We conducted a search in three English databases (PubMed, MEDLINE, Web of Science), two Chinese databases (CNKI and WANFANG DATA), and one gray literature database (ProQuest) from the inception of the databases until December 25, 2025. A Medical Subject Heading (MeSH) search was performed to establish all relevant literature on physical activity using wearable electronic devices. Specifically, the database searches were performed using the following keywords and truncations in conjunction with the following search criteria: (Wearable Electronic Devices OR Wearable Electronic Device OR Wearable Devices OR Wearable Device OR Wearable Technology OR Wearable Technologies OR Electronic Skin OR Wearable Computer OR Wearable Computers) AND Exercise[MeSH] AND Aged[MeSH] AND English[lang]. Additionally, the reference lists of relevant meta-analyses and articles were also screened. Two reviewers independently assessed the identified publications for eligibility, with no language restrictions. Any disagreements were resolved through consensus. For specific retrieval strategies, refer to the Supplementary material.

Studies were eligible if they met the following criteria: (1) Study design: (a) parallel-group randomized controlled trials or quasi-experimental studies with a comparator; and/or (b) single-arm pre–post studies included only for design-stratified, exploratory synthesis to summarize within-group changes where controlled evidence was limited. Importantly, two-arm and single-arm studies were not pooled together in the same meta-analysis. (2) Participants: older adults (≥60 years) with any health condition. (3) Intervention: any exercise/physical activity intervention delivered with the assistance of wearable electronic devices. (4) Comparator (for two-arm studies): usual care, normal daily activities, or exercise without wearable device assistance. (5) Outcomes: at least one quantitative measure relevant to older adult health, including lower-limb strength, balance, waist-to-hip ratio, cognitive function, endurance, or mental health.

Two reviewers independently extracted study characteristics and outcome data (means, standard deviations, and sample sizes). For two-arm studies, we extracted post-intervention values (or change scores when only change data were available) for intervention and comparator groups. For single-arm studies, we extracted pre- and post-intervention values to compute within-group changes. When necessary, authors were contacted for missing information; if essential data could not be obtained, the study was excluded from quantitative synthesis.

The quality of the included studies was assessed using the Cochrane Risk of Bias tool, with particular attention to seven domains: random sequence generation, allocation concealment, blinding of participants and personnel, blinding of outcome assessment, completeness of outcome data, selective reporting, and other potential sources of bias. Each domain was rated as having a low risk, high risk, or unclear risk of bias according to the standards outlined in the Cochrane Handbook for Systematic Reviews (31), and any disagreements were resolved through consensus.

Meta-analyses were conducted using Review Manager 5.4 and StataMP 15. Continuous outcomes were synthesized using either mean difference (MD) when studies reported outcomes on the same scale and unit, or standardized mean difference (SMD) when different instruments/scales were used. Therefore, WMD was not used. Because internal validity and comparability differ across designs, we performed design-stratified meta-analyses: Two-arm controlled synthesis (primary analysis): effect sizes were computed as the between-group difference at post-intervention (intervention vs. comparator). When only pre–post change data were available, we used change-score SMD/MD as recommended in the Cochrane Handbook. Single-arm pre–post synthesis (exploratory analysis): effect sizes were computed as within-group pre–post changes. These results were interpreted as supportive/implementation-oriented evidence and were not combined with the two-arm estimates.

A random-effects model was used as the primary approach due to expected clinical and methodological variability (participants’ health status, intervention dose, device type, feedback/behavior-change components, and outcome measurement). Effect estimates are presented with 95% confidence intervals (CIs), and statistical significance was assessed using Z-tests (two-sided α = 0.05).

To ensure consistent interpretation across outcomes, we aligned effect directions a priori. Negative SMD/MD values were defined to indicate improvement for outcomes where a lower value reflects better status. Conversely, for outcomes where higher values represent improvement, data were entered so that improved performance corresponded to the same direction across studies.

To avoid unit-of-analysis errors when a study contributed multiple measures within the same outcome domain (e.g., multiple balance tests) or multiple time points, we applied the following rules: One effect size per study per outcome domain per synthesis. When multiple instruments assessed the same domain at the same time point, we prioritized the most prespecified/clinically common measure (e.g., for balance: Berg Balance Scale when available), or otherwise selected the measure identified as the primary outcome by study authors. If multiple eligible time points were reported, we used the immediate post-intervention assessment as the primary time point. Longer follow-up data were summarized narratively or analyzed in sensitivity analyses when sufficient studies were available. When two distinct subgroups within a single study were eligible, we combined groups using Cochrane-recommended formulas to avoid double-counting the comparator.

