This scoping review was conducted using the recommended PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) guidelines published in 2018 [19].

This scoping review was conducted between February and April 2022. The databases examined included PubMed, Web of Science (WoS), Scopus, and Google Scholar. All papers available on these databases were exported to Mendeley software, which eliminated duplicate papers. Upon completion of this first step, the screening of evidence was initiated.

The search strategy was conducted following the Population/Problem, Intervention, Comparison, and Outcome (PICO) criteria [20], where the population was older adults, the intervention was AI, the comparison was compared or not with other diagnostic methods, and the observation was FS (reporting at least one of the following values: sensitivity, specificity, accuracy, AUC).

The search was conducted using the following keywords and Medical Subject Heading descriptors on PubMed, along with the following search elements on WoS, Scopus, and Google Scholar: “artificial intelligence,” “deep learning,” “machine learning,” “natural language processing,” “neural network,” “unsupervised learning,” “supervised learning,” “frail elderly,” “frailty,” “frail syndrome,” “diagnos*,” “recog,*” “prognosis,” “detect*,” “screening.” Moreover, the Boolean descriptors “AND” and “OR” were also included. The full search strategy is shown in Multimedia Appendix 1.

The papers included in this review contained information on the identification and diagnosis of FS in older adults through any type of AI. Textbox 1 shows the specific criteria used for paper inclusion or exclusion.

Four steps were implemented during the study selection process. First, entries were identified on the databases and independently reviewed after searching for and identifying duplicate works. Second, headings and abstracts of all registers were examined to identify those to be eventually included. Third, the full text of all candidate papers was read for determination of their final inclusion or exclusion in this study. Fourth, after selecting the final entries, an in-depth review of the studies in full was conducted. All four steps were independently reviewed by two researchers (AGM and DVD), while any inconsistency detected was discussed and settled by mutual agreement or engaging a third researcher (VPC).

According to the PRISMA-ScR guidelines, once the papers had been selected upon assessing that the inclusion criteria had been met, the abstraction of data of interest began, including: (1) general characteristics of the studies, (2) specific classification of different AI models used in each study, (3) characteristics of the data used in each study, and (4) type of data set in each study. The type of data used in each study was specifically indicated, including (5) data related to sensitivity, specificity, accuracy, and AUC; and (6) performance of each model provided by the different papers.

After conducting the search using the keywords and Boolean descriptors mentioned above, a total of 926 papers were identified from different databases. Once duplicate papers were eliminated, a first screening was implemented by analyzing the title and abstract. All papers not meeting the inclusion criteria were excluded; the main cause was the lack of identification or diagnosis of FS in older adults or not using AI. The full text was identified for 37 papers meeting the inclusion criteria.

Of these 37 papers, 11 did not report at least one of sensitivity, specificity, accuracy, or AUC, and were consequently excluded from this review. Finally, 26 articles [21-46] were included in the present review (Figure 1).

Among the most relevant countries of publication were the United States, where 30.7% of the studies were published, followed by Korea and Taiwan, with 11.5% each. The years in which more publications were released were 2021 and 2020, accounting for 38.5% and 26.9% of all articles, respectively. With the exception of two articles, the remaining articles identified FS in older adults from a general perspective, without specifying the type of frailty. Machine learning was the type of AI most frequently used for studies amounting for 69.2% of all studies (see Table 1).

Table 2 shows the detailed classification of the AI algorithms used in the included studies. The information contained was structured by classification models, linear regressions, and deep learning–based models.

Table 3 shows the characteristics of the different data used in the included studies. The information contained was structured by the size of the data set used, type of data (clinical or nonclinical), abstraction source (public/private), and type of subject within the study sample.

Table 4 shows the characteristics of the different types of data used in the studies included in this review. The information contained was structured by the set of clinical and nonclinical data of subjects under study, with a set of subtypes explored and identified.

Table 5 shows the ratio of different data explored and abstracted in the studies related to accuracy, sensitivity, specificity, and AUC.

Table 6 summarizes the AI models and type of data used in each study, along with specific data for sensitivity, specificity, accuracy, and AUC.

This scoping review primarily aimed to analyze existing scientific evidence on the use of AI in the identification and diagnosis of FS in older adults, as well as to identify which model provides enhanced accuracy, sensitivity, specificity, and AUC. The outcomes achieved showed that AI might be an acceptable, accurate, and reliable tool for the identification and diagnosis of FS in older adults. However, existing evidence showed highly heterogeneous results as far as the application methods are concerned as well as in relation to the data that need to be included in the programming being implemented.

A considerable number of different AI algorithms were identified. Therefore, it would be interesting to find out which types of algorithms have been used and determine which AI features provide enhanced accuracy, sensitivity, specificity, and AUC for identifying and diagnosing FS.

