A consensus statement on the use of digital twins in medicine

Abstract

Digital Health Technologies represent a marked shift from current medical technologies in use, the approach to health and healthcare and stakeholders engaged in healthcare delivery. What the digitalized future of medicine will look like and how it should be governed is unclear. A participatory process with interdisciplinary expert groups developed scenarios of Artificial Intelligence use in medicine and recommendations on their governance. The process included a patient-consumer focus group and the recommendations were validated by a representative population survey in Switzerland. Digital twins were identified as a pivotal innovation for personalized healthcare, with 62% of the Swiss population expressing interest, though 87% oppose mandatory use. Additionally, 75% view the state as responsible for ensuring necessary infrastructure. Digital twins are seen as an opportunity to support both the healthcare provider as well as patient-consumer directly in different modes of use and along functions, prevention, diagnosis, prognosis, and therapy.

Introduction

In 1994, the prospect of using computers to effectively assist physicians in medical decision-making was considered a rather remote possibility. That year, a seminal paper by Berner and colleagues in the New England Journal of Medicine compared and evaluated the at that time leading automatic decision support systems, with overall poor results leading some to doubt the usefulness of such tools for clinical practice^1–3. The following 20 years saw changes, including the broad introduction of electronic health records across healthcare institutions as well as a plethora of efforts around mobile health apps, direct-to-consumer tests, and other technologies. The resulting availability of large digital and clinical data sets, though, together with new types of artificial intelligence (AI) algorithms published in the 2010s (deep learning algorithms), rapidly changed the outlook for building robust computer tools that can emulate or even rival physicians across diverse medical tasks. In 2017, a group led by Sebastian Thrun showed that computers perform on par with dermatologists in the task of distinguishing images depicting malignant and benign skin lesions⁴. The systems used the latest innovation in AI, convolutional neural networks, and were trained on hundreds of thousands of digital dermatological but also everyday images, the latter being likely important to “pretrain” the system to recognize fundamental and repeating visual shapes and forms. In parallel, the field of Natural Language Processing, which aims at building computer systems to understand free text documents such as newspaper articles but also medical discharge summaries, and which until the advent of deep learning approaches was making steady but only marginal progress, showed massive gains in performance, particularly thanks to a new “Transformer” AI technology first introduced in 2017⁵. Astonishingly, Transformer systems, pretrained on large amount of everyday text (so-called generative pretrained transformers or Generative Pre-trained Transformers (GPTs)), showed emerging capabilities in text understanding, such as providing coherent textual explanations to complex natural language queries. In 2023, the latest versions of these systems were shown to answer medical licensing exam questions at the level or above the average medical student⁶. At this point, it is becoming increasingly clear that computers will touch every aspect of medicine, including complex decision-making, which was seen out of reach for computers 30 years earlier.

This paper is an empirical attempt to foresee such scenarios and changes triggered by the evolution of digital health technologies for different stakeholders across the healthcare system. Subsequently, we derived recommendations on how to deal with these changes and validated them by a representative survey of the Swiss population.

Results

Digital twins will play a key role in the future of healthcare

Four case studies were developed that forecast the changes in healthcare due to AI over time. The digital twin (DT) was identified in the first expert workshop as an important development in the far future for the provision of individualized healthcare across all medical functions and their support. It would integrate other technologies in development and use, such as GPTs.

DT technology is based on simulations of biological processes powering cells, organs, and ultimately, the body. Applied to medicine, a DT would represent an identical in silico replica of a patient, down to its most minute detail. The DT can simulate a patient’s future, identifying a departure from a healthy state towards one with a harmful outcome. To achieve this goal, DTs receive a regular feed of data from their physical counterparts, as is already the case with DTs of drones, buildings, or aircraft engines. A medical DT would thus be provided with health information of its biological twin through sensors that can detect changes in a person’s health state. After integrating the sensor data into its simulation, the DT will perform health forecasting and will suggest preventive measures to maintain health equilibrium in the biological twin. The technological foundation to power medical DTs is already being conceived and developed today. Research groups have established virtual microbial cells, simulations of human cellular processes, and in silico formations of complex multicellular structures^7–9. Other groups have tackled simulations of organ-level processes, such as in silico recreation of heart function¹⁰. In parallel, the computing infrastructure needed for powering DTs is being built. Figure 1 illustrates core characteristics of one of the cases built. The full cases are available as Supplementary Notes 1–4.

