Disability is a major public health issue. Currently, more than 1 billion people (about 15% of the world’s population) live with short- and long-term disabling conditions [1]. In a lifetime, almost all individuals will experience a temporary or permanent disability. Health conditions leading to disability can affect a wide range of systems, such as the musculoskeletal, metabolic, cardiocerebral and vascular, nervous, or respiratory systems. In most cases, exercise programs are either the primary treatment or an essential adjunctive treatment to reduce short- and long-term disabling conditions [2,3]. In this document, short-term and long-term disabling conditions refer to people living with acute or chronic functional limitation, regardless of the location or type of impairment [4].

However, the effective implementation of exercise programs is limited by many factors. For health care professionals, the prescription of clear and detailed exercises can be difficult in the absence of adapted tools (eg, to transmit detailed exercise modalities). For individuals with short- and long-term disabling conditions, the first difficulty may be the availability of the health professional [5]. Then, when a program has been prescribed, other challenges may arise when performing exercises independently; the main challenge is to remember which exercises to perform and to stay motivated throughout the treatment process. As a consequence, the implementation of exercise programs is generally suboptimal [6].

In recent years, we have seen the emergence of mobile health technologies (mHealth), which may be very useful to facilitate the implementation of exercise programs [6]. Mobile health, or mHealth, is a subset of eHealth and is defined as “the use of mobile wireless technologies for health” [6]. mHealth offers easily accessible interventions that can reach large populations. In addition, interventions are low cost, easily adapted, reduce costs (such as travel), and save time [7-9]. The recent pandemic has highlighted the need and potential for telehealth solutions in situations in which travel is limited, and the health system is restricted [10]. These solutions can take the form of web-based platforms or apps offering physical exercise programs (generic or individualized to the person’s needs) prescribed by a health care professional or automatically provided by an algorithm embedded in the app [11]. One type of solution may be particularly useful in helping the therapist prescribe exercises and promote patient compliance: digital apps that allow us to design personalized exercise programs (individualized to the person’s needs), prescribed by a health professional (unlike programs proposed by algorithms, which cannot propose fine adjustments), and are accessible in the form of videos for a more faithful execution of the exercises. We will call such apps: apps providing personalized exercise videos [2,12,13]. To increase the use of these apps so that they may benefit as many people as possible, their effectiveness must be demonstrated [6,14].

To date, no systematic review has specifically evaluated the effectiveness of apps providing personalized video exercises. Existing reviews mix a wide variety of digital solutions, ranging from texting and phone calls to rehabilitation sessions via videoconference [7,11,13,15-17]. However, the heterogeneity of the studies included in those reviews prevented firm conclusions from being drawn on the effectiveness of digital solutions [18].

This systematic review aimed to synthesize the recent scientific literature on the impact of apps providing personalized video exercises to support rehabilitation for people with short- and long-term disabling conditions, on functional capacity, confidence in exercise performance, health care consumption, health-related quality of life, adherence, and adverse events.

This systematic literature review was conducted in accordance with the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines [19] (Multimedia Appendix 1) and the Cochrane Handbook of Systematic Reviews [20]. The review protocol was registered with the International Prospective Register of Systematic Reviews (PROSPERO reference CRD42022323670) before commencement of the study. Deviations from the protocol are presented in Multimedia Appendix 2 [12,21,22].

We systematically searched the following databases: MEDLINE, CENTRAL, and Embase until March 1, 2022. We also searched clinical trial registries to identify ongoing studies, unpublished studies, and published studies not identified by the electronic search: ClinicalTrials.gov and WHO international clinical trials registry platform. We also searched the PROSPERO international prospective register for systematic reviews. Manufacturer’s websites of apps were searched (Physitrack, Kaia, Hinge, Swordhealth, and Caspar-health). One journal in the field (JMIR) was specifically searched. We manually searched the reference lists, studies that cited relevant studies as well as related studies in PubMed to identify any additional studies.

The search equations for the different databases are described in Multimedia Appendix 3. These search terms were extended with specific terminology and synonyms using Boolean operators and the respective Medical Subject Headings. No specific filters or limits were used, and the language was not restricted.