Of the 26 papers included in this review, the most frequently used models were usually the classification models, with random forest models accounting for 34.6% and support vector machine accounting for 30.7% of all models, followed by the nearest-neighbor classification algorithm (23%), naive Bayes algorithm (23%), and baseline logistic regression (23%). Following these, artificial models such as neural network and deep neural network algorithms were also identified (15.4%). Finally, other models were used, but featuring a lower ratio.

Another relevant variable of interest was related to the details on the data used in different models. In this sense, these details may be classified into the size of the data set used as well as the used source, mainly nonclinical data and clinical data of patients.

The size of the data set was not reported in numerous papers (42.3%). Among the papers reporting this information, the majority (23%) used a data set size above 2000. Following these, 11.5% used between 200 and 600 data, 11.5% had a data set size between 100 and 200, 7.6% had a data set of less than 100, 3.8% were between 700 and 1000, and 3.8% were between 1000 and 2000.

Most of the studies used nonclinical data (65.4%), which were obtained through different sensors monitoring an activity. The most frequently used sensors were portable sensors and inertial sensors (19.2%), followed by a kinetic sensor (15.4%), radar sensor (3.84%), ultrasound sensor (3.84%), and grip strength sensor (3.84%), represented by one study each.

To a lesser extent, clinical data were also used (34.6%). Most of these studies are based on data abstracted from electronic medical histories (23.1%). Data were also abstracted from questionnaires and heart rate dynamics, although at a lower frequency.

The retrieved and examined studies contained at least one of the four main measurements for the performance of a clinical diagnosis of FS based on AI models, including accuracy, sensitivity, specificity, and AUC. The analysis was focused on the indicators identified for diagnostic validity and was based on the type of data, along with the computation model used. The most frequently reported data attaining the highest values in the diagnostic validity of FS were those focused on assessing physical functions by parameterization of kinetics and kinematics.

Using inertial sensors for the monitoring of kinematic variables in clinical contexts was reported in five of the studies [30,31,42,43,45]. Although inertial sensors for kinematic analysis in clinical contexts were among the most frequently used tools for FS identification, most of these studies [30,42,43,45] only reported the data for accuracy, ranging from the lowest accuracy value of 73.3% from Panhwarr et al [30] with an artificial neural network to the highest value of 96.2% from the study by Jung et al [42] based on the deep neural network computational technique. Only Rahemi et al [31] reported the AUC value (88%).

Moreover, the use of inertial sensors for monitoring activities in a nonclinical context was also frequently used among the included studies [23,25,29,39,40]. In four [23,25,29,39] of the five studies using a kinematic data profile, at least three of the four performance metrics of interest were reported. In this sense, irrespective of the programming used, we found that the lowest accuracy value, reported by Park et al [39], was 73.2% and the highest value, reported by Abbas et al [40], was 99.72%. Concerning the sensitivity value, four of the studies reported these data [23,25,29,39], ranging from the highest value of 97.64% in Garcia-Moreno et al [25] to the lowest value of 71.8% in Park et al [39]. Four of the studies in this subgroup reported specificity [23,25,29,39], which ranged from 74.2%, reported by Park et al [39], to 99.47%, reported by Abbaset et al [40]. Finally, within this group of kinematic data, the AUC values ranged from 79.5% in Park et al [39] to 87% in Minici et al [23].

The use of kinetic sensors for activity monitoring was reported in four studies [22,35,36,39]. However, only one of these studies [39] reported all four metrics of diagnostic capacity, attaining values that were lower than those obtained with models based on other data types. Two studies [39,43] reported AUC data equal to or higher than 91.9%, although they did not report information on sensitivity or specificity. Due to the lack of reported information, we cannot confirm data based on kinetics as the most suitable for FS diagnosis. Future research reporting data on sensitivity along with specificity for this type of data profile is needed.

When focusing on the different AI methodologies applied and their diagnostic capacity, heterogeneity was also found in the results. Garcia-Moreno et al [25] showed the best ratios for the diagnostic capacity of FS in older adults by implementing a nearest-neighbor classification algorithm. However, this type of AI was only used in one other study [35] and the findings reported based on kinetic data were not aligned; thus, the identification and diagnosis capacity of this type of AI was lower. In addition, within the AI computation group showing the best results in relation to classification and diagnosis, we found one study [32] that implemented the C5.0 algorithm AI model, which was used for the analysis of patients’ clinical information sourced from their medical records and self-reported details provided by questionnaires. Other AI models that are frequently implemented [26,30,35,41] were support vector machine–based models, showing homogeneous findings with high diagnostic accuracy ratios. Another frequently used model was a neural network [22,23,31,34,44,45], in which heterogeneity was apparent in relation to the statistics used. The diagnostic accuracy, which was the most frequently reported metric, reached above 73.3%. Finally, logistic regression was reported as another AI method for this purpose, which was used in four studies [28,32,42,43], but showed the lowest overall ratios for diagnostic accuracy among all of the AI models reported in the included studies.

