The electronic databases used in this review included the PubMed and CINAHL databases. Table S1 in Multimedia Appendix 1 provides additional information about the search terms used for this review.

All authors participated in the planning of screening and data extraction, as well as in the construction of search strategies. One author (EH) evaluated the titles, abstracts, and full texts of all publications identified through systematic searches.

A data charting form was developed and refined, incorporating the final variables, such as study sample, country, the dependent variables used in the study, the independent variables used in the study, data source, statistical analysis, findings or conclusions, settings, and implications. Key study characteristics and detailed information were extracted from eligible studies and recorded, using the charting form. Furthermore, we categorized and synthesized the sources of evidence that contained explicit information on leveraging electronic data and addressing the practical implementation of the outcome of NHA risk prediction. The study selection and data charting processes were amended, with any disagreements among the authors being addressed and resolved consensually.

Based on the results obtained from data extraction and charting, we found that the identified studies covered varying health care settings and timings of nursing home utilization. Accordingly, we established distinct phases to differentiate the timings and settings of risk prediction (Figure S1).

The initial search identified 610 articles across all search engines. Searching through the references of past reviews provided an additional 46 articles. A total of 313 articles remained after removing the duplicates. Screening the titles and abstracts of the 313 articles resulted in the exclusion of 263 articles. The full texts of the remaining 50 publications were reviewed, which resulted in the exclusion of 16 articles. A total of 34 peer-reviewed original research publications were selected for final inclusion in this review (Figure 1). Tables S1-S4 in Multimedia Appendix 1 [8,18-50] present the search strings and the comprehensive characteristics of the sources of evidence included in this study.

In this study, we reviewed 34 studies that either assessed the association between predictors and outcomes or reported the development and validation of predictors of NHA via observational studies based on EHR or administrative data. These predictors were further categorized into sociodemographic, caregiver support, health status, health use, and social service use factors, consistent with previous studies [9,18-22,24,26,34,39,40,43,46,52,53]. Most studies (23/34, 68%) included multiple outcome variables in addition to NHA. All-cause mortality, hospitalization, and rehospitalization were frequently used as outcome measures. Cox proportional hazards models, logistic regression models, and a combination of both were frequently used to predict NHA. In terms of data sources, administrative claims data and EHRs were jointly used, rather than each source being used alone.

Overall, studies did not fully leverage the data sources of our interest. Administrative claims and EHRs may have strengths, including the abilities to help with recruiting larger sample sizes, incorporate expanded types of predictors or outcome variables, enable short-term prediction, and facilitate easy implementation in practice [54]. For instance, routinely coded EHR data can be used to establish a frailty index, thereby allowing for the classification of subtypes of frailty. These data can be readily integrated into clinical computer systems, automatically populated, and made available without requiring additional resource input [44]. Further, although we found some studies that used large sample sizes comprised of more than 200,000 participants [18,20,28,31,43,44,47], this did not necessarily mean that the predictors or outcome variables were distinctive. In addition, although many studies used administrative claims data, the use of billing information or the construction of cost predictors was rarely observed. Discussion on enhancing the quality of data source utilization was limited. For instance, only 1 study reported the variation in outcomes based on the use of structured EHRs versus nonstructured EHRs [32].

Developing a validated risk prediction model that is expressed as an index or score is one option to effectively leverage electronic data sources. Such a model would serve as a convenient tool for preplacement assessments and help address the challenges posed by big data sources, which often contain excessive information that may be disorganized or lack systematic summarization. As international guidelines recommend the routine identification of frailty to provide evidence-based treatment [44], several studies tried to identify frailty from routinely available electronic health data, which showed fair predictive value. Despite the identified NHA predictors, the development of a predictable index from routine data sources needs further exploration, as the predictive value of the identified factors was limited.

Although many of the reviewed studies included clinical predictors, social determinants of health, such as education, economic hardships, food security, and community factors [55], were used in a limited fashion. The limited use of social determinants of health might be partly due to the complexities of extracting such data from EHRs and the limited capture of these data in insurance claims [56]. Similarly, caregiving, which is an important aspect of social support for patients, was used as the predictor of NHA in a few of the reviewed studies. One study [21] collected information on a caregiver’s formality and qualitative characteristics of caregiving (eg, value of caregiving to and impact of caregiving on care recipients). In another study [40], a comprehensive assessment survey, which included the caregiver distress factor, was conducted and included as a predictor of NHA.