Statistical heterogeneity was assessed using the I² statistic. We interpreted I² values using conventional thresholds (25% low, 50% moderate, 75% high). When heterogeneity was very high (I² > 75%), we implemented a prespecified interpretation plan: verify data extraction, effect direction harmonization, and outcome-domain classification; conduct influence (leave-one-out) analyses to identify outliers; perform subgroup analyses when ≥2 studies per subgroup were available; if substantial heterogeneity persisted and clinical/methodological diversity was judged to be too large for meaningful pooling, we refrained from meta-analysis for that outcome and provided a structured narrative synthesis instead. Publication bias was assessed when ≥8 studies were available for an outcome using funnel plots and Egger’s test (Stata).

The initial database search yielded 5,017 publications. Subsequent screening resulted in 17 papers eligible to be included in the meta-analysis (Figure 1).

This study included 611 participants (age range: ≥60 years) from the included studies. The studies consist of 13 group comparison studies and 4 single-group analyses, which were mainly conducted in China, USA, Korea, and Japan (Figure 2). Detailed descriptions of the study participants are provided in Tables 1–3.

Table 1 provides an overview of the participant characteristics in the two-arm controlled trials included in this meta-analysis. The included studies involved healthy and frail populations, with participants aged 60 years and older. The sample size of the experimental group (EG) ranged from 9 to 40 participants, while the control group (CG) had a sample size ranging from 8 to 40 participants. The studies were conducted in multiple countries, including China, South Korea, the United States, and Japan, with designs including randomized controlled trials (RCTs) and quasi-experimental designs (QEs). The participants’ health status ranged from healthy and pre-frail to frail, and the outcomes assessed various aspects of physical health, including muscle strength, balance, endurance, and cognitive function. The diversity in sample characteristics, outcome measures, and measurement tools across these studies provides a broad basis for evaluating the impact of wearable device-based exercise interventions on the health outcomes of older adults.

Table 2 summarizes the intervention protocols in the two-arm controlled trials included in the meta-analysis. The table provides detailed information about the intervention methods, including the types of electronic devices used, the nature of physical or cognitive training, the frequency and duration of the interventions, and the intensity levels. Most studies involved structured physical interventions, such as aerobic exercise, body training, or cognitive training, with a frequency of 1–3 sessions per week, each lasting 30–90 min. The intervention duration ranged from 3 to 12 weeks, depending on the study. Additionally, various technologies were used in the studies, such as wearable devices, mobile phones, and virtual reality devices, providing multiple monitoring and intervention methods for improving the health of older adults. These differences in intervention designs reflect various approaches to improving health outcomes in older adults through wearable technologies and structured physical activity programs.

Table 3 summarizes the participant characteristics and intervention protocols in the single-arm trials. The table includes intervention information from four studies, which were conducted in different countries such as South Korea, China, and the United States, and all participants were aged 65 or older. The participants health statuses included healthy, frail, and hypertensive individuals. The interventions primarily involved a combination of exercise, resistance training, and aerobic exercise, with intervention periods ranging from 4 to 20 weeks, and frequency from 3 to 5 days per week, with session durations between 30 min and 150 min. The electronic devices used included assistive robots, AI devices, and digital health platforms to support exercise and health management. The results indicate that exercise interventions based on wearable electronic devices can effectively improve older adults individuals’ physical function, exercise capacity, and quality of life.

The methodological quality of the included controlled studies was assessed using the Cochrane Risk of Bias tool (RoB 1.0), following the guidance in the Cochrane Handbook (32). Seven domains were evaluated: random sequence generation, allocation concealment, blinding of participants and personnel, blinding of outcome assessment, incomplete outcome data, selective reporting, and other bias. Each domain was judged as low risk, unclear risk, or high risk according to the Cochrane criteria. Overall, most studies were judged to have low risk of bias in blinding of outcome assessment, completeness of outcome data, and selective reporting (Figure 3). Some studies were rated as unclear or high risk in random sequence generation, allocation concealment, and blinding of participants/personnel. For transparency, domain-level judgments for each study are presented in Figure 3. In addition, we defined an overall risk-of-bias classification rule a priori: studies were considered overall low risk if all key domains were low risk; overall moderate risk if at least one key domain was unclear risk but none were high risk; and overall high risk if at least one key domain was high risk. This overall classification was used only for descriptive summary purposes, while interpretation focused on domain-level assessments.

