Despite a surge in funded research in the application of digital technologies toward a higher assurance of quality nursing care, in times of aging societies and skill shortages [1], the application of artificial intelligence (AI) in nursing practice is still scarce. In this context, AI can be defined as algorithms that enable learning from data sets to achieve intelligent, goal-oriented action.

Recent systematic and scoping reviews on the application of AI in nursing research (as well as in practice and emerging trends), covering original research published until October 2019, identified papers listed in medical and multidisciplinary databases. These included studies focused on machine learning (ML) methods, such as deep learning [2], or on health technologies that incorporate AI approaches themselves, such as robots or clinical decision support systems [3]. Various application scenarios have been identified, including clinical or organizational outcomes (eg, falls), admission decisions in emergency medicine, high-definition image recognition, as well as socially assistive robots or health care assistant chatbots [2,3]. In addition, recent years have seen an increase in research highlighting possibilities for the future development of AI in nursing care while underscoring the importance of collaborative, interdisciplinary research, and representative, robust data sets [2].

However, as of today, a universally accepted classification of AI subfields relevant to health, which could act as a vantage point for AI in nursing practice, is missing [4]. Prominent AI approaches include ML, expert, and hybrid systems. ML, as a method of data analysis guided by algorithms, identifies patterns in data and learns from them using different approaches [4]. This is utilized in medical diagnostics, for example [5]. Expert systems build on a knowledge base and a rule-based reasoning engine [4], which, in combination, mimic the reasoning of a human expert who would solve a complex problem by applying predefined if-then rules drawing on a specific knowledge base [6]. These systems can be found in tools that support clinical decision-making and case-based reasoning [7,8]. Hybrid systems combine different AI capabilities by integrating ML with expert systems [9-11]. AI applications aimed at determining the meaning of texts, such as clinical notes, can be found in the AI subfield of natural language processing (NLP) [4,12]. AI applications for automated planning and scheduling can be used to improve the efficiency of human procedures [4], such as generating nursing staff rosters or care-related scheduling decisions [13,14]. Applications that target image and signal processing use algorithms that typically include signal feature analysis and data classification to analyze images or data produced by movement or sound [4]. These can, for example, aim at activity and health monitoring, wound detection, or pressure injury and fall prediction or prevention [15-19].

Turning our attention specifically to nursing care settings, the primary opportunities for applying AI include application scenarios such as decision support in complex care situations [3,18,20,21]. AI also holds great promise for supporting nurses in tasks considered to take place remotely from direct patient interactions [3,21]. High expenditures of nurses’ working hours are frequently reported as being used for the documentation of care processes, with some care facilities reporting up to almost a third of daily working hours being expended for documentation processes [3]. This represents one of the many starting points from which to develop AI solutions to consistently improve nursing care processes and support nurses efficiently in their daily tasks. AI applications for the direct support of care-dependent persons and their informal caregivers are another starting point, as studies with AI approaches in different community and home care settings have shown [4-6]. This is of particular need, given that most long-term care recipients in Germany are being cared for in their own homes [22]. Until now, little knowledge on the practical relevance and applicability of AI systems with setting-specific requirements in nursing care, for example, when introduced in care processes involving persons with limited cognitive abilities, exists thus far.

Furthermore, the transformative effect of AI, resulting from its ability to change the intrinsic nature of health care delivery, is accompanied by ethical risks, namely, concerning the validity of evidence, the fairness of outcomes, and the traceability of harm caused by algorithmic activity [23]. Furthermore, although consensus on the potential of health technologies powered by AI to enhance nursing practice has been reported [3], the critical ideological and ethical nature of nursing practice still needs to be considered, and the role of decision-making, enhanced and burdened by an amplified understanding of opportunities granted by AI applications remains uncertain in the context of providing ethical and transparent nursing care [21]. To our knowledge, an extensive overview of the evidence base and status quo of research on AI for application in nursing practice, including evidence from medical and computer science databases, is missing. By identifying promising application scenarios for AI in nursing practice, such an overview contributes to the systematic enhancement of research and development for AI in nursing practice.

This rapid review aims to synthesize the evidence base of application scenarios for AI in nursing care settings, namely, ambulatory and stationary (long-term) care, acute hospital care, and nursing education. We also address prominent adjacent aspects within the ethical, legal, and social discourse concerning AI in nursing care by addressing the following review questions:

Although our approach is broader than those of similar reviews that have focused exclusively on ML algorithms [24], it is also more broadly scoped in that it considers the ethical and regulatory context of the AI system deployed but does not focus exclusively on these aspects as other reviews have done [25].

