To identify the defining features of AI-based health care services as well as their corresponding archetypes, we formulated a bilateral methodological approach to develop our taxonomy in accordance with both the guidelines of Nickerson et al [24] and their extension by Kundisch et al [25]. In doing so, we achieved our dual aim of creating a versatile set of tools for the analysis of both conceptual and empirical observations. Taxonomies comprise various layers, which in turn include multiple dimensions, which can again contain diverse characteristics [24]. They are frequently used (eg, Fischer et al [26] and Gimpel et al [27] to show how different concepts are connected or to investigate their relationships [28].

As proposed by Kundisch et al [25], we began our taxonomy development by specifying the central phenomenon under investigation, that is, AI-based health care services. To pay due attention to all of their various design characteristics as well as those of our 2 target groups—researchers examining the field of AI-based health care services and practitioners in the health care sector who are already integrating or aim to integrate AI in their service offerings—we define the meta-characteristic [24], that is, the design characteristics of AI-based health care services.

In line with the best practices of Nickerson et al [24], we then define our objective and subjective ending conditions, both of which are checked after each iteration to assess whether the taxonomy development process is completed [24]. Nickerson et al [24] identified a set of ending conditions that can be applied or adapted to the taxonomy development process, four of which we selected for this study: (1) all objects of our dataset have been examined, (2) no new dimension or characteristics were added in the last iteration, (3) every dimension is unique and not repeated, and (4) at least one object from our dataset is classified as having every characteristic of every dimension [24]. Furthermore, we concur with Nickerson et al [24] regarding the subjective ending conditions in that the taxonomy should be concise, robust, comprehensive, extendible, and explanatory [24].

Our taxonomy development process consists of 4 iterations. Nickerson et al [24] propose 2 development approaches to this development. First, inductive, which is to say E2C, and second, deductive, which is to say C2E [24]. These 2 approaches are to be executed consecutively in several iterations. Table 1 depicts details on this taxonomy development process, including the development approach chosen for each iteration, the objects in question, the significant changes, and the evaluation of the ending conditions.

In our first iteration, we chose an E2C approach since this is generally agreed to be the best practice when a significant amount of data can be used to examine a specific research field. To collect the necessary empirical data when making our selection of AI-based health care services, we followed a similar approach to that of Labes et al [29] and Fischer et al [26]. In doing so, we were aware of the challenge of keeping up with the fast-changing field of AI in health care, which is why we identified so-called “top lists” that consolidate the most promising AI companies in health care. After compiling an initial list with the help of several internet search engines, using terms such as “top AI health care companies,” “ranking AI companies in health care,” and “top list health care and AI,” we synthesized our results into a single integrated list of AI-based health care services derived from 6 data sources. Our sample only included services that describe an activity involving at least 2 entities with different roles, for example, a patient and a health care professional, where both use their resources, be it information or technology [30], in a collaborative process that serves their mutual benefit. Furthermore, it was not our aim to compile a complete list of all existing health care start-ups that use AI in their portfolio. Instead, we wanted to collect a sample that was as representative and generalizable as possible to serve as a basis for the taxonomy development. For this reason, we used 6 different top lists as a data source, which are based on various evaluation criteria such as growth, funding, and global impact. Furthermore, we use top lists from the years 2020 and 2023 to also cover a temporal variance. This ensures that our sample reflects sufficient variance and can be considered representative. The final list comprised 268 AI-based health care services. Distinct subsamples were used in iterations 1, 3, and 4. Tables 2 and 3 illustrate the data sources as well as exemplary services.

