This review considered any study using quantitative methods relating to the predictive use of data analytics in the fields of telecare and telehealth. Qualitative studies were excluded. The Population, Concept, and Context (PCC) framework was applied. Database searches were conducted in August 2022 and restricted to papers published within 10 years of the initial searches being conducted. Only papers published in the English language were considered.

Papers focusing on any and all users were included. All populations of users (anyone using a telecare and telehealth device or systems) including both adult and child services were valid for inclusion since the focus of this review was on the methods of analytics being applied, rather than the specific reason for accessing telecare or telehealth.

Any telecare or telehealth innovation that gathers or generates data and electronically communicates it for use in an analytical manner was valid for inclusion. This could be “passive” technology, such as sensors and wearables, or “active” technology where data are intentionally entered into a device by a user. Papers investigating devices, which do not directly monitor a health element of an individual, such as an educational app, were excluded. Any data analytics that make inference or predictions from the data they receive were included in this review. This includes diagnosis, classification, and anomaly detection and does not exclusively consider predictions of future events. Additionally, this review only considers telecare and telehealth devices related to a somatic condition, that is, physical condition of the body. Papers focused on mental health and loneliness, for example, were excluded because these conditions may require a significantly different management approach.

Any paper which had a “care in the community” setting was suitable for inclusion (patient’s own home, assisted living facilities, and sheltered accommodation). In-patient and non–home-based settings were excluded with the exception of papers that focus on technologies clearly designed for at-home use that have thus far only been tested on individuals in an in-patient setting.

All reviews (systematic, literature, and scoping) were excluded as this would cause duplicate data to be reviewed and could lead to bias through overreporting. Any paper outlining an entirely conceptual framework and not detailing on how it would work in practice was excluded. The review also excluded editorials, summaries, and opinion pieces.

Databases relevant to health and social care—MEDLINE [OVID], Embase [OVID], and Social Science Premium Collection [ProQuest]—were searched.

The following 2 key domains were identified for inclusion in the search strategy: data analytics and telecare or telehealth (see Table 1). Search terms that were deemed most applicable to each database were applied. MeSH (Medical Subject Headings) terms and free-text entries were considered as appropriate. Boolean operators such as “AND,” “OR,” and truncation codes were used to refine and improve searches. A copy of the full search strategy employed while searching the Medline database can be found in Multimedia Appendix 1.

Results from each database search were imported to EndNote [13] where duplicates were removed. Studies were uploaded to Covidence [14] for screening. Title and abstract screening were completed by 6 reviewers (ML, NW, ED, DK, MR, and LL). Every paper was screened independently by at least 2 researchers, with conflicts resolved through discussion. A third reviewer was consulted when agreement could not be reached.

Full-text versions of the accepted papers were obtained for full-text screening. There were 537 papers considered for full-text screening by the lead author. Of these 537, approximately 15% (n=80) were screened collaboratively by the lead author (EA) and 2 other reviewers (NW and DK). Interrater agreement (all 3 reviewers coming to the same conclusion on inclusion or exclusion) was categorized through the following thresholds: <70%=poor, 70%-79%=fair, 80%-89%=good and ≥90%=excellent [15]. Of the papers that were collaboratively reviewed by all 3 researchers, there was an interrater agreement of 81%. This was a sufficient level of agreement for the remaining full-text papers to be independently screened by the primary author only. A second opinion was sought by the primary researcher during full-text screening when required.

A data extraction table was created in Microsoft Excel by the primary author. The data extraction table was piloted by the primary author for the first 10 papers before a discussion with secondary authors was conducted to ensure the appropriateness of the data being extracted. These discussions helped shape the table further with modifications made so that all relevant pieces of information were extracted. Data extracted related to key study characteristics, data analyzed in the paper, the technology employed, and the analytics techniques used.

Data were collected on paper characteristics (eg, title, authors, year of publication, location of publication, and country of origin) and study characteristics (eg, study design, stage of implementation, study setting, primary or secondary analysis, participant description, duration of study, and dropouts). Data were also captured relating to the technology in use (eg, what the technology is designed to assist with, the technology being employed, and its function), the data used in the analyses (eg, data streams, where the data are sent, and what it is being used for), and the methods of analyses employed (eg, the statistical method of analysis, the actions taken as a result of the analysis and outcome measures). Information on the key findings from each study and any potential limitations with the studies were also collected. A summary of the data extracted for each paper can be found in Multimedia Appendix 2.

This review also highlighted that there has been significantly more research into predictive analytics in telehealth (n=58) compared to telecare (n=28). Telehealth data may be more suitable to the application of predictive analytics because they are often more structured and numerical in nature whereas social care data more frequently rely on unstructured case notes.

Studies which considered a system or technology aimed at a specific disease or condition made up the majority of papers identified, with the most common disease of focus being Parkinson disease [31-42]. The extensive focus on Parkinson disease in research may be attributed, in part, to its features and symptoms and their suitability for being measured by sensors and then modeled by data analytics techniques. For example, slowness of movement, uncontrollable shaking, and gait problems are very common symptoms of Parkinson disease and are all well suited to being captured through wearable sensors. Such remote monitoring or assessment is also useful in diseases like Parkinson disease where clinical features of the disease may be intermittent in the early stages and thus may not be present during a scheduled assessment [103].