The web-based data manager Rayyan.ai (Rayyan) was used for the selection process. After removal of duplicates, selection of relevant trials was performed by 2 independent authors (TD and PM) following the eligibility criteria defined above. The selection was made on the basis of title, summary, and then full text individually. A consensus was then sought on the papers finally selected. A third person was to be contacted in case of disagreement on the consensus but this was not necessary.

Data were independently extracted by 2 authors (TD and PM) using a data extraction form previously tested with 3 trials [12,23,24]. When data were missing or unclear, we contacted the corresponding author for clarification. Information related to the intervention and the app was gathered from information available in the paper and the manufacturer’s website.

The data extracted included general characteristics (authors, year, study design, country of origin, and context of inclusion), the study population (pathology, number of participants, health conditions duration, body mass index, age, and sex), intervention details (app used and mode of delivery, duration, whether the device was currently on the market or not, components related to the digital therapeutic tool, and the exercise dosage), and general information about comparator group and outcome measures used in relation to our objectives (the measurement tool and the corresponding scale, the evaluation time at the end of the intervention and the mean, SD, and number of participants analyzed in each group at the end of the intervention). If there was an imbalance at baseline, we collected the difference between baseline and end of the intervention in each group. When several analyses were presented for the same outcome and assessment time point, we extracted the data from the intention-to-treat analysis and the analysis that used the most robust missing data imputation technique.

The methodological quality of the included studies was assessed using the Cochrane Risk of Bias tool for quality assessment of randomized controlled trials (RoB 2) by 2 independent authors (TD and PM) [25]. Any discrepancies were resolved by discussion. To account for bias caused by deviations from the intended interventions, this review focused on the effect of assignment to the interventions at baseline (regardless of whether the interventions were received, or the level of participant adherence during follow-up, known as the “intention-to-treat effect”). The risk of bias was assessed according to the following domains: bias arising from the randomization process, bias due to deviations from intended interventions, bias due to missing data, bias in measurement of outcomes, and bias in selection of the reported result. The RoB 2 Excel tool was used to complete the risk of bias assessment. We judged each outcome as being at low risk, some concerns, or high risk according to the RoB 2 algorithm. The risk of bias assessment was incorporated in the Results section of the review, and it was also part of the Grading of Recommendations, Assessment, Development, and Evaluations (GRADE) assessment of the certainty of evidence (along with precision, directness, consistency, and publication bias).

For quantitative outcomes, the magnitude of effect was assessed with the mean difference (MD) when possible or with the SMD. Data from medians and IQRs were converted to means (SDs) [26]. For binary outcomes, the measure of effect was a risk ratio.

If the data were sufficiently homogeneous, the effects of the interventions were pooled using random effects models in RevMan (version 5.3; Cochrane Collaboration). To interpret the SMD, the following thresholds were used: no effect (<0.2), small effect (0.2-0.5), moderate effect (0.5-0.8), or large effect (>0.8) [27]. Heterogeneity was evaluated visually on the forest plots, using the heterogeneity test (P<.05 indicating significant heterogeneity) and I² statistic that measures the proportion of variation (ie, inconsistency) between studies that is caused by heterogeneity rather than chance [28]. An I² value of 0% to 40% might not be important; 30% to 60% may represent moderate heterogeneity; 50% to 90% may represent substantial heterogeneity; and 75% to 100% considerable heterogeneity [20]. Where there were insufficient data to pool studies, a narrative synthesis of the studies was conducted. As there were fewer than 10 trials per meta-analysis, we were unable to draw a funnel plot and perform an Egger test to assess small study effect [29].

The quality of evidence for each outcome was assessed using GRADE to evaluate the following domains: study limitation (risk of bias), inconsistency, indirectness, imprecision, and publication bias [30].

A total of 1641 titles and abstracts were screened after excluding duplicates, of which 1599 records did not meet the inclusion criteria (Figure 1). The full texts of 42 potential eligible records were read, and 10 papers published between 2017 and 2021 were included [12,21-24,31-35]. The list of papers excluded at the full-text selection stage with reasons for exclusion is presented in Multimedia Appendix 4. The 10 selected papers included a total of 1050 individuals (range across studies 20-305). A summary of the study characteristics is presented in Table 1. Eight studies involved adults, 1 involved children (aged 6 to 17 years) [32], and 1 involved people aged >60 years [23]. Overall, 7 studies included people with musculoskeletal disorders, and 3 included people with neurological disorders. Among these, 2 studies concerned employees and their dependents [22,24], and 3 involved rehabilitation after surgery [31,34,36].

