisions, even when using AI. Legal expert Thomas Kuster pointed out that professional standards may change when new technology demonstrably leads to better patient outcomes. As AI becomes part of generally accepted medical practice, he believes the responsibility of healthcare providers may also change. That does not mean that every new AI application must automatically be used. The clinical added value and safety must first be sufficiently demonstrated.

The relationship between data protection and information security was also discussed. According to Kuster, strict privacy requirements can sometimes hinder the deployment of technically secure solutions. He therefore called for a broader discussion on balancing these two interests. Another point of concern is dependence on vendors. AI that is directly integrated into the EHR can be scaled up relatively easily, but at the same time can make hospitals more dependent on a single technology provider.

Various real-world experiences show that the implementation of clinical AI cannot be reduced to simply purchasing new software. Hospitals must make decisions regarding data infrastructure, collaboration, training, and responsibilities. In this context, it is not the number of AI applications implemented that is decisive, but whether they demonstrably contribute to better, safer, and more effective patient care.

The full event can be viewed on the Kaiko website.

Add ICT&health on Google

Show more content from ICT&health in Google Search.

Add on Google

This topic will also have a prominent place at the ICT&health World Conference 2027. Want to be there and stay ahead of what’s next in healthcare? Reserve your ticket today.