We conducted a rapid review to identify and synthesize publications promptly [26]. A protocol describing the rationale and methods of this review was published in May 2020 [27]. This paper follows the guidelines outlined in the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) Statement [28].

We included all designs of quantitative and qualitative original research, systematic reviews, and discussion papers or essays on ethical, legal, and social aspects that address the application of AI, specifically in:

We included publications in the English language from 2005 onward, as we expect publications on AI to become quickly outdated and updated. Publications focusing on improving the functionality of medical diagnostic or therapeutic technologies, without clearly describing nurses as a relevant target group being affected by the application of the AI system or being directly involved in the application process, were excluded.

Publications had to either designate at least 1 of the following groups of persons as main users or as benefactors of an AI application:

Publications using inconclusive terms, such as the elderly or health care professionals, without further information on target groups of users or benefactors, were also assessed for inclusion.

Care settings encompassed ambulatory and stationary long-term as well as acute outpatient and hospital care (including rehabilitation facilities) and nursing education settings. Studies assessing care in community settings or assessing the populations mentioned above, but in a laboratory setting, were also included.

As there is no conclusive definition of specific AI abilities or subfields that are relevant for health [4] or nursing care as of yet, all types of AI systems or approaches, ranging from clearly stated types (ML, expert system, hybrid system), to any type of approach combining ML and an expert system and algorithms, across to rather vague descriptions, such as smart system or AI in health care, were deemed as eligible.

We searched the following databases in June 2020: PubMed, CINAHL (including Embase), Association for Computing Machinery Digital Library, Institute of Electrical and Electronics Engineers Xplore, Digital Bibliography & Library Project, computer science bibliography, and Association for Information Systems Library. In addition, we searched digital libraries of leading conferences identified through expert consensus within the study team. These conferences were specifically the Association for the Advancement of AI conference, the Association for Computational Linguistics Conference, the Conference on Computer Vision and Pattern Recognition, the International Conference on Machine Learning, the International Joint Conferences on AI Organization, the conference of the Association for Computing Machinery’s Special Interest Group on Knowledge Discovery and Data Mining, the Conference on Neural Information Processing Systems, the International Conference on Principles of Knowledge Representation and Reasoning, the Conference on Uncertainty in AI, the International Conference on Autonomous Agents and Multiagent Systems, and the European Conference on AI. The search strategy based on the block-building approach [29], combined terms for nursing and artificial intelligence and their respective synonyms. If applicable, we searched the titles, abstracts, and all fields of publication. In the first step, single terms for each block were searched. Second, all terms of a single block were combined using the Boolean operator OR. We initially deemed publications in German or English to be eligible. As no publications in the German language fulfilled the inclusion criteria, we focused only on the English. Finally, the results from the second step were combined for the 2 blocks using the Boolean operator AND. The hits were recorded for each step. To circumvent imprecisions between concepts described in the titles or abstracts, regarding the components of the search strategy, and to identify a large number of potentially eligible publications, we developed a preferably sensitive search strategy. Multimedia Appendix 1 contains the search strategy, search terms, and the number of hits for all databases and conference libraries.

Considering the type of AI approach, we found the vast majority (228/292, 78.1%) of studies employing ML approaches. Rule-based expert systems were used in 11.6% (34/292) of the publications, whereas hybrid systems were used in only 3 studies. For the remainder of the publications, the specific AI approach used was either not identifiable or this attribute was not applicable, mostly because of publications not employing specific AI approaches in basic or applied research. Studies incorporating real-world settings made use of ML approaches comparably in 71% (24/34) of cases, whereas expert systems were only covered in 2 studies and hybrid systems not at all (Table 3).

Most AI approaches have been described as solutions for image and signal processing (178/292, 60.9%), that is, the processing of large amounts of signals, such as audio and video data, for feature analysis and data classification [4]. AI approaches have been used for automated planning and scheduling. This category entails approaches used to organize and prioritize activities and that “can be used to improve the efficiency of human procedures” [4]. Studies have focused less often on the processing of human language (NLP). Image and signal processing were also performed in most studies conducted in a real-world setting with automated planning and scheduling, and research on NLP has been less frequently reported (Table 4).

The areas of support for nursing care targeted by the AI approaches are shown in Table 5. With regard to the intended effects from the described AI approaches, 47.6% (139/292) of the studies focused on the support of the direct, immediate process of care. The support of the organization of care services and the support of care-dependent people themselves, as well as risk estimation and prevention, are further prominent purposes. Risk estimation and prevention potentially pose a cross-sectional topic, where the type of support manifests at multiple levels. The health of the caregiver and education were addressed in only a few cases. For studies in real-world settings, risk estimation or prevention, support of direct care, and support of care organization were each a focus of 29% (10/34) of the studies, whereas the support of care-dependent people and education did not play a prominent role, and no research conducted in a real-world setting focused on the health of caregivers.

