This was a cross-sectional study.

We included 587 people from the cohort of the Otassha study, a longitudinal study that has been ongoing since 2011 [30], who underwent an invitation-based health checkup and completed sarcopenia, frailty, and sit-to-stand measurements from September 2023 to October 2023. All study participants lived in the community and were aged over 65 years. Participants living in care facilities, as well as individuals with contraindications for bioelectrical impedance analysis, were excluded from the study. Participants underwent comprehensive geriatric health examinations regarding their physical, cognitive, psychological, nutritional, and lifestyle habits.

Body composition was measured using a bioimpedance body composition meter (InBody 770; InBody Co, Ltd) to estimate appendicular skeletal muscle mass. Although dual-energy x-ray absorptiometry is recognized as the gold standard for measuring body composition, bioimpedance analysis has been evaluated against it and verified to provide correct measurements [31].

The participants were instructed to stand still with their arms at their sides while using the bioimpedance equipment without talking and to hold a sensor for measurement. The skeletal muscle index (SMI) was calculated as the appendicular skeletal mass, estimated from the body composition measurement, divided by the square of the individual’s height. Smedley hand dynamometers were used to measure grip strength.

For grip strength measurements, participants were asked to extend their arms straight down and grip the dynamometer as strongly as possible with their preferred hand when the tester instructed them to do so.

Walking speed was measured on a 5-m course with a preceding 3-m acceleration and subsequent 3-m deceleration walkway. Participants were asked to walk the course once at a comfortable speed. The time taken to walk through the 5-m section of the course was measured using a digital stopwatch.

A smartphone with our app installed was secured to the lower abdomen of each participant using a commercially available sports smartphone holder (Figure 1). Participants folded their arms across their chests to minimize arm motion and trunk variability. Participants received auditory prompts from a Bluetooth speaker positioned 2 m in front of them via the smartphone app, which instructed them to stand up as quickly as possible and then hold the position for 3 seconds. After the practice attempt, the participants were evaluated twice, transitioning rapidly from a seated resting position (seat height; 42 cm) to a standing position. Acceleration during this action was captured using the smartphone app.

The app is web based and uses JavaScript to call the operating system application programming interface and measure the smartphone’s acceleration on 3 axes over a given period. To validate the accuracy of the smartphone accelerometer for use in biomechanical research, 3 healthy volunteers performed sit-to-stand trials at 3 different speeds (fast, moderately fast, and normal) to simulate a range of physical performance levels (9 trials in total). A research-grade reference accelerometer (MVP-RF8-BC; MicroStone Corp; sampling at 1000 Hz) was attached to the smartphone holder on the participants’ lower abdomen to measure acceleration simultaneously with the smartphone. The resultant acceleration signals from both devices were synchronized using cross-correlation to align the waveforms temporally. We then extracted the specific “movement phase” (from the start of the rise to stabilization) and assessed the waveform similarity using the intraclass correlation coefficient (ICC(2,1); absolute agreement). The analysis demonstrated excellent agreement between the smartphone and the reference device despite the difference in sampling rates. The mean ICC(2,1) was 0.982 (SD 0.019), with a range of 0.932 to 0.995 across all trials and speeds, confirming that the smartphone’s accelerometer measurements were reliably comparable to those of the dedicated biomechanical research device.

The smartphone app recorded the sit-to-stand motion at 60 Hz, and this motion was analyzed in the cloud. Data were processed using an inverse fast Fourier transform algorithm with a low-pass filter at 7.5 Hz to mitigate the impact of noise outside the physical activity domain. In traditional settings, the maximum force exerted during sit-to-stand motion is measured using a floor reaction force meter with a fixed coordinate system established upon installation. However, when using a smartphone accelerometer attached to the body, the coordinate system is affected by the user’s movements. The norm of the triaxial acceleration was computed to account for these variations and ensure that the measurement was less affected by changes in the orientation of the accelerometer. The peak force exerted during the sit-to-stand motion was calculated by multiplying the resultant peak acceleration by the body weight.

