Age and environment-related differences in gait in healthy adults using wearables

Gait speed is widely recognized as an informative clinical measure, often termed the sixth vital sign, and is associated with outcomes such as cognitive decline, falls, and mortality. Traditional, episodic in-lab or clinic-based gait assessments can be subjective and may not reflect real-world performance due to observer effects. Wearable inertial sensors enable continuous, free-living measurement of gait, but their validity and sensitivity can depend on population and sensor placement. There is growing evidence that clinic-based gait measures differ from free-living measures, with real-world gait often slower and more variable, potentially offering richer clinical insights. An important unresolved question is the minimal at-home monitoring duration needed to reliably estimate gait speed, with prior work suggesting 3–6 days for various physiological measures and only limited evidence for gait speed. This study investigates healthy younger (18–40) and older (65–85) adults to: (1) validate a single lumbar-worn accelerometer method (GaitPy) against a six-sensor system (APDM) and an instrumented mat (GAITRite); (2) test whether in-lab versus at-home gait speed (median and 95th percentile) differentiates age groups; and (3) determine the minimal at-home monitoring needed for reliable gait speed estimation.

Prior studies validate wearable inertial sensors against ground-truth systems (instrumented mats, motion capture) but note reliability can vary by population and sensor location. Free-living gait assessments often reveal slower speeds than clinic measurements in older or frail adults, and free-living gait features can better distinguish disease states (e.g., Parkinson’s disease) compared to controls. Regulatory interest is increasing in at-home digital endpoints (e.g., EMA qualification of 95th percentile stride velocity in DMD). Evidence on required monitoring duration is mixed: studies across activity measures suggest 3–6 days for reliable estimates; a recent gait study in slow-walking older adults with sarcopenia found 3 days needed for gait speed. Open questions include standardized processing of at-home data, interpretation of day-to-day variability, and the optimal minimal acquisition period for gait speed in different populations.

Design: Cross-sectional study with two in-lab visits (~2 hours each, 7–14 days apart) and continuous at-home monitoring between visits (~6–15 days). Participants: 65 healthy adults recruited at Pfizer Innovation Research Lab (PfIRe Lab), Cambridge, MA, USA; 33 younger (29.2±4.6 years; 17F; BMI 23.4±2.6) and 32 older (72.3±5.8 years; 16F; BMI 24.5±2.6). Inclusion criteria: no significant health problems (physician-reviewed), BMI ≥18.5 and <30 kg/m² or weight <125 kg, and predefined VES-13 scores. Ethics: Advarra IRB (Pro00029419); informed consent obtained. In-lab instrumentation and tasks: Participants wore six inertial sensors (Opal, APDM; accelerometer, gyroscope, magnetometer; 128 Hz) on sternum, lumbar (L4), and bilaterally on wrists and feet. Gait task: three laps on GAITRite instrumented mat; standard SPPB timing with stopwatch also collected. Gait speed was computed via: (1) GAITRite proprietary algorithm (ground truth), (2) APDM Mobility Lab (six-sensor set), (3) GaitPy (single lumbar sensor; open-source, Python v3.6). GaitPy algorithm: wavelet-based detection of heel strike and toe-off from vertical acceleration; vertical displacement via integration; inverted pendulum model for speed. At-home monitoring: GENEActiv device worn on lower back (and wrist, but only lumbar data used), 3-axis accelerometer at 50 Hz, continuous wear for ~7–14 days (mean 8.72±1.88). GaitPy detected walking bouts using a binary classifier; bouts <3 s apart concatenated; gait speed estimated stride-by-stride. Exclusions: bouts <10 s or >3000 s and bouts with <4 gait cycles. Statistical analysis: Per in-lab visit, median of gait metrics across steps per lap, then across laps. Agreement analyses: Bland–Altman plots; ICC2,1 (two-way random effects, absolute agreement) for device agreement; Pearson’s r. Group analyses: linear mixed-effects models (R 3.5.2; lme4, car) with repeated measures. In-lab model fixed effects: method (GAITRite/APDM/GaitPy), age group, sex, height, muscle mass; random effects: participant/visit and participant/device. At-home models: for median gait speed, fixed effects age group, sex, day type (weekday/weekend), covariates height and muscle mass; random effects participant/type of day/day. For 95th percentile gait speed, summarized per participant; same fixed effects excluding day type and its random effects. Multiple comparisons corrected by FDR. In-lab vs at-home association: linear regression predicting in-lab gait speed from at-home median or 95th percentile; Spearman’s rho reported. Reliability and minimal data: Bootstrapping with replacement. Steps-based reliability: participants with ≥25,000 steps (n=62); subsets of 5,000–25,000 steps; compute median and 95th percentile gait speed per subset; ICC vs full data (all available days). Days-based reliability: participants with ≥5 days (n=65); subsets of 1–5 days; ICC vs full at-home data. ICC thresholds: ≤0.40 poor, 0.40–0.59 moderate, 0.60–0.74 good, ≥0.75 excellent. Group distinguishability vs duration: bootstrapped t-statistics (1,000 samples per duration) for age-group contrast compared to full-data t-statistic; p-value as proportion of t-bootstrap > t-original. Minimum daily steps threshold for inclusion set at ≥100 steps/day; sensitivity to thresholds (10, 250, 1000) assessed and found not impactful.