In the first iteration, we analyzed a random sample of 25 AI-based health care services to derive relevant dimensions and characteristics. In the second iteration, we followed a C2E approach informed by the insights of a structured literature review [60]; Tables 2 and 3 contain the details. With this 2D approach to data collection from practical and academic data sources, we were able to account for the multiple developments in the rapidly expanding field of AI-based health care services. To broaden our understanding accordingly and heed the lessons learned from the research already conducted, we performed a literature review on the Web of Science. We used an iterative search process for the literature review. Initially, we had a very narrow search string to gain an initial understanding of the topic, which we later expanded again to ensure a more comprehensive literature review. For the final search, we limited our search to papers published between 2015 and 2023 in the English language and excluded book chapters or proceeding papers. By using the exact search string “(“artificial intelligence” OR “AI” OR “machine learning” OR “deep learning”) AND (“healthcare” OR “health care” OR “clinical” OR “medicine”) AND (“service” OR “application”),” we were able to exclude any literature irrelevant to our study. This left us with 9318 relevant results. Furthermore, 2 members of the research team (MB and PK) then analyzed the title and abstract of each paper. After that, all those who did not directly address the topic under investigation were excluded. Since reliability was a guiding principle of this process, the team screened and classified a random sample of 30 papers in close collaboration. After that, 30 papers were screened independently, and their classifications were compared to certify the analytical adequacy. All subsequent papers were screened individually by one of the team members. Speaking in numbers, this meant that from the initial set of 9318 papers, 276 had their full text analyzed, and 20 of these were considered for our taxonomy. In addition, 8 papers were included through forward and backward searches, which brought the total to 28 papers.

For the third and fourth iterations, we again followed the E2C approach with 25 randomly selected AI-based health care services and 218 remaining from our initial sample. After the fourth iteration, the team agreed that the objective and subjective ending conditions were all met.

As Kundisch et al [25] pointed out, a rigorous demonstration and evaluation of a newly developed taxonomy require a 2-step approach. In the first step, 13 researchers with a domain-specific background in AI or health care, or indeed both, classified multiple real-world services by using our taxonomy. These demographics were selected because they constitute the intended future users of our taxonomy. We measured the agreement among the individual classifications by means of the percentage agreement and Fleiss κ [61]. The latter was chosen by virtue of the fact that it can be used to determine the degree of agreement between the assessments of more than 2 raters, which is to say that it works as an extension to Cohen κ [62], provided that the number of raters per assessed unit remains constant (described in detail in “Taxonomy Evaluation” in Multimedia Appendix 1). The unweighted κ [61] is determined, that is, each nonmatch is weighted equally. The calculation of κ [61] () results from the calculation of the relative agreement of the raters () and the calculation of the probability of random agreement ().

n describes the number of raters (n=13), N is the number of considered service offerings per rater, and k is the number of possible evaluation categories. Aside from being a well-established procedure to evaluate all manner of taxonomies, it allowed us to ensure that classifications made with our taxonomy are not too susceptible to personal bias [27,63].

In the second step, we classified all available services to learn more about the occurrence rate of each characteristic in our dataset. This was followed by cluster analysis, a statistical method of similarity analysis performed to identify groups of similar objects, so-called clusters, within a dataset [64-66]. The matching objects are grouped in such a way as to achieve high homogeneity within clusters and high heterogeneity between clusters [67]. In our case, this cluster analysis allowed us to identify archetypes of AI-based health care services by observing the three necessary specifications: (1) a proximity measure to calculate similarities or distances between objects, (2) a clustering method with underlying grouping methods or fusion algorithms, and (3) the order in which the clusters are split or merged. It is worth noting, however, that the choice of the clustering procedure is closely related to the question of the optimal number of clusters. To determine an appropriate number of clusters for AI-based health care services, we calculated several measures discussed in the literature (described in detail in “Clustering” in Multimedia Appendix 1). For the various numbers of clusters, each analysis is performed with an agglomerative hierarchical method that creates a corresponding cluster solution for any number of clusters between 1 and n, where n provides the number of services considered in the sample [68,69]. To this end, we chose the Ward algorithm [70] as our clustering algorithm and used the Manhattan distance as our distance measure, seeing as the latter can be applied to nominally scaled data and has proven to perform well in connection with the Ward algorithm [70,71]. To ensure the correct application of the distance measure, we first dichotomized and then standardized the underlying data by accounting for the number of characteristics per dimension, thus also ensuring that all dimensions are equally weighted.