Fig. 1. Digital twin scenario for treatment.

Treatment as a medical function may see paradigmatic shifts along key dimensions of medicine and its wider societal contexts. Digital Twins enable fundamental shifts in key dimensions of care provision. This example based on cancer treatment illustrates one of four cases with three points in time (what would currently be possible, what the near future could bring and what the far future could look like) developed.

Patient-consumers have a broad spectrum of requirements

In the following, major themes that emerged during the focus group discussion of scenarios for the future of medicine and the roles of medical professionals are described. They are divided according to the temporal stages of the scenarios in a current state, near-future state, and far-future state.

Regarding today’s scenario, participants see the potential benefits of DT but have concerns centered around autonomy. There was a general optimism that digital means of information-gathering could enable positive effects on health outcomes, but also make the process of healthcare delivery less uncertain, less dependent on physicians. The patient would be empowered to advocate for himself based on technology informing him and find himself in a moderating role between different physicians as opposed to a dependent role. This increased autonomy was also seen as a risk, as some users would be better suited to operate AI than others. Undue influence by other entities on patients employing digital technologies and abuse of data was seen as a possibility that needed to be acted against, e.g., by fine-tuned data access levels and documented access histories. There was agreement that such technologies could only augment and not replace healthcare provided by physicians at this time, and that non-participation and the right to not know should be ensured.

For the scenario concerning the near future, uncertainty increases with DT technology in a transition stage. As a continuation of the key theme of autonomy from the earlier scenario, patient-consumers want to have a say in how much they participate in digitalized models of care and have the option to see a flesh-and-blood physician. This is noticeably different from the much-heralded idea of the digital divide, i.e., being left out without having a choice. Participants re-emphasized this in another aspect: “life style optimization stress” where opting out of nudging/recommendations still should be possible without financial or social detriment. Trust was mentioned in that some entities would be better geared towards operating the technological infrastructure needed with a distinction made between big tech vs. government/healthcare players, as well as temporally limited support vs. continuous monitoring.

In the far future, the benefits and risks of DTs are perceived to unfold fully. Polarization occurred in participants’ opinions of the far future scenario. On one hand, large benefits were seen to materialize to the individual, such as the pre-testing of pharmacological or other interventions to optimize health after the occurrence of disease and fewer reservations were present versus preventative DT use. Inputs to the DT model could be seamlessly collected, providing convenience. On the other hand, societal scale risks were perceived. Quality of life could be affected negatively if life choices were adjusted according to DT suggestions. Individuals may be strived of their trial-and-error approach to life inherent to human development. The DT may alter/hinder this approach and societal consequences would be unclear. Commitments to follow recommendations for (reasonable) incentives may be acceptable, but the incentives may not be designed in a way to tie up and limit the autonomy of the individual.

Stakeholder group-specific priorities and recommendations

During the second expert workshop, a broader spectrum of stakeholders, encompassing diverse groups, actively participated. This led to the identification and definition of four distinct stakeholder perspectives, namely: patient-consumers, healthcare professionals, manufacturers/service providers, and regulators and payors. Within the workshop, each stakeholder group drafted a set of recommendations. The number varied in the groups (up to 14 recommendations), with overlap between the groups, for example, education was mentioned by all groups. After written consensus rounds with the workshop participants, 12 goals and recommendations for the responsible use of DTs in medicine were developed. They were structured in three per stakeholder group (see Table 1).

Table 1.

