A randomized controlled trial was conducted, with participants randomly assigned to three groups: (1) experimental group (personalized exercise); (2) active-control group (World Health Organization [WHO] general exercise guidelines); and (3) control group (no intervention).

Based on a statistical power analysis using G*Power [23], our target was to enroll 300 participants, with 100 in each of the 3 treatment groups (see more details in our protocol paper [14]). However, to maintain methodological rigor, we later conducted a more precise power analysis based on the actual number of participants included in the statistical analyses (see Results section).

The study was conducted in a community setting (private homes and independent living facilities) in central and northern Israel, including both urban and agricultural settlements. Recruitment included local flyers and lectures delivered by the principal investigator in community clubs, senior centers, cultural centers, or other local gathering places. Participant enrollment was carried out by the study team, which consisted of 6 testers and a study manager.

From a total of 317 volunteers, 239 community-dwelling, healthy older adults (155 women), aged 72.63 (SD 5.38) years, completed the 8-week study intervention, and 230 completed the full 12-week follow-up. Inclusion criteria were (1) age 65+ years, (2) independent living, (3) independent walking, (4) fluent in Hebrew, and (5) smartphone competency. Exclusion criteria included (1) cognitive impairment (Mini-Cog score <3) [24], (2) any hospitalization (>24 h) or emergency room visit within past year for cardiac (heart failure, rhythm disorder, ischemia, valvular disease) or neurological conditions (dizziness, cerebrovascular, vestibular, progressive diseases affecting gait or balance), and (3) high fall risk (≥1 positive answer from 3-item test), validated in community-based exercise intervention and fall-prevention studies [25]. Data collection included demographics, self-reported habitual physical activity, depression status (Geriatric Depression Scale) [26], and frailty index [27].

The study design, procedures, and informed consent were approved by the Hadassah Hebrew University Hospital Ethics Committee, Jerusalem, Israel (Trial ID 0074‐19-HMO), and were conducted according to the ethical standards of the Helsinki Declaration.

We used a permuted block design [7] with 2 stratification factors (age: 65‐74 and 75+ years; gender: men and women), each block included 6 participants, 2 per group. Block distribution was adjusted using the minimization principle [28] in order to address randomization imbalance. More specifically, as there were 6 testers, and each tester had his or her randomization table. This process was supervised by an advanced student (EB).

At a certain point, we realized that few testers had the first participants on the control group, and the control group became a lot larger than the other groups. Based on the minimization principle, we changed the block distribution to 2 experimental, 2 active-control, and only 1 control (5 participants in a block).

The motor components chosen for remote fitness assessment were postural control (stability) while standing (static balance) or moving (dynamic balance), strength (muscle endurance) of upper and lower body, and range of motion of upper body (flexibility). The following standard movement performance tests were selected to assess these components:

Instructions were incorporated into a smartphone app developed by Montfort Brain Monitor Ltd. Movement was captured by smartphone accelerometer and gyroscope sensors by attaching the mobile phone to the relevant body part using a simple band (see Figure 1: tandem walk forward, arm flexion, arm extension, arm strength).

The specific digital markers (DMs) generated from the phone for each test are as follows:

Detailed information and graphical demonstrations of the DMs have been previously published [14,15].

Data collection started in November 2020 and ended in September 2023. Following informed consent, demographic/clinical data collection, and randomization, baseline fitness (T0) was assessed using the study app. Fitness levels across the study measures were determined (low versus high), and assessment was repeated at 4 weeks (T1), 8 weeks (T2), and 12 weeks (follow-up T3). The assessment was conducted by 6 qualified physical activity teachers, each trained to use the app for testing participants individually and supervised by the study manager, a senior physical activity teacher. The app was installed on the teachers’ (testers’) smartphones, and the test results were automatically streamed to the database. All study participants (the personalized exercise group, the active-control group, and the control group) were tested. The testers and participants were blinded to the test results. Participants received weekly phone calls from the teachers to maintain contact and motivation. There were no issues of safety or other issues reported.