This study did not require approval from an ethics review board, as it does not constitute human participants research. In accordance with the guidelines of the University of Hohenheim, ethical review is mandated only for research involving human participants, biological samples, or identifiable personal data. The focus of this study was on identifying the design characteristics of AI-based health care services, relying primarily on a review of existing literature and analysis of established AI-based health care services. Data from individuals were collected solely for the purpose of evaluating the proposed taxonomy, and as such, the study does not meet the criteria for human participants research as defined by the relevant policies of the University of Hohenheim (Satzung der Ethikkommission der Universität Hohenheim).

To validate the robustness of classifications made with our taxonomy, we recruited 13 researchers in the fields of AI, health care, or both. The overall demonstration phase took 3 weeks and was analyzed by 2 of the authors (MB and PK). We tasked the 13 researchers to classify 4 service offerings from our dataset anonymously. In choosing the service offerings for the application of our taxonomy, we made sure to cover a wide range of characteristics to provide a holistic evaluation of the taxonomy. Each service provider supplied the participating researchers with a brief description of their service along with a few images to illustrate how it works. As a further quality control measure, we made sure that there could be no confusion about the task assigned to the participants by giving them descriptions of all the dimensions and characteristics that they would have to consider when using the taxonomy, and we tested the comprehensibility of the questions as well as the appropriateness of the supplementary texts in advance. As a result of these measures, the percentage agreement among all authors and concerning all dimensions was an excellent 93.4%. Complementarily, the overall percentage agreement among all participants was 88% [72]. Fleiss κ is 62%, indicating a “substantial” agreement among participants [73]. For a detailed overview of percentage agreement and Fleiss κ across all individual dimensions, see “Taxonomy Evaluation” in Multimedia Appendix 1 [61]. It is suffice it to say that we conclude the taxonomy to be concise, robust, comprehensive, extendible, and explanatory [24].

In the following phase, we classified all 268 health care services to reveal the absolute and relative frequencies of all characteristics. When analyzing the relative frequencies, as presented in Table 5, some significant observations can be made, such as the health care professional is a direct recipient in 70% (n=186) of services, and the service recipient triggers the exchange of information in 79% (n=211) of services. Also noteworthy is that 74% (n=198) of data generators are objects, such as wearable devices. In 95% (254) of services, the primary data target is the patient, while 70% (n=188) of the data underlying the service are unstructured. As for the role of AI, 84% (n=223) of services are hardware agnostic, which is to say that they do not require a specific device to deploy the service. Also of interest in this context is the statistic that 83% (n=222) of the services in our sample were offered as a stand-alone solution.

A deep dive into AI capabilities revealed that 96% (n=257) of services can recognize patterns within audio, video, or other health-related data. At 33% (n=89), the second most frequent AI capability is to generate something new based on available input data. Examples are structured consultation protocols based on sound recordings or descriptions of injuries based on image files. Approximately 28% (n=75) of observed services include the AI capability to perform decision-making, for instance, to decide between multiple treatment alternatives. Another 13% (n=34) of observed services have the AI capability of reasoning, which offers the medical professional the significant benefit of delegating the task of explaining the link between bad habits and potential diseases. The rarest observed AI capabilities were acting (n=30, 11%), for example, calling emergency staff for help, and predicting (n=28, 10%), for example, being able to tell in advance whether people with specific characteristics will develop a particular disease. The final statistic of note is that most of our sample services are used for diagnostics (n=163, 61%). Second in demand (n=83, 31%) is the type of service used in treatment and care. At 87% (n=231), the vast majority of our sample services are used to treat physical illness or injury. In comparison, 9% (n=25) are used for medical triage. Only a minority of our sample services use AI to improve mental (n=18, 7%) or social (n=4, 1%) health.