Recommendations by interdisciplinary expert group

Patient-consumers	Healthcare professionals	Manufacturers/providers	Regulators/payors
Patient-consumers decide on the generation, data sources, design, type of use, and lifetime of their personal digital twin services.	Digital twin services are integrated into interprofessional treatment teams, where the necessary competencies are available and responsibilities are clarified.	Digital twin service providers have access to as much anonymized health data as possible according to open data principles (open standards, interoperability).	The state ensures the provision of a data infrastructure by means of which patient-consumers can bring together data sources from all areas of life.
The relationship of trust between patients and other healthcare stakeholders is preserved through the use of digital twin services.	Healthcare continues to enable care for individuals who do not want to use digital twin services.	Medical services are billed to reimburse for recommendations of high-quality and appropriately certified digital twin services.	The state establishes benchmarks for the quality and security of digital twin services.
Patient-consumers are empowered through measures in education to understand personal digital twins in the best possible way and to be able to use them in accordance with their values and interests.	The setup and operation of a digital twin service infrastructure works internationally, so that location-independent use of digital twins is possible.	Authorization procedures, certification, and regulation of digital twin services, and information requirements for providers of such services are defined and agile.	Certain health-related data generated or made available through digital twin services that are of greatest public health importance will be made available in anonymized form to third parties through this data infrastructure.

Patient-consumers confirm (representative survey)

The answers of N = 1472 persons (51.0% female, 48.6% male, 0.4% other/71.5% German-speaking, 24.5% French-speaking, 4.0% Italian-speaking/mean age: 50.4 years) were analyzed. With respect to the general acceptance of DTs (Fig. 2a), we find that 61.5% of respondents would welcome a DT of themselves—older persons, male persons, and persons with experience in using digital tools (assessed by frequency of using ChatGPT) are more willing to use a DT in the future. The most accepted reasons to use a DT (Fig. 2b) are directly related to potential medical benefits such as “Coordinating medical treatments effectively” (80.9%), “Predicting the course of a disease and initiating countermeasures” (76.4%), or “Detect health risks early” (76.2%); being “No longer dependent on a doctor” (23.3%) receives only low acceptance. With respect to trust in potential providers of DTs (Fig. 2c), we find that “Universities with medical research” (79.0%) and “Public hospitals” (74.6%) are trusted most, whereas “Pharmaceutical/medical technology companies” (7.9%) and “Tech companies” (9.1%) are hardly trusted.

Fig. 2. Digital twin acceptance, reasons to use, and trust in different providers.

a General willingness to use the digital twin is high overall. Older persons, male persons, and persons with experience in using digital tools (assessed by frequency of using ChatGPT) are more likely to be willing to use digital twins. This does not necessarily correlate with the ability to use them, however. b Reasons to use the digital twin are driven by perceived benefit to the individual. Both preventive and corrective aspects of health and healthcare are the focus of personal motivation to use digital twins. c Different entities are trusted to different degrees to provide digital twin services. Public Hospitals and Universities/University Medical Centers are trusted most, Tech and Pharma/MedTech companies least to provide digital twin services and infrastructure. However, the latter are likely to actually develop and market these technologies.

The assessment of arguments in favor or against DTs confirms that they are considered by the population mainly as a new and effective tool for medical experts to provide better prevention, diagnosis and care (Fig. 3). For example, 74.1% agree with the argument that “Thanks to DTs, doctors will be better able to treat their patients”. However, 68.8% also believe that “The use of DTs means that people are forced to share their health data because otherwise they will be excluded from the healthcare system”.

Fig. 3. Agreement to arguments in favor and against using digital twins.

Both potential positive and negative effects digital twins may have are equally considered plausible, pointing to a rather differentiated view of this technology.

Consequently, with respect to our recommendations (Fig. 4), we find that 87.0% agree with the statement “There must be no obligation to use DTs, even if this could result in the individual being treated less favorably” and 79.0% agree with the statement “DTs should be a tool used exclusively by doctors or other healthcare professionals.” 64.1% of the participants agree with the statement “People in Switzerland should make their health data available to researchers free of charge in anonymized form so that better methods for diagnosing and treating diseases can be developed using DTs.” Three-quarters see the responsibility for creating the necessary technical and organizational conditions for the use of DTs by the state.

Fig. 4. Swiss population agreement to recommendations proposed by the expert group.

Patient-consumers agree with governance recommendations by the expert group. A strong majority would oppose mandatory use of this technology—the health system should still offer options that are not digital.

Discussion

The DT is a generic instrument able to integrate different technologies applicable across the disease (and indeed health) spectrum underlying future scenarios of medicine. A significant level of acceptance for DTs as a transformative technology in medicine exists. This finding aligns well with the broader trend of increasing acceptance of AI in healthcare¹¹. However, this acceptance is accompanied by specific conditions and concerns, which are reflected in our study’s recommendations. DTs are widely viewed as tools that should augment healthcare professionals rather than replace them. Respondents emphasize their role in improving fundamental aspects of medical practice, such as coordinating treatment and identifying health risks. However, the perception of DTs as a personal health management tool remains limited. A strong preference for maintaining human medical expertise highlights the need for careful integration of DT technologies within existing healthcare frameworks.