A more detailed summary of the purpose of the AI approaches is presented in Table 6. The most prominent purpose was activity and health tracking (monitoring or classification) in 30.1% (88/292) studies. Care coordination and communication are frequent topics, which, among others, include AI approaches classifying information in nursing documentation, supporting decision-making, and yielding information for coordination and continuity of care. Fall detection, fall prevention, and fall risk classification are also frequently mentioned purposes for topics in AI. In contrast to falls, other mobility-related aspects were of lesser interest and were mentioned in only a few studies. Further purposes with a high degree of specificity are the recognition, classification, reduction of alarms, and risk prediction and classification of pressure ulcers. Addressing nurse rostering or scheduling problems was the purpose of an AI solution in 4.1% (12/292) studies.

Requirements or challenges and barriers for the application of AI in nursing care outside of technological infrastructure or reliability, precision, and validation of data have been reported in 6 studies, including real-world scenarios [18,35,37,41,58,314] and 4 reviews [59-62]. Requirements included compliance with the EU General Data Protection Regulation and preferences of target users concerning usability and complexity, and also requirements stemming from implementation in specific care settings [35,60]. Furthermore, providers need to concern themselves with their capacity, ability, and willingness to generate data inputs required to achieve high accuracy, in contrast to the clinical burden of false positive or negative results [58]. In addition, the quality of administrative databases should not be affected by AI implementation in the care context [58]. Inclusion of caregivers, user engagement, and commitment to further the participation of older adults in the development and testing of AI systems as well as successful implementation are required [18,41]. Reported challenges and barriers target accuracy of recognition, integration with sensor networks, privacy, security, human–machine interaction, and cognition impairment of users, acceptance, and costs [59,61]. The physical appearance and programmed behavior of hardware hosting AI systems when presented in a humanoid form may seem confusing, unpredictable, or frightening and limit the interpretability of the system’s action for nursing home residents [41]. In addition, the appearance of sensors when they are nondisposable and bulky may pose a burden to caregivers [63]. Underreporting of relevant events and scarce public availability of databases holding sufficient data and information to compare one’s results, as well as limitations to data sets due to regional data protection laws, constitute barriers to the accuracy of algorithms and the external validity of results [37,62].

We identified 6 publications [7,21,64-67], specifically focusing on the ethical, legal, and social aspects of AI in nursing care. In addition, 7 publications [35,41,57,59,60,68,69], which were either reviews or studies including real-world settings, addressed selected ELSI aspects when discussing results or limitations of their work. Recurring aspects were consent (of care dependents or nurses), data privacy and safety, acceptance and implications for work processes and workforce, such as lack of human interaction and communication skills or the fear of replacement of nurses by technology and the implications arising from choosing humanoid designs for hardware hosting intelligent technologies [41,67]. Peirce et al [21] focused on relevant ethical, legal, and social aspects of nursing as a profession and highlighted implications arising from the type of data used, such as possible sampling bias, correlational false positives, and hidden discrimination, as well as the values and interests of companies building huge data sets that should be kept in mind. They pointed out the importance of nurses’ understanding of the underlying motivations and goals for creating algorithms as well as the learning mechanisms and potential to mediate, as AI-generated knowledge should not be regarded as universally valid and the potential of algorithms to limit nursing actions and cause the loss of human dignity should be regarded of utmost importance within the discourse on AI in nursing [21].

The results of this rapid review explicate the application scenarios for AI systems in nursing care. Hospitals, followed by independent living at home, were the most frequently investigated settings, whereas nursing homes and ambulatory long-term care were less often examined. The vast majority of studies applied ML, whereas expert and hybrid systems were used only in about every 10th publication. This implies that the current instance of AI is mainly ML driven. For instance, Taddy [33] described the evolution of ML toward status as a general purpose technology and, consequently, as the main driver of the current rise in AI. More than half of the publications focused on image and signal processing and one-third on automated planning and scheduling, whereas NLP appeared in less than 1 in 10 publications.