To ensure that our taxonomy can identify separable archetypes despite its extensive range, we performed the cluster analysis analogous to Gimpel et al [27], which is to say that we performed it for each of the identified dimensions separately. This cluster analysis revealed 13 archetypes of AI-based health care services that differ in their dominant design characteristics. We identified 2 archetypes in the agent perspective, 5 archetypes in the data perspective, 2 archetypes in the AI perspective, and 4 archetypes in the health impact perspective. Table 6 presents the respective number of services that can be assigned to each archetype as well as the absolute and relative frequencies of their characteristics.

First, the agent perspective includes the 2 archetypes, “AI-initiated” and “recipient-initiated.” Service offerings of the AI-initiated archetype actively request information from the user. They are designed for use by both the patient (27/57, 47%) and the health care professional (35/57, 61%). In contrast, the recipient-initiated archetype targets the medical professional in the majority of cases (152/211, 72%).

Second, the data perspective includes 5 archetypes. These mainly differ from one another in the origin and the type of the used data. What the archetypes have in common, however, is that their data are primarily related to the patient. The “structured object” archetype is characterized by the use of structured data (100%) that are provided by an object (100%). Almost one-quarter of services also use the patient (8/44, 18%) or medical professionals (3/44, 7%) as a source of information. In contrast, services in the “unstructured object” archetype work with unstructured data (100%) that is primarily available in audio or image formats. Occasionally, this data include information about the patient’s environment (13/145, 9%) but rarely about the responsible medical personnel (1/145, 1%). The “structured nonobject” archetype contains structured data concerning the patient (100%). These data are primarily provided by the patient (32/34, 94%) and less so by the treating physician (6/34, 18%). Services of the “unstructured nonobject” archetype work exclusively with unstructured data (100%). However, compared to the “structured nonobject” archetype, they are almost twice as likely (9/28, 32%) to be generated by the responsible health care professional. As for the “structured-unstructured” archetype, this one is characterized by the simultaneous processing of structured and unstructured data (100%) that are obtained from objects (12/17, 71%), health care professionals (8/17, 47%), and patients (7/17, 41%).

Third, the AI perspective comprises the 2 archetypes, “device-dependent” and “device-independent.” The former includes the majority of services in the sample (222/268, 82.8%) and operates with generic hardware. Services belonging to this archetype are primarily offered as stand-alone solutions (206/222, 93%). The underlying AI can recognize patterns and other correlations (215/222, 97%), decide between separate alternatives in approximately one-third of the cases (64/222, 29%), or generate output (79/222, 36%). To a lesser extent, it is also able to provide reasons for correlations (27/222, 12%), make predictions (23/222, 10%), and execute actions on its own (24/222, 11%). Services of the “device-dependent” archetype require the use of application-specific hardware. If supported by such hardware, this archetype can recognize contexts in a similar number of cases (44/46, 96%) as the previous archetype, but it makes decisions (10/46, 22%) or generates objects (10/46, 36%) in far fewer cases.

Fourth, the health impact perspective comprises 4 archetypes: “physical,” “diagnostic-physical,” “treatment,” and “triage.” Health care applications of the “physical” archetype provide material benefits to the patient (30/40, 75%). These are often achieved by means of preventive measures (28/40, 70%) or measures that help the patient to be more independent and self-determined in everyday life (11/40, 28%). Services of the “diagnostic-physical” archetype also focus on physical complaints (100%). In contrast to the previous archetype, however, they almost exclusively aim at disease diagnosis (131/132, 99%). Typically, services classified as belonging to the “treatment” archetype are used after the patient has been diagnosed (75/76, 99%). Another typical feature of these services is that they have a physical benefit in as much as 87% (66/76) of cases and a psychological benefit in as little as 13% (10/76). As for services of the “triage” archetype, these address patient referral to the appropriate health care specialist (100%). Given the comparatively small number of services in our sample (n=20), the underlying data may not be representative of real-world services. However, in three-quarters (15/20, 75%) of the cases, the occurrence of this archetype is closely linked to the diagnostic process. Other use cases are found mainly in the prevention (4/20, 20%) and treatment (9/20, 45%) of previously diagnosed diseases.