The study findings have implications for the development of prevention strategies in communities. Several studies reported the utilization of home care to delay NHA [8,36-39,41,48,49]. Prediction models for older adults may play a distinctive role in screening the at-risk population or assessing the need for home services in primary care settings. Efficient NHA prediction may involve making the short-term decision of whether a person is suited for living in the community or for institutionalization and applying this approach to a specific population instead of applying it broadly to all community-dwelling older adults. In addition, such complex decision-making may require informed communication among older adults, caregivers, and care providers about resource allocation between community-based care and institutional care. Therefore, NHA prediction in communities should be in alignment with the continuum of care, considering the benefits and costs of institutionalization.

Unlike other recent studies that primarily targeted risk prediction for nursing home residents [57,58], this study examined the preplacement status of individuals, with many of the study samples residing in communities. Our focus was on exploring how the study results could assist in developing prevention strategies at the population health level within these communities.

The main findings of this study are consistent with a previously conducted meta-analysis [9] and cohort study [53], which identified predictors of NHA, specifically ADL dependency, cognitive decline, and prior NHA. Studies on the predictors of readmission to nursing homes also showed similar trends [59]. Nevertheless, in comparison to these prior studies, our study revealed an expansion in the types of predictors reported since 2010, including frailty and geriatric syndromes.

Previous studies on risk prediction via EHR data or administrative claims data discussed various strengths and challenges [48,54,60,61]. EHRs may provide an alternative to the longitudinal cohort studies that are traditionally used to construct risk models [60]. In this regard, clinically available risk prediction models for the early prevention of NHA that are based on EHRs are more appropriate for a clinical context than those based on other types of electronic data sources [44]. However, variations in EHR implementations and data standards across health care systems may limit the practical adoption of EHR-based NHA prediction models [62]. Future research should pay attention to the development and use of predictors of NHA across different electronic data, health systems, and patient populations.

To our knowledge, this is the first study to systematically review existing evidence on identifying predictors of NHA based on administrative claims or EHRs. Since NHA is a considerable burden to health care delivery systems, identifying and developing NHA prediction models can enable health systems to both reduce overall health service utilization and improve health care outcomes among older adults. We reported distinctive characteristics of existing predictors and prediction models, including how they leveraged specific data sources. These findings have implications for delaying NHA in communities.

This review has several limitations. First, this review is limited to the literature on EHRs and administrative data concerning the older adult population and NHAs. Therefore, the findings from the literature on EHRs and administrative claims data may not be generalizable to other data sources. Second, this review only included studies that researched potential factors or predictors of NHA in a retrospective manner. Thus, studies that evaluated a specific program for reducing NHAs were excluded in this study. These studies may have included predictors that could have been added to the list of predictors found in this review, although such predictors are often not available in routine EHR data or claims data sources. To mitigate this limitation, we searched and reviewed the list of excluded studies during the multiple stages of the final selection process. Third, we limited this review to studies that used administrative data or EHR data as their primary sources of data to assess predictors of NHA. However, the differences between administrative data, administrative claims data, and EHR data are not always clear in different clinical settings. Future studies may refine data sources beyond EHRs and claims data, although the application of such data sources in population health management efforts should be further evaluated. Finally, our review followed a scoping review guideline and thus may have inadvertently excluded a few relevant studies. A systematic review of relevant studies is recommended to provide a comprehensive list of predictors for NHA among older adults.

Considering that NHA is not an end point but rather a transitional point [9] in the care continuum, an NHA prediction model can be used in a risk prediction and prevention framework (Figure S1 in Multimedia Appendix 1). In the first phase of preventing NHA, adults may undergo routine screenings while maintaining independence in their communities. If they experience functional decline or reach a certain level of dependency, preventive interventions can be implemented based on the results of prediction models to delay NHA timing in the second phase of prevention. Even after NHA occurs, prevention strategies to reduce the risk of other adverse outcomes during nursing home residency are still important. During this phase of prevention, we believe that NHA risk prediction models can evaluate whether an individual is prepared to return to the community, if relevant. Thus, risk scores calculated by using a frailty index or a comprehensive geriatric assessment index can also be used for newly admitted nursing home residents. In future studies, other factors may be controlled to examine the impact of using NHA prediction scores (Figure S2 in Multimedia Appendix 1).

Of note, some studies focused on continuous care and primary care qualities [8,44,55]. As prevention programs for delaying NHA at earlier stages can be implemented in primary care settings, further studies are warranted on this topic.

Several studies have investigated the potential predictors of NHA by using EHRs or administrative data. Frequently used predictors included sociodemographic factors, caregiver factors, health-related factors, medications, and medical and social service utilization. The most frequently used models for predicting NHAs were Cox proportional hazards models and logistic regression models. Prediction tools based on EHR data or administrative claims data may assist clinicians, patients, and policy makers in making informed decisions and allocating public health resources.