The onset of standing was identified as the point at which the accelerometer registered the lowest acceleration from the start of the motion (Figure 2). Peak acceleration (measured in m/s²) was defined as the highest acceleration recorded after starting. A stable stance was considered when the acceleration variation decreased to 1% or less of the gravitational acceleration after peak acceleration. The duration from the start to peak acceleration was termed the “rising time” (T1; seconds), whereas the duration from peak acceleration to attainment of a stable stance was termed the “stability time” (T2; seconds). The peak force (in newtons) was calculated by multiplying peak acceleration by the weight of each participant.

Sarcopenia was diagnosed based on the criteria outlined in the 2019 consensus statement of the Asian Working Group for Sarcopenia [32]. Grip strength below 28 kg for men and below 18 kg for women, walking speed below 1.0 m/s, and SMI below 7.0 kg/m² for men and below 5.7 kg/m² for women were the thresholds used to diagnose sarcopenia. Participants who met the criteria for SMI and grip strength or walking speed were classified as having sarcopenia. However, sarcopenia severity was not classified, similar to our previous study [22].

Frailty was diagnosed using the criteria outlined in the Japanese version of the Cardiovascular Health Study [14]. The criteria were an adaptation of the original Fried frailty index [33] tailored to the Japanese population. A person was classified as frail if they met 3 or more of the following criteria: an unintentional loss of 2 kg or more of body weight in the previous 6 months, handgrip strength of less than 28 kg for men and less than 18 kg for women, a feeling of exhaustion, a walking speed slower than 1.0 m/s, and a lack of engagement with regular physical activity.

To confirm the reproducibility of the smartphone-derived measurements (peak force, T1, and T2), we calculated the ICC(2,1) using a 2-way mixed-effects, absolute-agreement model based on 2 consecutive sit-to-stand trials. For all subsequent analyses, the average of the 2 measurements was used.

To identify and classify sarcopenia and frailty, we developed predictive models following TRIPOD (Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis) recommendations. Five logistic regression models were constructed for each outcome to enable direct comparison with established clinical indicators:

All models were implemented as elastic net penalized logistic regression (penalty=“elasticnet”; l1_ratio=0.5; solver=saga) to mitigate overfitting and address multicollinearity among biomechanical and functional predictors. Continuous predictors were standardized using z-score transformation.

Internal validation was conducted using bootstrap resampling (500 iterations). For each model, we calculated the following performance metrics: apparent area under the receiver operating characteristic curve (AUC), bootstrap AUC (mean), bootstrap 95% CI, optimism-corrected AUC, corrected AUC 95% CI, calibration slope, calibration intercept, sensitivity, and specificity.

Calibration was assessed using a logistic recalibration model, from which the calibration slope and intercept were derived. We reported apparent and bootstrap optimism-corrected calibration metrics (500 resamples; percentile 95% CIs). Calibration slopes close to 1.0 and intercepts near 0 indicated good agreement between predicted and observed risk.

To contextualize the performance of the smartphone-based model, we additionally compared it against established screening measures—walking speed, grip strength, and age and sex—using identical model structures and validation procedures.

All statistical analyses and model development were conducted using Python (version 3.11; Python Software Foundation), including scikit-learn (Google Summer of Code project; logistic regression and scaling), NumPy, and pandas. The reproducibility analysis (ICC) and descriptive statistics were performed and calculated using SPSS (version 27; IBM Corp).

This study was conducted as part of the aforementioned comprehensive cohort study. The research plan was discussed and approved by the ethics committee of the Tokyo Metropolitan Institute for Geriatrics and Gerontology (R22-034). Written informed consent was obtained from all participants. Strict confidentiality procedures were implemented. All analyses were performed using only anonymized datasets. Participants did not receive any incentives.

A total of 587 older adults participated in the assessment, of whom 569 (96.9%) completed the smartphone-based sit-to-stand measurement (Figure 3).