In the context of direct nursing care, AI is used to organize care processes and support care-dependent people or family caregivers through tracking, monitoring, or classifying activities and health data. This was followed by applications to support care coordination or communication, as well as nurse rostering and scheduling. Detecting, classifying, and preventing falls, as well as recognizing, classifying, and reducing alarms, and predicting and classifying pressure ulcers were further purposes of introducing AI to nursing care. Only a few publications went beyond proof-of-concept studies or laboratory experiments and applied AI in real-world scenarios. Few studies have assessed the effects of AI on clinical and organizational outcomes. In addition to technical or computational requirements, further requirements concerning the specific context of nursing care are scarce and mainly tackle overarching topics, such as data privacy, safety, and acceptance. The same holds for ELSIs, which, for instance, have not been reflected or discussed in most studies using real-world scenarios.

Most studies describe AI applications in hospital settings, particularly ICUs. This may be attributed to the availability of such data. Besides electronic medical, nursing, or health record data, real-time sensor data on vital parameters are more frequently available from ICUs than from regular wards, facilitating a multidimensional approach, which is, for instance, being used to classify risks or identify care needs. This finding is in line with a previously reported increase in the publication rate of studies using ML to analyze routinely collected ICU data [70]. The availability, quality, and quantity of data might also be limited due to differences in digitalization activities in specific care settings, as well as the sufficient inclusion of study participants and duration observation periods to generate large data sets from sensor data. Furthermore, the heterogeneity of different data sets complicates the use of data for AI development.

Considering the development of digital technologies in general, nurses themselves have reported that they feel that regular hospital wards or long-term care settings are being considered too little [315], which is consistent with our results. It should be noted that some of the included application scenarios cannot be attributed unambiguously to the nursing domain. Although an impact on nurses and patient care is evident when trying to reduce false alarm rates in monitoring [71,316] or to improve mechanical ventilation and sedative dosing processes [72], there is a blurred line between AI systems to support medical diagnostics and therapy, and AI systems to support nursing care. Considering the variety of applications for tracking, monitoring, and classifying health and activity, nurse rostering and scheduling, or detecting falls or fall risks, it is remarkable that so few studies went beyond testing the efficacy of AI approaches. This points to a gap in the existing evidence regarding the effectiveness of AI in real-world scenarios. This is also reflected in the results of the LOE ranking, with only a few studies using designs that could test effects on real-world nursing or on patient outcomes.

In addition, the explicit operationalization of nursing tasks or care processes, and desired clinical, psychosocial, or organizational outcomes, has not been addressed in most of the included studies. On the one hand, this might point to undiscovered possibilities for AI support in nursing care. On the other hand, it might limit the perceived benefit of AI in nursing care and, subsequently, the participation of providers, nurses, care dependents, and family caregivers in developing (as well as sustainably integrating) AI in care processes and everyday activities. Integration of AI in nursing care might also be limited by the lack of a sound description of outcomes, benefits, or values, which will influence the adoption or nonadoption of technologies in nursing care [315,317]. Our review points to a gap in published research on possible application scenarios for AI in nursing care on the one hand, and on the availability of evaluation results regarding already implemented AI systems on the other. This raises the presumption that such evaluations have so far been of less interest. This is particularly troublesome because some studies suggest little or no extra benefit of using AI when compared with alternative or existing solutions [45]. Concerning nurses’ need for technologies providing enhanced technological support of direct nursing care tasks to reduce physical burdens and mental stressors [315], there seems to be room for research on AI-sensitive outcomes in nursing care.

Concerning the requirements and barriers for AI in nursing care, we expected to find topics such as data quality and access, as well as factors associated with measuring primary data and obtaining and sharing routine data, more frequently reflected in the included publications. However, only a few studies have addressed these concerns. Most requirements and barriers mirrored topics that are not only relevant for AI systems, but also for digital technologies in nursing and health care in general, such as data privacy, safety, and user acceptance [318]. On the one hand, this could indicate that there are few nursing-specific requirements or barriers to consider. However, this seems unlikely given the heterogeneous origins of the included studies, which have been conducted in different societal and health systems, including, for instance, different regulations on data protection or storage. In contrast, the lack of data and access-related factors could be attributed to the fact that the descriptive or conceptual nature of most studies led to authors addressing requirements or barriers less frequently.

The ELSIs discussed in the included publications addressed prominent topics in the discourse on the use of technology in health care, such as data privacy and protection, consent, acceptance, and implications for communication [20,25]. These aspects were not addressed in most studies in the subsample of studies, including real-world settings. Only one publication focused specifically on the ethical implications arising from the knowledge generated by AI systems in the context of nursing care. Other ethical principles incorporated in existing AI guidelines, such as trust, sustainability, justice, and fairness [25], are covered superficially if at all. Even though the limited uptake of ELSI aspects in published research might be biased by the fact that the remaining publications were not screened for ELSI aspects, there seems to be room for researchers to incorporate the discussion of ELSI aspects in their work, contributing to building trustworthy and trusted AI solutions [319]. Furthermore, publications describing stakeholder processes, surveys, interviews, or focus groups involving care dependents or nurses and accessing their perspective on AI, were underrepresented in our sample. This indicates that there is room for implementing and facilitating the concept of participatory development and testing, which contribute to the demand orientation and acceptance of AI systems.