Notably, there is strong opposition to any mandatory implementation of DTs. Individuals value the option of receiving traditional, non-digital healthcare, even if it may not be as optimized as AI-driven alternatives. This preference carries implications for healthcare policy and medical education, as future professionals must be trained to navigate both digital and conventional care pathways¹².

While the potential benefits of DTs are widely acknowledged, concerns regarding privacy and data security remain prevalent. A majority of participants support anonymized health data sharing for medical advancements, but simultaneously express fears about coercion to share personal data. Financial considerations also play a role in public perception. If DTs enhance medical care quality, most respondents find financial compensation for these services acceptable. However, skepticism persists regarding the role of private entities, such as tech and pharmaceutical companies, in developing and deploying DT technology. Instead, trust is primarily placed in public institutions, universities, and hospitals, reinforcing the need for state-led digital health initiatives. This echoes previous data¹³.

Our research highlights distinct priorities among stakeholder groups, which must be carefully balanced. Patient-consumers want to be empowered to make well-informed choices while being able to maintain justified trust towards other healthcare stakeholders; healthcare professionals emphasize the need for the integration of DT services in interprofessional teams and international interoperability while allowing their patients to decline DT services; developers aim for good access to (training) data through open access schemes, reasonable reimbursement for their products as well as for clear regulation that is responsive to new developments; finally, regulators and payors are thinking of a centralized data infrastructure, access to anonymized data in case of significant public health benefit, and benchmarks for quality and security. Interestingly, these goals do not necessarily conflict. There might be tensions—e.g., between patient-consumers and regulators on the question of when the anonymous release of data for public health interests is justified, or between payors and developers on what reimbursement schemes for DT services are appropriate, or between physicians and payors on patients’ ability to opt out of DT services that have been proven to be cost-effective. Still, it seems that being aware of these groups’ key concerns should allow for implementing DT services in a way that does justice to patients’ rights, enables healthcare professionals to provide efficient care of high quality, sets incentives for innovation and sustained interest of companies in the field and contributes to a sustainable, fair health system for all.

We take the findings as support for our recommendations as follows: people are generally positive towards DTs, but they want to keep agency with respect to whether or not to use this technology, which should be enabled. Healthcare professionals are expected to be sufficiently competent in using this technology for the benefit of patients, which should be fostered in curricula. People are, in principle, willing to provide the necessary resources (data, financial compensation models) to develop DTs; but private actors, who may actually develop those technologies, are confronted with considerable skepticism. To counteract this skepticism, the state and public institutions should have a key role in developing a health system that critically relies on DTs; including preserving options that do not need such technologies.

DT technology represents a clear departure from the core idea of today’s healthcare system, which states that physicians and the research system that supports them are the ultimate arbiter not only on how to treat patients, but more importantly, on defining what constitutes a disease in the first place. This conceptualization of disease and treatment rationales is by necessity a human affair, typically the result of consensus-finding processes among physicians organized in specialist societies. While data-driven approaches, particular evidence-based medicine, are attempts to objectivize the process, a large swath of medicine is still a very human intuition-driven enterprise with vast differences in available care pathways and ultimately health outcomes¹⁴. The core of this enterprise is an international, national, regional, or individual understanding of disease and treatment strategies. By necessity, the AI tools that are being introduced today reflect this situation, that is, the AI tools, trained on current medical data, are a mere reflection of current medical concepts, definitions, and approaches^15,16. Constrained as such, current AI cannot evolve beyond what is happening in medicine today. In contrast, DT technology, in its most advanced state, applies a predominately data-driven approach for predicting a future state given a patient’s current health data. At its core, a DT should “know” the principles of how processes at different scales evolve over time, but it will not be constrained by them. Rather, DTs will use machine-derived states to trigger corrective actions in the form of health interventions. It is unclear at this point whether humans will be able to understand the meaning of these states, and whether these states can be translated into current medical concepts. This begs the question: if DTs will emerge as a key tool in healthcare, in the form of personal digital replicas guiding our treatment choices, what is the role of the traditional bearer of medical knowledge and wisdom, the physician? Can physicians remain the epistemic authority in healthcare, given that they operate using a scheme of thought that may no longer allow them to understand machine predictions of a patient’s health state? How should patients deal with the technology, both in the role of data providers and consumers of DT services? Is there an emergent role for technology companies that build and maintain digital medical twins? And finally, what are sensible regulatory approaches to ensure fair and equitable access to the technology? Our paper provides possible answers to these questions, boldly proposing a future healthcare system with new stakeholders, such as tech companies and novel roles of traditional players like physicians, but also contrasting the current perspective of the population versus a different outlook on technological development. Realistically, changes will appear gradually, as will be the capabilities of the DT technology, which, in its first incarnation, will likely be built on today’s AI capabilities using only little multi-scale simulations. The paper acknowledges this fact by discussing the near-, mid- and long-term effects of DT technology on the healthcare system.