A summary of the characteristics of interventions is presented in Table 2. The duration of intervention ranged from 3 to 48 weeks. All intervention groups received an exercise program with videos, tailored to each participant’s needs and available on an app. All trials except one [31] allowed communication between participants and therapists through the app. Exercise programs were designed by a physiotherapist in 8 studies [12,21,22,24,31,35,36,38], by a medical doctor in 1 study [34], and by an occupational therapist in 1 study [23]. Three studies provided real-time biofeedback through wearable motion sensors for the group using the app [22,24,36], 1 study provided feedback on steps per day, graphically over time for the intervention group and with simple information on the number of steps per day for the control group [35]. Seven studies used commercially available apps [12,22,24,35-38]. Six of the 10 included studies involved a comparison group receiving a similar exercise program as the intervention group but without the app (eg, on paper and with screenshots) [12,21,31,35,37,38]. The other 4 studies involved comparator groups that were significantly different from the intervention groups: more supervised sessions for 1 [36] and fewer components for 3 [22,24,34].

The results of the risk of bias assessment of the included studies are presented in Table 3. All included trials were rated as high risk of bias for at least 1 domain. The 2 domains with the highest proportion of high risk of bias were “bias due to deviations from intended interventions” and “bias in measurement of outcomes” because of the impossibility of blinding. The 2 domains with the highest proportion of low risk of bias were “bias arising from the randomization process” and “bias in selection of the reported result.”

A summary of findings for the comparisons with GRADE ratings is presented in Table 4. Detailed meta-analytic forest plots are also presented in Figure 2 and Multimedia Appendix 5.

Eight studies involving 644 participants reported the effect of an app providing personalized video exercises to support rehabilitation compared with the absence of apps on physical function at the postintervention time point (range 3 to 48 weeks, median 12 weeks; Figure 2). Details on how physical function was evaluated in included trials are reported in Multimedia Appendix 6. Overall, a significant small effect in favor of the intervention was observed (SMD 0.35, 95% CI 0.19-0.51); heterogeneity was low (Phet=.86; I²=0%). Using the GRADE approach, the quality of evidence was rated as low because of the high risk of bias in most studies.

Two trials (358 participants) assessed confidence in the ability to undertake exercise on an 11-point numerical rating scale (0=strongly disagree and 10=strongly agree). Overall, a significant, small effect was found in favor of the intervention (MD 0.67, 95% CI 0.37-0.96); the heterogeneity was moderate (Phet=.22; I²=33%). The quality of evidence was rated low because of the high risk of bias in most studies.

Health care consumption was assessed in 3 studies (637 participants). All reported a statistical difference in favor of the intervention group. No meta-analysis was carried out owing to the high heterogeneity of the outcomes (Multimedia Appendix 7 [12,22,24]).

Overall, no significant difference was found between groups in the 3 studies assessing health-related quality of life (237 participants; SMD 0.42, 95% CI −0.09 to 0.93); heterogeneity was substantial (Phet=.08; I²=60%). The quality of evidence was rated very low because of the high risk of bias in most studies, small sample size, and substantial heterogeneity.

Among the 6 studies assessing adherence, 3 studies (n=375 participants) reported a statistical difference between intervention and control groups, while the 3 others (n=135 participants) reported no significant difference. No meta-analysis was performed because the adherence criterion was measured differently each time (Table 4).

Adverse events were assessed in 7 studies (n=959 participants). Overall, no significant difference was found between groups (risk ratio 0.68, 95% CI 0.37-1.23); heterogeneity was low (Phet=.38; I²=5%). The quality of evidence was rated as low because of the high risk of bias in most studies. Most studies did not explicitly mention the search for adverse events in the Methods section or in the protocol. No serious adverse events were reported in the intervention group of the 7 studies.