While patient-reported outcomes (PROs) were one of the more commonly featured tools in this review (n=17, 20%), they are not commonly used in telecare predictive data analytics models (n=3/28 telecare papers, 11%). PROs can provide more nuanced information than solely using clinical indicators which can lead to an underestimation of the impact on a patient in combination with an overestimation of the effectiveness of treatment being provided [104,105]. As such, there is an argument to be made for further use of PROs in predictive data analytics models, especially in the field of telecare.

Including PROs in predictive modeling is challenging as it involves the integrating both objective and subjective data. However, this integration can enhance model results by capturing the true reported experiences and outcomes of patients. Indeed, evidence shows that PRO measurements are of comparable accuracy to many objective clinical measures [106]. Appropriate testing, validation, and re-evaluation of PROs can help improve the quality and consistent collection of data while the move toward standardization of PROs through the use of tools such as the National Institute of Health’s Patient-Reported Outcome Measurement Information System (PROMIS) can enable a rise in data quality levels across the board, facilitating a greater integration of PROs in predictive modeling work [107].

Routinely collected data can be defined as data that has not been specifically captured for research purposes. There are only 3 studies featured in this review using data that have been routinely collected in real-world health and care practice, with all of these papers considering telehealth systems [31,42,68]. From a telehealth perspective, a lack of use of routinely collected data makes sense due to these systems focusing on highly specific features that need to be extracted about a given condition or illness. As such, the data considered in these systems tend to originate from bespoke, highly targeted data collection methods.

However, a significant amount of data is being generated by providers of telecare services globally as they deliver care, and the application of data analytics in these real-world data sets needs to be explored further than it has been to date. One key barrier to the analytical use of routinely collected telecare data is that these data are typically siloed in different locations, with systems lacking interoperability. For example, call handling data are frequently maintained in a different system than other social care data, resulting in the outcomes of calls being inaccessible to social care organizations. This has been identified by the Scottish Government as being a key issue preventing the use of data-driven care [108].

Additionally, work must be done to improve other issues surrounding the use of routinely collected data such as patient consent and data governance and security [109]. If researchers, care providers, and any commercial suppliers in control of these rich data sources can collaboratively overcome these identified issues, then a whole new avenue for the use of predictive data analytics will be opened.

Limitations noted by researchers were typically specific to the technology employed. These limitations include low quality data being captured [84]; the technology being uncomfortable to wear and with a short battery life [86], and there being a limited number of sensors employed [93]. Limitations related to the analytics techniques included low impact falls being missed by a model [89], large volumes of missing values [61], and a model that struggled to differentiate between an individual sitting and standing [101].

The main limitation identified in this review is that a significant number of papers are trained on very small data sets or samples. In total, there were 32 papers (47% of the total papers reporting limitations) that acknowledged this as an issue. This is a critical problem as having a small sample size could undermine the legitimacy of the findings of the paper—particularly when the outcome of interest is rare. Small sample sizes make it harder to accurately train, validate, and test ML models with the findings less conclusive and less reliable.

To ensure that the strongest evidence base possible sample size calculations should be conducted prior to the study, however only two of the papers featured in this review reported prior sample size estimation [31,79]. This may be attributed to the pragmatic nature of recruitment, where it is difficult to recruit sufficient numbers of individuals with a certain condition, but it remains critical for ensuring the validity of the findings.

The other limitation that was identified a significant number of times was the possibility of the introduction of bias to the models, as can be seen in Table 6. Bias could invalidate study findings as the model is trained on a group that is not representative (eg, gender, age) of the target population meaning that its performance may not translate in reality. In the field of telecare and telehealth, it is critical that data sets consider individuals of appropriate age—generally elderly—and that disease-specific systems have been trialed on individuals with the illness or condition of interest. For example, a study using young, healthy volunteers to classify falls—and other activities—requires participants to simulate falls [100]. This may have an impact on the accuracy of the model, and a data set featuring genuine falls captured by elderly individuals would be significantly more appropriate. The key sources of bias identified in this review are the use of exclusively young, healthy adults to trial technologies that are designed for an older population and data sets, which are dominated by women.

The quality of the studies selected for inclusion in this review was not assessed using any official appraisal tool. This is typical of a scoping review, which seeks to synthesize the available literature rather than provide a systematic analysis; however, this means that the quality of the papers featured in this review cannot be guaranteed. Another limitation of this review is that it may have missed commercially developed data analytics tools that have been implemented in practice, as these may not necessarily be documented in research literature. Finally, only papers available in the English language were considered, which may preclude a number of relevant papers from this review.

Predictive data analytics have been widely used in the field of telecare and telehealth but all of the studies featured in this review are still small-scale pilot studies and must be extended to larger trials. Additionally, opportunities for predictive analytics revolving around routinely collected data and PROs should be explored further. Using larger and more diverse “real world” data will enable models to be built that have less bias, can predict more accurately, and could be adapted more widely within other telecare or telehealth settings. Ultimately, appropriate consideration of these factors could lead us to more predictive and preventative data driven models of telecare and telehealth.