Data from the pilot study [15] served as a baseline for large-scale data collection, and the baseline (T0) fitness level of participants in the current study (low or high) was determined using machine learning principles. Based on the fitness level, the app determined the appropriate level of exercise difficulty (A or B) for each fitness component. For example, a participant in the personalized exercise group might have been assigned level A for balance exercises, level B for flexibility, and level B for strength. Figure 3 illustrates an example of a study participant’s unique fitness profile, based on the DMs, graphically displayed alongside the average profile for the entire study sample. Repeated fitness assessments at 4 weeks (T1) allowed for the personalized exercise program to be adjusted according to the updated DMs.

For comparability reasons due to different units of measurement, we transformed the DM data into normalized (z) scores. For balance measures, we calculated the mean of the 5 generated DM scores to one for each balance test. Results are presented as z scores.

Since participants needed to access the study app in order to watch and perform exercise videos, it was therefore possible to determine the actual and accurate measurement of adherence. Although they were instructed to exercise 5×/week, the actual adherence varied. To explore whether adherence influenced improvements, we conducted an exploratory analysis categorizing participants as adherers or nonadherers. The following cutoff points, along with their rationale, were examined:

A consistent trend favoring adherers over nonadherers emerged across all cutoff points (see Multimedia Appendix 1). The most notable differences were observed with the ≥2 sessions/week versus <2 sessions/week and ≥1.5 sessions/week versus <1.5 sessions/week cutoffs.

Ultimately, we selected 1.5 sessions/week as the final cut-off for two reasons:

Subsequently, we conducted statistical analyses using four groups: (1) personalized exercise adherers (n=66), (2) personalized exercise nonadherers (n=26), (3) general guidelines exercise (active-control; n=80), and (4) control (n=67).

We applied statistical analyses to the 3-time measurements during the intervention period, T0, T1, and T2, and examined whether improvements observed after 8 weeks (T2) were maintained at 12-week follow-up (T3). Specifically, a mixed repeated measures ANOVA (3 test dates×4 groups) was conducted for each fitness component, and Eta squared (η²) was calculated to assess the effect size. Fisher LSD was used for pairwise post hoc analyses, and Cohen d coefficients were calculated to reveal the standardized differences between means when effects reached a significance level (P<.05).

To examine whether improvements were maintained at T3, we examined participants from the general group and personalized adherers who demonstrated improvement from T0 to T2 (between the baseline and the end of the 8-week intervention period). Improvement was defined as a change of ≥0.12 in z score (zT2-zT0), corresponding to the 55th percentile with normal distribution. More specifically, individuals with T2-T0 z scores ≥0.12 were considered “improvers.” Two-way ANOVA (2 test dates×2 groups) with repeated measures was applied on the outcome measures.

Due to incomplete compliance during 1 or more measurements, or failure to upload measurements due to poor internet connectivity, specific data points for certain individuals were excluded. Additionally, scores greater than 2.5 SDs from the raw data means were omitted. As a result, the number of participants in each ANOVA (assessing the effect of the intervention on each fitness component) ranged from 189 to 231. The participation rate (%) for each group was calculated for the main outcomes (see Multimedia Appendix 2). χ² analyses revealed no significant differences. The power calculation was based on a sample size of n=189.

Figure 4 illustrates the participant flowchart across the 3 original study groups. Of those who completed the 12 weeks, 90 were from the personalized exercise group, 78 from the general exercise group, and 62 from the control group.

Table 1 presents baseline data.

We used G*Power [30] to perform power analysis for a 2-way ANOVA (3 test dates×4 groups) with repeated measures on the outcome measures. Our sample (n=189) provided 91% statistical power to find an interaction with small effect (Cohen f=0.1), 91% power to find group differences of moderate effect (Cohen f=0.25), and 97% power to find within measurements differences of small effect (Cohen f=0.1). All power analyses used a correlation of 0.685 among repeated measures because, among all our primary outcomes, this was the lowest (the highest was 0.856) and thus the most conservative value to use.