In summary, the agent archetypes differ mainly in how and by whom information is requested. It is worth noting that approximately 8 (80%) out of 10 services require manual process initiation, which means they need human interaction. This reliance on the interaction between humans and AI is also predicated on the fact that both archetypes are directed at trained health care professionals in more than half the cases. When it comes to the “data” perspective, we found the distinguishing characteristics to be the type of data and the data generator. Potentially, this can be explained by the fact that all 5 archetypes focus predominantly on patients as data targets. This is logical since information about a patient’s complaints, symptoms, or medical history concerns that patient directly. It is questionable, however, whether AI-based applications in the health care sector are still making insufficient use of some of the available data, for example, environmental information about a patient’s working or living conditions. As for the archetypes of the “AI” perspective, they distinguish between health care service offerings according to their dependency on specific hardware. This is consistent with the aforementioned high proportion of applications for image file interpretation, as these typically do not involve hardware components. Another reason for the high proportion of device-independent applications could be the complexity and cost of product development. This brings us to the final point—the archetypes in the “health impact” perspective. Their analysis focuses on patient benefit and shows that, to date, AI-based applications are primarily used for the diagnosis of physical illnesses. Given the increasing life expectancy of most industrialized nations, it is conceivable that the importance of AI in treatment and care, especially in geriatric care, might grow considerably in the near future. As for the other seismic shift in the technologically advanced West—our rapidly increasing awareness of mental health as a medical, social, and indeed political issue, it remains to be seen whether the number of services for the diagnosis and treatment of mental illnesses will also increase in the future.

To account for the phenomenon of AI-based health care services, we developed a new taxonomy based on the latest science [24], a structured literature review, and an extensive sample of real-world instantiations. As our rigorous demonstration and evaluation confirmed, this due diligence in the development of our taxonomy ensured its broad applicability and central positioning in the intended research field. As a result, this study contributes to the knowledge of how to incorporate AI into health care services successfully. Furthermore, it provides a clear and comprehensive structure to a fast-developing research field. Although there may be questions about the usefulness of taxonomy in such a dynamic field, this study indicates that a phenomenon does not have to be static for it to benefit from a taxonomy. As we have seen in similarly fast-developing research areas, such as the Internet of Things [26], FinTechs [27], analytics-based services [44], or Blockchain [74], there have been considerable benefits as a result of taxonomy development.

Meanwhile, our research results have theoretical as well as managerial implications. From a theoretical point of view, our research is the first to investigate the structure of AI-based health care services based on perspectives, dimensions, respective characteristics, and overall archetypes. The taxonomy supports academics and practitioners alike in that it enables them to recognize and re-evaluate the multiple forms of AI-based health care services, either to use those services to their full practical potential or to use the taxonomy for further theorizing. The archetypes we have identified in each perspective reveal various service offerings to have recurring patterns. These patterns advance the current understanding of how digital technologies transform the health care sector with the use of AI. Most importantly, this study incentivizes more extensive research on application areas in which AI has not yet been used to its full potential. Doing so encourages scholars to explore the multiple opportunities to apply AI in the ever more significant sector of digital health care.

