The advancement and development of artificial intelligence (AI) in health care hold several promises, such as the heightened quality of diagnosis and treatment [1-4], improvements for clinical workflow processes [5], and a reduction of costs [3,6]. These hopes also apply to health systems’ reactions to global health emergencies, such as the COVID-19 pandemic or future challenges [7]. For the purpose of this research, we used the definition by He et al [8], who define AI as a “branch of applied computer science wherein computer algorithms are trained to perform tasks typically associated with human intelligence” [8].

Although AI development and research are conducted in many countries nowadays [5], a report from the Joint Research Centre of the European Commission [9], as well as original research, shows that AI development and implementation in the European Union and in Germany are still in their early stages [10,11]. In part, this could be explained by the low availability of the data needed for the development of AI tools [12,13] and by regulatory and legal uncertainties [10]. In its efforts to support the further development of AI tools within the country, the German government deemed the following aspects as especially important: (1) data sovereignty, (2) patients‘ protection-worthy interests, (3) patients‘ rights, and (4) compliance with ethical requirements for the protection of sensitive health data [14].

Both patients’ interests and their rights concerning the use of their routine medical data have been studied before. In a systematic review, Aitken et al [15] synthesized the results of qualitative research on data sharing for the purpose of health research. They identified overall widespread conditional support for this purpose. The conditions for support included, inter alia, the assurance of individuals’ confidentiality, a preference for the anonymity of data, and assurances of data security [15]. However, it is unclear whether these results and conditions can be applied to AI—a fundamentally new general-purpose technology. A study by McCradden et al [16] aimed to generate insights into public perceptions on using health data for AI research. They concluded that the general views about AI in their studied sample of the general population were mostly negative. Still, the participants were able to describe the potential benefits for health research that could be gained through the use of AI. Important conditions for supporting the use of health data for AI research were consent, the transparency of AI use, and assurances of data privacy [16]. These findings were based on a sample from the general population in Canada. It is plausible to assume that these preferences and opinions could vary among different cultural contexts and between the general population and current patients. Hence, we decided to involve patients in the first part of our project, even though they will not be direct users or beneficiaries of our generic infrastructure. To facilitate later discussions on patient and public involvement in research, the Guidance for Reporting Involvement of Patients and the Public [17] framework will be used.

The availability of large quantities of structured, high-quality data is fundamental to the development of AI tools [4,18]. Hospitals and other health care facilities inherently store the large amounts of data needed for AI development, but access to these data can pose legal and ethical challenges [12,13]. Besides that, legal certainty for the (partly) automized testing of routine medical data against a defined gold standard is missing [19]. Therefore, databases for the development and testing of new AI algorithms are often assembled manually. This requires high efforts from hospitals and product developers and slows down development and evaluation processes. In turn, patients can only benefit from new innovations after significant delays.

These problems and questions are addressed by the Protected Artificial Intelligence Innovation Environment for Patient-Oriented Digital Health Solutions for Developing, Testing, and Evidence-Based Evaluation of Clinical Value (pAItient) project, which aims to establish the AI Innovation Environment as a proof-of-concept extension of the already existing Medical Data Integration Center [20] at the Heidelberg University Hospital. The project partners include the German Cancer Research Center, the German Research Center for Artificial Intelligence, and Mint Medical GmbH (Heidelberg, Germany).

We also propose that knowledge of the requirements and needs of stakeholders, such as patients, health care providers, and industry partners, regarding the planned, generic infrastructure is an important antecedent for high acceptance and usefulness. This infrastructure is a novel concept and thus warrants a close analysis of factors influencing stakeholder acceptance.

This paper presents the protocol for our study, which aims to explore stakeholders’ requirements for developing AI in partnership with an academic hospital and granting AI experts access to anonymized personal health data.

Study participants were recruited from the following six stakeholder groups: (1) researchers from a biomedical research institute, (2) researchers from an AI research institute, (3) employees from start-up companies in the field of AI development, (4) employees from an AI imaging company, (5) patients at the Heidelberg University Hospital, and (6) physicians actively working in inpatient or outpatient health care.

Potential study participants from groups 1 to 4 were invited via email, following a snowball sampling approach. Patients (group 5) were recruited through a purposive sampling approach. Recruitment took place in different departments (the Department of Obstetrics and Gynecology, the Department of Internal Medicine, and the outpatient department at the National Center for Tumor Diseases Heidelberg) at the Heidelberg University Hospital, which is one of Europe’s largest medical centers [21]. Patients were approached by a study team member and invited to participate. In addition, informational materials, such as leaflets containing information on the study and the contact data of the study team, were provided. Recruitment procedures for physicians (group 6) are described elsewhere [22].

The inclusion criteria for groups 1 to 4 and group 6 were employment or the conduction of research activities in one of the defined areas. To be able to participate, patients had to report either a self-defined chronic disease or at least 3 visits to the hospital or an ambulatory health care provider in the last 3 months. The exclusion criteria were persons aged under 18 years and persons whose command of German or English was not sufficient for conducting an interview.

The following paragraphs describe the qualitative and quantitative methods that were used for data collection (Figure 1 depicts an overview). The methods used for data analysis are described in their respective paragraphs.

The design for the stakeholder requirements analysis study was approved by the Ethics Committee of the Heidelberg University Hospital (approval number: S-241/2021) in March 2021.