The methods employed in this study have inherent and study-specific limitations. Although we attempted to have multidisciplinary diversity in our 20-member expert group that forecasted future scenarios of medicine, they ultimately represent a group hailing from academic and/or tertiary-level care backgrounds of Zurich universities. This may introduce bias, and therefore, the scenarios may have been different if we had attempted to include other groups, such as primary care physicians or patient-consumers. Similarly, although the second 22-member expert workshop diversified the group to include non-academic/non-physician stakeholders, the resulting policy recommendations may still suffer from selection bias. Although invited, patient representatives unfortunately canceled their participation in the second workshop, which represents a limitation. While the presence of a patient representative would have strengthened the second expert workshop, substantial input was provided by other participants taking the perspective of patients. We also remind that the patient's view is included in the process earlier through the focus group that consisted of persons with patient/chronic disease experiences. A limitation here is that the six German-speaking patient-consumers only represent a convenience sample and the issues raised may not be representative of the wider Swiss population. This limitation was partially mitigated by the representative survey of the Swiss population that included German-, French-, and Italian-speaking Swiss. Last but not least, the findings of this research are based on Swiss perspectives and may not apply to other healthcare systems.

Methods

Aims of study and overall approach

The aims of the study are (1) to identify likely future directions of the practice of healthcare expressed in case scenarios, (2) to assess patient-consumer attitudes towards these scenario practices, (3) to identify required policy actions and derive recommendations in anticipation and (4) to assess public opinion in regard to the identified future practices and policy recommendations. To fulfill these aims, a sequential mixed-method approach was chosen, consisting of two expert workshops to develop scenarios, a focus group with patient-consumers (six persons; all with chronic disease experiences), a second expert workshop to develop recommendations on governance, and a representative survey of the Swiss population to validate these. Expert workshops were inspired by a Delphi method-based approach (12 of 20, respectively 22 workshop participants participated in both workshops; all 22 experts of the second workshop were invited to provide feedback to the final recommendations) to allow for collaborative refinement. No ethics approval for this study was necessary under applicable Swiss law. The internal ethics and data protection policies of the university unit were reviewed and followed.

Case scenarios derived by an academic and clinician expert workshop

An interdisciplinary full-day workshop of twenty researchers developed case-based scenarios for the future of medicine. Three medical functions and one support process were pre-chosen and cases developed during the workshop: prevention, diagnosis, therapy, and resource allocation. The cases included three distinct points in time: a scenario that would be possible now, in the near future (around ten years from now), and far future (around 25 years from now). No prior content scope was defined beyond disease areas, oncology, and depression, that were chosen by group poll at the beginning of the workshop. The participants were drawn from the University of Zurich Faculties of medicine, economics and informatics, science, philosophy, and law, and the four University Hospitals (University Hospital Zurich, Psychiatric University Hospital, University Dental Clinic, and Balgrist Orthopedic University Hospital). We aimed at recruiting professorial-level clinicians from surgical and nonsurgical medicine, psychiatry, and dentistry, as well as professorial-level non-clinicians from other backgrounds to ensure interdisciplinarity. Workshop participants rotated through 3 of 4 case stations, ensuring breadth of thought. One moderator per table ensured continuity and transfer of information. Case scenarios were then presented in plenum and subsequently circulated in writing, edited, and agreed for release. The case scenarios are offered as Supplementary Notes 1–4.