• Validation against GAITRite: Both APDM (six sensors) and GaitPy (single lumbar sensor) under-estimated in-lab gait speed relative to GAITRite with consistent bias. Mean biases: GAITRite − APDM = 0.07 m/s (~5%); GAITRite − GaitPy = 0.17 m/s (~13%). Correlations with GAITRite were high/moderate (APDM r=0.98; GaitPy r=0.72). ICC across methods showed moderate agreement (overall ICC=0.66 [0.27–0.83]); GAITRite vs GaitPy ICC=0.49 (−0.07–0.77), improving to 0.72 (0.63–0.79) after mean-bias correction.
• In-lab age-group comparison: No main effect of age group on gait speed (χ²=0.28, p=0.6) using GAITRite, APDM, or GaitPy; methods differed significantly (χ²=199, p<1e−16). Pairwise: APDM and GaitPy < GAITRite (all p<1e−10). Trending visit effect (χ²=3.79, p=0.051); significant age group by sex interaction (χ²=5.43, p=0.02) and age group by sex by method (χ²=14.67, p<1e−3).
• At-home age-group differences: Significant differences between younger and older groups for both median gait speed (χ²=12.54, p=0.006) and 95th percentile gait speed (F=22.59, p=1e−5); older adults walked slower. Day type effect for median speed (weekday/weekend χ²=42.08, p<1e−5) with group×day-type interaction (χ²=13.38, p=0.002): significant group differences on weekdays (χ²=21.81, p<1e−5) but not weekends (χ²=3.23, p=0.33). No group difference in number of walking bouts (p=0.8).
• In-lab vs at-home association: Weak-to-moderate relationships. At-home median predicting in-lab: adjusted R²=0.18, F(1,62)=14.77, β=0.57, p<1e−3; Spearman’s rho=0.35 (p=0.004). At-home 95th percentile predicting in-lab: adjusted R²=0.25, F(1,62)=21.45, β=0.47, p<1e−5; Spearman’s rho=0.42 (p=0.0005). Intercepts significant in both models (approx. 0.65 and 0.54 m/s respectively).
• Minimal data for reliable at-home gait speed: Days-based ICC vs full data showed excellent reliability with ≥2 days for both median (median ICC=0.85 [0.65–0.93]) and 95th percentile (median ICC=0.89 [0.66–0.95]); improvements minimal beyond 3 days. Steps-based reliability: good agreement at ~5,000 steps, excellent at ≥15,000 steps for median and ≥10,000 steps for 95th percentile, with less variance for 95th percentile. Group distinguishability bootstrapping: median gait speed required ≥2 days to match full-data distinguishability (p≥0.05), whereas 95th percentile required only 1 day (p>0.05).
• Additional observations: At-home gait speed was generally lower than in-lab for both groups; in-lab stopwatch 4-m times showed no age-group differences (younger 3.8±0.5 s; older 3.7±0.5 s; χ²=1.08, p=0.3). Other in-lab gait metrics (e.g., step time, stride length) did not differ by age. Estimated average at-home median gait speed difference between groups was ~0.09 m/s.

Findings address the primary questions: a single lumbar-worn accelerometer (GaitPy) provides acceptable accuracy versus gold-standard GAITRite with consistent, correctable bias; in-lab gait speed does not differentiate age groups in healthy adults, whereas at-home measures (median and 95th percentile) do, particularly on weekdays; and only 2–3 days of at-home monitoring suffice to reliably estimate gait speed and capture age-related differences. The weak association between at-home and in-lab gait speeds indicates they capture distinct constructs—real-world mobility likely reflects cognitive load, mood, fatigue, and environmental context absent in laboratory tasks. The results support prioritizing free-living, sensor-based endpoints in clinical research for greater ecological validity and sensitivity to subtle functional differences, enabling decentralized trial designs and reduced participant burden. The trade-off in sensitivity from single-sensor setups appears minimal given the practicality and participant compliance benefits, though multi-sensor or inclusion of gyroscopes could improve characterization of complex movements (e.g., turns).

This study demonstrates that a single lumbar-worn wearable sensor and open-source GaitPy algorithm can validly estimate gait speed relative to a gold-standard instrumented mat. In healthy adults, at-home gait speed (median and 95th percentile) discriminates between younger and older groups, whereas in-lab gait speed does not, and at-home and in-lab measures show only weak-to-moderate association. Reliable estimation of at-home gait speed requires only 2–3 days of monitoring. These findings advocate for using free-living gait speed as a sensitive, ecologically valid endpoint in clinical trials and suggest that short monitoring periods can reduce burden while maintaining reliability. Future work should validate performance in diverse and advanced disease populations, examine disease-specific variability (e.g., motor fluctuations), optimize algorithms for single-sensor configurations, and assess longer monitoring to further characterize day-type effects and compliance.

• Monitoring duration averaged ~9 days; reliability analyses compared 1–5 randomly selected days to a full dataset averaging 9 days, which may limit generalizability to longer baselines.
• Participant compliance for continuous wear with a lumbar-worn accelerometer could not be robustly verified; inclusion threshold set at ≥100 steps/day, though sensitivity analyses with different thresholds showed minimal impact.
• Day-type effects (weekday vs weekend) influenced gait speed, but bootstrapping for minimal days did not stratify by day-type, potentially affecting reliability estimates.
• Single-sensor approach exhibited higher variability than multi-sensor APDM; though biases were consistent, sensitivity trade-offs may exist.
• Study population included only healthy adults; algorithm performance and minimal monitoring requirements may differ in advanced or specific disease populations.
• In-lab tasks were short and may not represent longer continuous walking, which could relate differently to at-home performance.