From a practical point of view, our taxonomy and archetypes provide multiple use cases for practitioners. First, it is worth noting that practitioners who have already integrated AI into their health care service may find our taxonomy helpful in classifying this service and comparing it to any other offerings by their competitors. Second, practitioners who wish to introduce new AI-based health care services can use our taxonomy to gain an overview of possible applications, be it to find suitable options that are already out there or to find inspiration for potential new applications in their working fields. By analyzing numerous real-world examples, we were able to shed light on the intricacies of AI-based health care services and the general distribution of their design characteristics. By using our archetypes along with our insights into their frequencies, practitioners can identify which potential AI adoption fields are already widely used and might, therefore, be easier and more accessible areas in which to implement their services. While this focus provides limited potential for future innovations, our taxonomy facilitates the assessment of alternative possibilities in less common adoption fields, where it can be used for the dual benefits of orientation and structure. Third, the insights derived from our taxonomy on the characteristics of AI-based health care services carry substantial implications for health policy and delivery. By categorizing AI services based on their unique features, policy makers can make informed decisions on resource allocation, regulation, and integration into existing health care systems. This taxonomy facilitates the identification of targeted interventions for improving patient care, such as personalized treatment plans and enhanced diagnostic accuracy. Fourth, understanding the characteristics of AI-based services aids in the development of policies that address ethical considerations, privacy concerns, and the need for transparency in health care AI implementation. Fifth, in the context of population health, the taxonomy enables the strategic deployment of AI technologies to address specific health challenges, reduce disparities, and optimize overall health outcomes. This discussion underscores the practical relevance of our taxonomy, extending beyond theoretical considerations to directly influence health care policy making and health care delivery.

Of course, this study is also subject to certain limitations that require careful consideration to ensure the correct interpretation of our results. The first of those limitations to address is that our classification, as to whether or not a service is AI-driven, is based on the external appearance given by the service provider’s website or information supplied by other media. It cannot, therefore, be guaranteed that all of the services considered in this study meet the academic criteria to be classified as AI-driven. There is a hidden benefit in this putative limitation, however. Our research approach does not focus on how AI is seen in research but rather on how AI is perceived and applied in practice. The second limitation to discuss is that AI is a rapidly evolving field, and as new developments overtake the research presented in these pages, our taxonomy may need to be adapted and extended accordingly. For example, the sample of service offerings that we analyzed has a noticeably high frequency of diagnostics-related services due to our definition of the term service and AI algorithms that were primarily used for making decisions and classifying symptoms. However, current developments in the topics of generative AI or robotics might shift this trend toward treatment and care or patient enablement. With that in mind, we were especially thorough in our structured analysis of current AI adoption and emphasized the extensibility of our taxonomy to provide the necessary guidance for future adaptations and extensions in research and practice. The third limitation worth considering is that we based our research on empirical insights from top lists. While our data collection was rigorous, our selection of the publicly available AI-based health care services only represents a subset of the current services. As for the fourth limitation of this study, while we are confident that our taxonomy offers a good representation of the current AI landscape in health care, it will undoubtedly need to be changed and adapted due to rapid developments at a technical level [75]. This could particularly affect the dimensions of AI capability and the combination with the application area. However, the most significant change will probably be seen in the archetypes. A comparison between the current and future status would undoubtedly be interesting for future research.

Being a relatively new paradigm in health care, AI has been heralded as a panacea, yet only the future will tell if it can deliver on those lofty expectations. In the meantime, we have developed a taxonomy that serves as a structuring tool for researchers in the field of AI and as an innovation map for practitioners. It provides a simple yet potent method of classifying current and future AI-based health care services. Such classification allows the user to better situate and evaluate these services by attributing their distinct characteristics to 10 potential dimensions, such as service recipient, mode of interaction, data generator, data target, data type, hardware agnostic, AI portfolio integration, AI capability, application area, and health benefit. Having tested our taxonomy with representatives of its intended future users in order to classify 268 AI-based health care services and identify all relevant archetypes, we were able to ascertain its applicability and assess the potential for AI integration in current health care services.

By way of conclusion, then, it remains only to be said that this study was conducted in the intellectual tradition of Nickerson et al [24] and Kundisch et al [25]. As such, it contributes to the expert literature by providing a common understanding of AI-based health care services and offering a taxonomy that promises to assist in the development of AI-based health care services and the assessment of their potential in this competitive field. It bears repeating, however, that the implementation of AI in health care is a young and rapidly developing research field, so although we are convinced that this taxonomy can serve as a useful analytical tool in the present, its future usefulness will depend on its adaptation to relevant advances in the field. We hope that our work here will inspire the future research required to identify those advances and make the necessary adaptations to our taxonomy.