Conventional FS assessment has traditionally been based on two principal sources: (1) self-reported questionnaires such as the Groningen Frailty Indicator, Tilburg Frailty Indicator, Sherbrooke Postal Questionnaire, Vulnerable Elders Survey, and Strawbridge Frailty Questionnaire and (2) clinical assessments measuring frailty, such as the Clinical Frailty Scale, Clinical Global Impression of Change in Physical Frailty, and Short Physical Performance Battery. The latter assessments have been the most frequently used in clinical contexts. More recently, specific FS assessment tools have been developed, such as the Edmonton Frailty Scale, Frailty Index derived from the Comprehensive Geriatric Assessment, and Triage Risk Screening Tool, based on models including different domains [47]. This conventional stream for FS diagnosis explains the fact that numerous models identified in this review used data associated with physical function through movement and other sensors. The combined models are booming tools that imply a longer assessment but a better diagnostic success rate.

These conventional tools showed an average accuracy cut-off point at 83% with respect to FS identification. However, the sensitivity ranged from 56% to 89.5% and the specificity varied between 52% and 91.3% [47]. The results of these conventional models were lower than those achieved by AI for most of the records, with many of the AI computational programming tools obtaining performance values higher than 95% for the accuracy, sensitivity, and specificity metrics. Traditionally, FS diagnosis has mostly been focused on biological variables. In this sense, the identification and classification of geriatric syndromes are extremely likely to fail owing to high variability in the underlying biological factors. Given the multifactorial nature of FS in particular and geriatric syndromes in general, the use of AI and data science is being considered in the future when identifying and diagnosing complex health conditions.

Integrating AI into the identification and diagnosis of FS requires the development and validation of precise algorithms that incorporate multiple data sources such as wearable devices and sensors. To ensure the highest level of integrity, it is crucial to address data quality and privacy concerns. AI should be used as a complementary tool to support clinical judgment rather than as a replacement. After validation, AI systems should be user-friendly and integrated into clinical practice. The performance of AI algorithms should be regularly monitored to ensure their continued relevance and accuracy in clinical practice.

To the best of our knowledge, this is the first scoping review analyzing diverse tools for FS identification and diagnosis using AI, as well as the accuracy, sensitivity, specificity, and AUC values from different models.

A set of limitations were identified. First, there is a scarcity of papers reporting all statistical data on accuracy, sensitivity, specificity, and AUC of AI tools. Second, the high heterogeneity shown by the various AI models did not allow us to standardize findings on which type of AI would be the most suitable for having our goal in mind. Another inherent limitation of these AI systems is that their programming, together with the collection of the data to be used, requires a huge amount of expertise and time compared to many of the conventional systems. Finally, a meta-analysis would be interesting to obtain more quantitative information from this scoping review.

Future research will be necessary to report all the parameters needed to conveniently establish the validity, accuracy, and reliability of these AI computational systems. Further research is needed, using sufficient data, as well as an analysis based on data cross-checking from different variables, and a meta-analysis could be interesting to offer more qualitative information about the main topic of the present scoping review. A point of interest for future research would be the use of combined diverse data sources rather than only a single type of data, as demonstrated by the diverse studies included in this review. The operational nature of AI implies that the higher the volume of data used and the higher the variety of variables included, the more likely it will be to determine a convenient functioning and learning development when diagnosing multifactorial constructs such as FS. It should be mentioned that all studies included are framed in the development of basic research. It would be interesting for future studies to implement these AI systems through translational research in health systems and different environments.

This scoping review showed that AI may be an acceptable, accurate, and reliable tool for the identification and diagnosis of FS in older adults. AI could be a useful tool for identifying and diagnosing FS in older adults in both clinical and nonclinical contexts. However, reported evidence shows highly heterogeneous results as far as the application methods are concerned as well as in relation to the data that need to be included in the programming being implemented.

Future research is needed to assess the validity, accuracy, and reliability of AI computational systems using adequate data sources and diverse variables. Combining diverse data sources would be beneficial, and implementation of these systems in health systems and different environments through translational research would be interesting.

The data collected proved that the most well-performing analysis lies in the combined use of electronic medical histories along with kinematic information from inertial sensors monitoring activities in a nonclinical context (activities of daily living). Currently, AI-based computational systems are valid, accurate, and reliable tools that when implemented in health care systems can help to reduce direct and indirect social and health costs associated with dependency and disability.