Patient-consumer concerns on future scenarios by focus group

A focus group-based approach was chosen to assess attitudes by patient-consumers towards these scenario practices. Focus groups are a qualitative method that aims to collect information that may not be easily collected in structured individual interviews or surveys. They offer the opportunity to yield deep insights into typically implicit reasoning and attitudes. A 2-h focus group held in German took place in September 2022 with six participants (1 female, 5 male) across age and political spectrums that were recruited through an external provider (gfs.bern). Hard criteria for inclusion were German language skills, experience as patients with and without chronic disease and legal age (18+). We also aimed at a diverse group in terms of age (at least one pensioner), political orientation (both left, center, and right), digital skills, and gender. Written informed consent was obtained and no conflicts of interest were declared. Three temporal scenarios were presented with a focus on oncology. The transcript was then single-coded in two subsequent rounds using qualitative analysis software VERBI MaxQDA 22.

Policy recommendations by wider stakeholder expert workshop

Another full-day workshop with 22 participants from both clinical and non-clinical academic backgrounds, as well as representatives of pan-Swiss stakeholder groups (physicians, digital business, med tech and pharma industry, consulting, and regulators), discussed the case scenarios developed earlier. We aimed at including key stakeholder groups required for the acceptance and implementation of digital health interventions into clinical practice. While we had also invited patient representatives, they unfortunately canceled their participation. Given a dedicated focus group with patient-consumers, those perspectives have been incorporated, however. Four groups taking the perspectives of “patients and consumers”, “healthcare providers”, “payors”, and “industry” were defined and participants assigned according to their individual backgrounds. Two rounds of discussion were held—centering first around expectations and limitations for those scenarios and subsequently deriving recommendations to deal with them. Moderators at each table summarized and documented the findings and presented them in plenum, where they were discussed. Drafts of the recommendation text were subsequently circulated and all workshop participants were invited to provide feedback; 11 experts provided input to the draft. The text was finalized in five editorial iterations with the experts and then later in the core editorial team (five persons).

Survey validation of recommendations in the Swiss population

To assess attitudes and opinions of the Swiss population both with respect to the identified core technology in general as well as with respect to our recommendations, we conducted a representative population survey from July 26 to August 22, 2023. The trilingual questionnaire (German, French, and Italian) was pretested in several rounds (both qualitative and quantitative pretesting) in cooperation with gfs.bern and implemented using Qualtrics XM. The questionnaire (available as Supplementary Note 5) was structured as follows: after providing informed consent, the study participants provided demographic information (gender, age, language, place of residence, education, and political orientation). In a second part, the participants answered questions to assess their digital knowledge and skills. In the third part, participants were made familiar with the concept of the key technology through an explanatory video; correct understanding of the concept was checked with two test questions; persons who answered those questions incorrectly were excluded from the analysis. After the video and the test questions, we assessed willingness, expectations, and reasons to use the technology, trust in potential providers, acceptance of arguments in favor or against, as well as acceptance of statements that relate to our recommendations. The study results have been weighted with respect to age, education, gender, language, and political orientation to match the statistical properties of the general Swiss population with respect to those parameters.

Author contributions

Conceptualization: J.D.I., M.K., C.W., N.B.A., M.C.; Methodology—data acquisition: J.D.I., M.K., C.W., N.B.A., M.C.; Formal analysis and investigation: J.D.I., M.K., C.W., N.B.A., M.C.; Interpretation of results: J.D.I., M.K., C.W., N.B.A., M.C.; Writing—original draft preparation: J.D.I., M.K., N.B.A., M.C.; Writing—review and editing: J.D.I., M.K., C.W., N.B.A., M.C.; Funding acquisition: M.K., C.W., N.B.A., M.C.; Supervision: M.K., C.W., N.B.A., M.C.; All authors have full access to all the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis. All authors gave final approval.

Data availability

Focus group data used to construct the analyses are not publicly available because of Swiss laws guarding the personal integrity of its citizens. All other data can be made available upon reasonable request.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41746-025-01897-4.