Figure 5 describes the main results. Group×time interaction was significant for dynamic balance (mean tandem walk forward and backward, F_6,404=3.232, P=.004, η²=0.046; Figure 5A). Pairwise analyses indicated significant improvements among personalized exercise adherers (Adherers) from T0 to T2 (M_diff0,2=0.228, P=.002, d=0.404) and group differences in favor of the Adherers in T2 (M_grp1,3=−0.357, P=.02, d=0.456; M_grp1,4=−0.383, P=.01, d=0.474; Figure 5A). The interaction on static balance (mean of leg stance and tandem stance) was not significant, but group differences favored the Adherers at T1 (M_grp1,3=−0.375, P=.01, d=0.430; M_grp1,4=−0.361, P=.02, d=0.446; Figure 5A).

Group×time interactions were revealed on both arm flexion (mean right and left, F_6,448=2.527, P=.02, η²=0.033) and arm extension (mean right and left, F_6,450=2.753, P=.01, η²=0.035; Figure 5B). Follow-up pairwise analyses indicated significant improvement in the Adherers on arm flexion (left and right) from T0 to T2 (M_diff0,2=−0.227, P=.007, d=0.356) and on arm extension (left and right) from T0 to T1 (M_diff0,1=−0.221, P=.02, d=0.302) and from T0 to T2 (M_diff0,2=−0.210, P=.03, d=0.290; Figure 5B).

Group×time interaction was also indicated on arm strength total (mean of lifting right arm forward, left arm forward, left arm to the side, and right arm to the side; F_6,424=2.394, P=.03, η²=0.033; Figure 5C). Follow-up pairwise analyses indicated significant improvement among the Adherers on arm strength total from T0 to T1 (M_diff0,1=0.228, P=.008, d=0.438) and from T0 to T2 (M_diff0,2=0.217, P=.008, d=0.384; Figure 5C).

More results are presented in the tables below. Regarding balance, group×time interactions were significant for tandem stance left (left foot forward, η²=0.033), tandem walk forward (η²=0.039), tandem walk backward (η²=0.034), and balance total (mean of all balance scores, η²=0.040; Table 2). Pairwise analyses indicated improvements (time differences) only in the Adherers from T0 to T2 (M_diff0,2=0.163, P=.009, d=0.259), with group differences favoring the Adherers at T1 (M_grp1,3=0.344, P=.007, d=0.470) and at T2 (M_grp1,3=0.399, P=.003, d=0.520, M_grp1,4=0.347, P=.01, d=0.480; Table 2).

Regarding flexibility, group×time interactions were revealed on right arm flexion (η²=0.038), right arm extension (η²=0.032), left arm (mean flexion and extension, η²=0.035), and right arm (mean flexion and extension, η²=0.054; Table 3). Pairwise analyses indicated significant improvements only among the Adherers on right arm flexion from T0 to T2 (M_diff0,2=−0.321, P=.002, d=0.425), on right arm extension from T0 to T1 (M_diff0,1=−0.264, P=.02, d=0.293), and from T0 to T2 (M_diff0,2=−0.262 , P=.03, d=0.309), on right arm (mean flexion and extension) from T0 to T1 (M_diff0,1=−0.205, P=.008, d=0.341) and from T0 to T2 (M_diff0,2=−0.283, P<.001, d=0.625), and on left arm (mean flexion and extension) from T0 to T1 (M_diff0,1=−0.192, P=.009, d=0.359; Table 3). No change was observed on torso rotation (Table 3).

Concerning arm strength, a group×time interaction was observed for lifting right arm to the side (η²=0.034; Table 4). Specific improvements were noted in the Adherers from T0 to T1 (M_diff0,1=0.247, P=.02, d=0.395) and from T0 to T2 (M_diff0,2=0.213, P=.02, d=0.327; Table 4). Pairwise differences were observed for arm strength variables, although no interactions were found: the Adherers improved on right arm forward from T0 to T1 (M_diff0,1=0.271, P=.006, d=0.444) and from T0 to T2 (M_diff0,2=0.308, P=.002, d=0.466), on left arm forward from T0 to T1 (M_diff0,1=−0.215, P=.03, d=0.332) and from T0 to T2 (M_diff0,2=−0.266, P=.005, d=0.399), on left arm to side from T0 to T1 (M_diff0,1=0.201, P=.03, d=0.331) and from T0 to T2 (M_diff0,2=0.179, P=.04, d=0.272; Table 4).