A major strength of this review is the sensitive design of the search strategy, which led to the inclusion of a large sample of study designs and publication types, giving an extensive overview of published works on applications of AI systems in nursing care, as well as considering publications from medical- and informatics-databases and conference archives, which to our knowledge is the first of its kind.

The decision to focus on published works limits the results, as our strategy did not include AI systems already in use for which scientific empirical evidence has not been published and which might be directly introduced to clinical practice by developers. Another notable limitation is our decision to use a rather broad definition of nursing care, care recipients, and care settings, which also included independent living of elderly people. As some of the included publications did not define nursing care as the primary application context, and often included nurses or nursing care facilities as possible users among others, publications dealing with borderline examples of AI application scenarios that might or might not be attributed to the domain of nursing depending on the originating context of the publication, such as medical diagnostics, were not included.

As we used the criterion of conducting field experiments or using real-world data to group studies into basic, experimental, or real-world scenarios, studies in the basic or experimental subsample tested applications that may be considered extending well beyond basic research topics when using other criteria to classify studies. We chose this classification primarily to show the extent of the AI solutions applied in real-world practice. It also needs to be noted that we chose, primarily, absolute and relative numbers of studies and categories to map the existing literature; however, this provides an overview of prominent research topics and points to gaps in the existing literature; however, this does not allow for an assessment of the sophistication of research on AI approaches done in the context of nursing care. Although an assessment of scientific impact, quality of content, originality, and clarity, as it had been done in the field of sensors, signals, and imaging informatics to identify research works that exemplify recent developments [320] was not part of this review, a number of papers have demonstrated useful techniques to improve the generalizability, interpretability, and reproducibility of increasingly sophisticated models. As we only included publications in English, language bias must be noted. The same holds for the decision to limit the publication range to 2005, and to the selection of the searched databases and conferences, which restrict the sensitivity of the search strategy. However, the increase in publications during the last 5 years indicates that our search managed to cover a relevant period of research and development in AI for nursing care.

To our knowledge, this is the first review of its kind to systematize a broad literature base on AI in nursing care, and previous relevant work on this topic is scarce. A direct comparison with the application of AI approaches in other related domains, such as medicine or global health, is difficult, owing to, among other things, reported heterogeneity in AI reporting and the lack of a standardized benchmark [321], which was also present in our sample of studies. Even though disciplines such as biomedicine seem to be more active than nursing in terms of publications targeting AI applications [322], overlapping areas and specific aspects to study such as ethical aspects as well as the impact of digital health interventions and the changes and requirements for the professional role can be noted, for example, in the medicine domain [323]. In relation to the nursing domain, Kikuchi [2] reviewed studies on AI technologies in nursing research that focused on clinical outcomes, such as fall prediction, surgery-related injury, nausea, depression, and survival of patients, as well as on managerial themes addressing bed allocation, decision support, communication risks, nurse burnout, nurses’ intention to quit, nursing diagnostics, and knowledge acquisition for nurses. Without including publications outside of medical databases, similar to most studies included in our review, the results indicate a focus on the performance capability of AI algorithms compared with standard statistical methods, underlining the assumption of a lack of evaluation studies on existing AI solutions. Buchanan et al [3] conducted a scoping review on emerging trends in AI-powered health technologies and their implications for domains of nursing, such as administration, clinical practice, policy, and research. In contrast to our current review, specific types of AI approaches were not reported. The described emerging trends are parallel to the application scenarios in this review and entail predictive analytics, clinical decision support systems, smart homes, and health care assistant chatbots, but lay a focus on robot applications [3], which were only sparsely included in this review, as most studies on robots included in the screening process did not meet the inclusion criteria for full text review. The potential of AI to enhance nursing practice, as well as the need for nurses to take on the shared responsibility to influence and take part in the way AI is integrated into the health system, are highlighted by the results [3], contributing to the importance of engaging in the discussion from the perspective of clinical practice and nursing science. Congruent with Shillan et al [70], who reviewed the use of ML approaches in the ICU context, a lack of methodological reporting guidelines for AI approaches using health care data or being conducted in a health care or nursing setting impedes the identification of relevant studies, as well as with the evaluation and rating of real-world relevance, confidence in reported findings, and translation into clinical practice.