The participants’ characteristics are listed in Table 1. Among them, 95 participants (n=25, 26.3% men and n=70, 73.7% women) met the criteria for sarcopenia, and 51 participants (n=22, 43.1% men and n=29, 56.9% women) met the criteria for frailty. Compared with individuals without sarcopenia, those with sarcopenia were older (mean age 77, SD 7 years vs 73, SD 6 years; P<.001) and shorter and had less weight, weaker grip strength (mean 16, SD 5 kgf vs 26, SD 8 kgf; P<.001), slower walking speed (mean 1.25, SD 0.22 m/s vs 1.40, SD 0.23 m/s; P<.001), and lower peak force (mean 70.87, SD 14.98 N vs 92.56, SD 25.20 N; P<.001). Similarly, individuals with frailty were older (mean age 77, SD 8 years vs 73, SD 6 years; P=.003) and had lower grip strength (mean 19, SD 6 kgf vs 24, SD 8 kgf; P<.001), slower walking speed (mean 1.13, SD 0.25 m/s vs 1.40, SD 0.23 m/s; P<.001), and lower peak force (mean 78.56, SD 20.49 N vs 89.93, SD 25.25 N; P<.001) than individuals without frailty (Table 2).

The reproducibility analysis based on 2 consecutive sit-to-stand trials demonstrated excellent reliability for peak force (ICC(2,1)=0.863). Rising time (T1; ICC=0.270) and stabilization time (T2; ICC=0.278) showed lower reproducibility. Therefore, the mean of the 2 measurements was used for subsequent analyses.

The distribution of the 5 frailty phenotype components is shown in Table 3.

Participants classified as frail exhibited a substantially higher prevalence of each phenotype component compared with individuals without frailty. Specifically, participants with frailty demonstrated markedly higher proportions of slow walking speed (16/51, 31.4% vs 12/536, 2.2%), weight loss (33/51, 64.7% vs 59/536, 11%), exhaustion (46/51, 90.2% vs 127/536, 23.7%), low physical activity (25/51, 49% vs 77/536, 14.4%), and weakness (44/51, 86.3% vs 133/536, 24.8%). These differences are consistent with the Japanese Cardiovascular Health Study frailty phenotype criteria and confirm that the frail group exhibited a substantially higher burden of functional and symptomatic deficits.

Model performance for sarcopenia is summarized in Table 4. Among the benchmark models, the grip strength model demonstrated the highest discrimination, with a bootstrap AUC of 0.847 (95% CI 0.797‐0.893). The smartphone model (age, sex, peak force, T1, and T2) showed good predictive ability, achieving a bootstrap AUC of 0.800 (95% CI 0.754‐0.848; Figure 4). After internal validation using 500 bootstrap samples, the model achieved an optimism-corrected AUC of 0.781 (95% CI 0.733‐0.826). This performance exceeded that of the walking speed (corrected AUC=0.663) and age and sex (corrected AUC=0.656) models. Calibration of the smartphone model was modest (slope=1.04; intercept=0.06). The results for the smartphone reduced model were similar to those for the full model.

Model performance for frailty is shown in Table 5.

Walking speed demonstrated the highest discrimination (AUC=0.750, 95% CI 0.672‐0.826), followed by grip strength (AUC=0.704, 95% CI 0.615‐0.781). The smartphone model achieved a bootstrap AUC of 0.720 (95% CI 0.640‐0.796; Figure 5). The optimism-corrected AUC was 0.659 (95% CI 0.587‐0.736). This indicates moderate discrimination, outperforming the age and sex model (corrected AUC=0.604). Calibration of the smartphone model remained modest (slope=1.09; intercept=0.19). The smartphone reduced model had a slightly lower AUC than the full model.

Across outcomes, the smartphone-derived sit-to-stand measurements consistently outperformed demographic models and walking speed for sarcopenia and provided moderate predictive value for frailty. Grip strength and walking speed remained the strongest predictors of sarcopenia and frailty, respectively, whereas the smartphone model functioned as a strong secondary predictor of sarcopenia and a useful complementary indicator of frailty.