No change was observed in lower body strength (sit-to-stand; Table 4).

No group×time interaction was observed on any of the outcomes. Significant (P<.05) time effects were revealed, indicating significant deterioration in both groups—the personalized Adherers and the general groups—during the 4-week follow-up.

We examined the hypothesis that remotely delivered, personalized multicomponent exercise for older adults—based on a simple yet reliable and accurate smartphone motor fitness assessment and individually tailored using machine learning—can improve balance, flexibility, and strength among older adults, all without the need for a laboratory or professional supervision.

The results of this randomized controlled study demonstrated that a multicomponent personalized exercise program among older people was more effective when compared to active controls who were counseled to perform regular exercise according to WHO guidelines, as well as a passive control group (who received no intervention). Furthermore, improvements were achieved with as few as 1.5 tailored exercise sessions per week over 8 weeks and in many fitness measurements, even within 4 weeks.

The study tool has been refined from the initial prototype and proof of concept [14-16] and in its present form was proven to be sufficiently sensitive to measure both improvement and decline across the motor fitness measures, in as few as 4 weeks following the initiation or cessation of the exercise intervention, respectively.

The improvement in balance is of particular importance, as the relationship between balance impairment and falls [31] and the positive effect of exercise on balance and fall prevention among older people is well documented [18]. However, exercise interventions typically involve group sessions in neighborhood clubs or physical therapy laboratories [32], requiring a specialized trainer or therapist, and mostly target participants at risk of falls [32,33]. Balance control is multifactorial [17,34], and the novelty of the study tool presented in this study is not only its potential widespread application but also the fact that it includes 4 static and 2 dynamic balance tests, each comprising 5 DMs assessing body sway in all directions (mediolateral, anterior-posterior, and superior-inferior), offering a complete overview of the participant’s balance. Simultaneously, the balance-focused segment of our exercise programs addresses all components of balance, encompassing 5 subcategories: static, dynamic, vestibular, leg strength, and leg flexibility [35].

Flexibility is a fitness component known to typically decline with advancing age, presenting a major contribution to reduced functioning [19]. Although amenable to improvement through focused exercise [20], nonetheless, current guidelines provide only minimal detailed attention, and consequently, this element of fitness often remains neglected [9,10]. Our findings revealed significant improvement in flexibility in as little as 4 weeks of tailored exercise for upper body (arm) flexibility (both extension and flexion), which was most pronounced among the personal adherers, while measures of torso rotation remained unchanged. In all likelihood, this lack of observed change in torso flexibility is an artifact due to the technical difficulty measuring this vector with the smartphone.

The importance of upper body strength is notable with advancing age, with weakness associated with painful symptomatology [36], upper limb dysfunction, and decreased performance measures [21]. It should be noted that upper extremity strength can be represented by several measurements, such as the widely used arm curl test [37,38]. We chose to focus on the shoulder joint because age-related decline in maximal upper extremity torque is greater at the shoulder than at the elbow and wrist joints [39].

In contrast to the shoulder, we did not observe significant differences in leg strength, although a trend for improvement was seen in the intervention group. In our pilot study, we previously observed improvements in leg strength using sit-to-stand repetitions over a 30-second period [14,15]. However, to simplify testing, we modified this measure to a time score for 10 repetitions, which led to a ceiling effect. Importantly, leg strength is essential for functioning in older adults [40,41], and evaluating the impact of exercise on this component is crucial. Since the sit-to-stand repetitions (as many repetitions as possible over 30 s) were shown to be sensitive to changes following exercise, we plan to return to this method in future assessments.

Our study is unique in using a simple yet accurate smartphone platform to incorporate a remote assessment of balance, flexibility, and strength to serve as the baseline for a subsequent matched range of exercises, spanning these modalities of motor fitness. The existing models rely heavily upon fitness trackers, built-in accelerometers, and on occasion GPS to remotely monitor steps, endurance, and energy consumption. However, the critical elements of balance, flexibility, and shoulder strength are neglected, despite their crucial importance among older people. It is worth mentioning that recent evidence is emerging from the rehabilitation setting, supporting the use of smartphones for assessing balance or flexibility (range of motion) in older adults; however, mostly for highly specific conditions requiring clinical supervision [42-45].