For older adults, eight two-arm studies evaluated lower-limb strength. Wearable-supported exercise was associated with an improvement in lower-limb strength compared with control (SMD = −0.60; 95% CI [−1.15, −0.05]; I² = 81%). Three single-arm pre–post studies reported within-group improvements (SMD = 0.51; 95% CI [0.05, 0.97]; I² = 23.5%). As prespecified, single-arm estimates are presented separately and were not combined with controlled effects.

Four two-arm studies assessed balance outcomes. A pooled improvement in balance was observed (SMD = −0.43; 95% CI [−0.81, −0.06]; I² = 17%). Three single-arm studies reported no statistically significant pre–post change in balance (SMD = 0.42; 95% CI [−0.48, 1.33]; I² = 79.8%). Given the small number of studies and substantial heterogeneity, these findings should be interpreted cautiously (Figure 4).

Subgroup analyses were originally conducted to explore heterogeneity for lower-limb strength; however, several subgroups contained very few studies (often n = 2). In line with best practice, we now present these subgroup results as exploratory, the results showed that the best improvement was observed with one intervention per week, lasting 10–45 min per session, and a total intervention duration of 8–12 weeks. No other subgroup differences were found. The subgroup tables have been retained for hypothesis generation (Table 4). Meta-regression did not identify intervention frequency, session duration, or intervention period as statistically significant moderators (p > 0.05), noting the limited power given the number of available studies.

To explore the sources of heterogeneity, we conducted a sensitivity analysis (Figure 5), and found that excluding individual studies had minimal impact on the overall effect estimate, indicating high consistency and reliability of this meta-analysis. A funnel plot (Figure 6) was used to assess publication bias. The funnel plot appeared nearly symmetrical; however, given that some outcomes involved fewer than 10 studies, the interpretation of the plot should be considered with caution, as small sample sizes can influence its symmetry. Additionally, the Egger test showed a p-value of 0.224 (p > 0.05), suggesting that there was no significant publication bias in the study, although this result should be interpreted with caution due to the limited number of included studies. To further examine potential sources of heterogeneity, we performed a meta-regression analysis, which revealed that intervention frequency (p = 0.640), duration (p = 0.387), and period (p = 0.143) were not significant sources of heterogeneity (p > 0.05). This strengthens the robustness of the findings and indicates that the observed effects are not influenced by these factors.

Three two-arm controlled studies reported endurance outcomes. The pooled estimate was not statistically significant and was highly heterogeneous (SMD = 2.48; 95% CI [−0.57, 5.52]; I² = 97.6%). Two two-arm controlled studies reported cognitive outcomes. The pooled effect was not statistically significant (SMD = 0.53; 95% CI − 0.02 to 1.08; I² = 0.0%). Five two-arm controlled studies reported upper-limb strength outcomes. The pooled effect was not statistically significant (SMD = −0.07; 95% CI −0.59 to 0.46), with high heterogeneity (I² = 75.8%). Two two-arm controlled studies reported mental health outcomes. The pooled effect was not statistically significant (SMD = 1.09; 95% CI −1.05 to 3.23), with very high heterogeneity (I² = 94.5%). Two single-arm pre–post studies reported waist-to-hip ratio outcomes. The pooled within-group estimate was not statistically significant (SMD = −0.03; 95% CI − 0.50 to 0.90; I² = 0.0%). For outcomes with very high heterogeneity (e.g., endurance and mental health), pooled estimates were interpreted cautiously as they reflect substantial between-study variability and imprecision due to the small number of studies (Figure 7).

Future research should prioritize well-designed controlled trials with adequate sample sizes and standardized outcome assessments for strength and balance, and should also expand high-quality evidence for other domains. To reduce heterogeneity and improve interpretability, trials should report intervention components in detail (exercise modality, supervision, device feedback functions, behavioral strategies) and predefine primary outcomes. Mechanistic questions—particularly whether personalization features add benefit—should be addressed using factorial or head-to-head designs that compare tailored vs. non-tailored wearable support. Finally, longer follow-up and consistent reporting of adverse events will be essential to evaluate sustainability and safety in older populations.