Notably, while the platform demonstrated sufficient sensitivity to detect changes at a low threshold (1.5 sessions per week over 4 weeks), it remains uncertain whether this frequency is adequate to produce meaningful clinical health benefits. This concern is further emphasized by the small effect sizes observed—despite their statistical significance—and by the higher frequency recommended in WHO guidelines. Based on our findings, we cannot assert that this level of activity is sufficient for inducing clinical health improvements.

Similarly, the deterioration observed after the 4-week follow-up period without exercise was also subtle, suggesting that short-term inactivity may not necessarily lead to clinically relevant declines in fitness. This aligns with the broader detraining literature, which indicates that the rate of fitness loss is highly dependent on the initial training volume—encompassing factors such as intensity, duration, and frequency. Individuals with a higher training volume tend to experience a more gradual decline in fitness, whereas those with lower initial levels may exhibit more rapid losses [46-48].

In our study, the smartphone app was able to detect subtle yet significant improvements in fitness after 4 weeks of exercise, which then disappeared following 4 weeks of inactivity. While neither of these changes is likely to translate into meaningful clinical health effects, they underscore the application’s sensitivity in capturing even small fluctuations in fitness over time—both in improvement and decline. This capability highlights its potential as a tool for remotely monitoring fitness trends that might otherwise go unnoticed outside of laboratory settings.

An interesting finding of our study is the notable relationship between participants’ habitual physical activity and sedentary time before starting the intervention and their subsequent adherence to the exercise regimen. Significant differences were observed between adherers and nonadherers, with adherers engaging in more habitual physical activity and nonadherers spending more time sedentary. It is important to note that the classification into adherers and nonadherers was made post facto, only after the intervention ended. This suggests that nonadherence to physical activity, often associated with a more sedentary lifestyle, may reflect an underlying behavioral predisposition that persists even when individuals voluntarily enroll in an exercise program.

Several limitations to our study deserve mention. First, the rate of adherence to the exercise program was accurately measured via the smartphone among the interventional group, while we had to rely upon self-reported data among the controls. Attempts to categorize the active control group into adherers and nonadherers based on self-reports, using the same cutoff points as described for the personalized group, showed no clear trends. To address this limitation, we compared the overall percentage of improvers in the intervention group to those in the active control group over the 8-week program, irrespective of the degree of adherence. Compared to the controls, the intervention participants showed higher rates of improvement for static balance (46% vs 33.3%, P=.07), dynamic balance (53.6% vs. 28.4%, P=.002), arm flexion (52.8% vs 43.4%, P=.48), and arm extension (46.1% vs 34.6%, P=.30).

Second, in the current study, in order to generate accurate data for the study, the trained study personnel performed the testing of participants. However, the ultimate goal of the application is that participants will be able to assess themselves (or possibly with the assistance of a lay person, eg, family member).

Another limitation is the sample’s bias toward participants from higher socioeconomic status. As this is a randomized controlled study, this bias does not affect the main outcome—the improvement of the intervention group over the controls. However, it may impact the generalizability of the study results to other populations.

In addition, in the current study, we applied only 2 levels of difficulty for each of the 3 movement components. As large-scale data accumulate, there is potential to expand and further define more variations of movement profiles.

Furthermore, physical fitness may include more components than the ones included in our study. For example, for assessing strength, while the selected tests did represent the major muscle groups, nonetheless, not all muscle groups were included. We also did not include other components such as eye–hand coordination or fine motor abilities. However, we selected the most appropriate tests for assessing motor and physical fitness in old age, taking into consideration medical aspects and safety precautions (for more information on designing the tests, see [14]).

The interaction with the technology—the smartphone—may also be problematic. For example, using the smartphone screen was inconvenient for some participants, and larger screens would improve the ease of performing the exercises. This could be solved by connecting the smartphone to a large screen. In addition, the videoed exercise program was running in a continuous manner on the phone screen; however, if the participant wanted to repeat an exercise before moving to the next one, or to stop for some reason, he or she had to change his or her current position, approach the phone, stop the program, and restart it again.

It may be argued against the smartphone sensors that there are settings where more accurate assessments may give more accurate results. Although applicable for research, these settings are inaccessible to the general population as they are scarce and expensive to use. Prior work [16,49] has already demonstrated that smartphones offer a reliable tracking of motor functions and therefore offer a sufficient, even if not perfect, means for evaluation in large scales.

The study sample was generally healthy at the time of enrollment, as assessed by the study physician. Enrolled participants were cognitively intact, free from any hospitalization in the last year for any cardiac or neurological reason, independent in daily function, able to walk without assistance, and with a low fall risk. Furthermore, we included a 41-item frailty assessment (see Table 1), which confirmed very low levels of frailty across all participants. Although comorbidities or other current health conditions were not formally assessed, nonetheless, the study sample’s health profile of high cognitive function, independent functional and mobility status, very low frailty levels, and no recent cardiological or neurological illness for hospitalization was suggestive of an overall health status.

Generating optimal personalized exercise program should rely not only on assessing the level of fitness but also considering basic aspects of the individual health status and behavior patterns. In line with personalized genomics and big data accumulation, a possible future upgraded version of this application might be used in conjunction with data drawn from genetic, molecular, clinical, social, and behavioral domains in order to achieve an optimal prescription of personalized exercise programs.

Furthermore, large-scale data will enable an in-depth analysis of individual movement capacities, a more accurate assessment of a wide range of fitness and mobility parameters, and, subsequently, more specific personalized exercise programs. For instance, when assessing balance in individuals with stability issues, it is beneficial to analyze the specific directions of sway—mediolateral, anteroposterior, and superior-inferior. Similarly, an arm strength test can be broken down into separate actions of “lifting the hand up” and “lowering it down,” while the sit-to-stand test can be divided into “getting up” and “sitting down.” Accurately localizing movement impairments has important clinical implications.

The potential ramifications of this study are multifold. First, we provide a model whereby a home-based, highly detailed remote analysis of fitness and mobility parameters provides objective quantitative data as a basis for subsequent tailored exercise interventions at the individual level. Furthermore, the sensitivity of the remote sensors was shown to be sufficient to enable an adaptive, precise, and updating exercise program, determined by the participants’ progress over time. In addition, the continually updating database serves to constantly improve the machine learning precision of decisions concerning the selection of appropriate exercises. Second, our study included only healthy, independent adults aged 65 years and over. We provide evidence to support a platform that has the potential to be adapted to meet the specific needs and abilities of various healthy and patient populations. Specifically, among older people, it is plausible that targeted and individualized exercise interventions are likely to play a beneficial role in improving intrinsic capacity and resilience, moderating frailty and deconditioning. Similarly, individualized programs can be further refined to address specific needs of populations with mobility limitations, such as those at risk of falls or confined to wheelchairs, as well as the potential applicability among participants with cognitive impairments.

In this randomized controlled study, we examined a novel home-based approach for personalized exercise programs for older adults, utilizing a simple smartphone that obviates the need for a laboratory or professional intervention. Through smartphone accelerometer and gyroscope sensors, we remotely assessed key components of motor fitness, including balance, flexibility, and strength. Based on these assessments, a machine-learning-generated personalized exercise program, tailored to each individual’s needs, was delivered via video directly on the smartphone. The results demonstrated that a multicomponent personalized exercise program was more effective when compared to active controls who were counseled to perform regular exercise according to WHO guidelines, as well as a passive control group who received no intervention. Furthermore, improvements were achieved with as few as 1.5 tailored exercise sessions per week over 8 weeks and in many fitness measurements, even within 4 weeks. The remote delivery of user-friendly multicomponent exercise programs to older people has the potential for widespread health benefits. Given the potential of this approach to be extended to various groups of older adults, including those with mobility and cognitive impairments, and considering the advantages of cost-free digital technology (assuming widespread smartphone ownership), health care providers may adopt this approach to provide home-based strategies for health promotion.