Free-living wrist and hip accelerometry forecast cognitive decline among older adults without dementia over 1- or 5-years in two distinct observational cohorts

Alzheimer's disease and related major neurocognitive disorders (ADRD) affect over 50 million people worldwide and are projected to grow as the population ages. ADRD disproportionately affects socioeconomically disadvantaged groups and minorities and is associated with worse quality of life and increased mortality, while few FDA-approved treatments exist, underscoring prevention as the mainstay of management. Cognitive trajectories vary widely among older adults and differ across racial/ethnic groups, motivating sensitive forecasters of early decline to trigger monitoring and preventive interventions. There is an acute need for easily deployed, noninvasive clinical tools to identify cognitively intact older adults at risk of cognitive decline. While demographic, clinical, social, and genetic factors can aid forecasting, genetic measures can be more invasive to collect. Wearable sensors enable remote assessment of free-living activity and sleep and are associated with other aging-related outcomes (frailty, disability, social disengagement, mortality), but their relationship to cognitive performance is less studied and prior findings vary across racial/ethnic groups. Prior work has seldom leveraged high-resolution accelerometer data to recognize individual patterns, and clinical translation has been hampered by differences across device locations, protocols, and manufacturers. The objective of this study was to advance forecasting of early cognitive decline among older adults without dementia by discovering prognostic, free-living 24-h accelerometry patterns. We considered a comprehensive set of accelerometry measures and harmonic features.

Prior studies have linked higher free-living activity to better cognitive performance cross-sectionally (e.g., greater active minutes/day associated with better processing speed; steps/day associated with executive function) and longitudinally (higher percent of moderate-to-vigorous physical activity predicting lower risk of impairment and better maintenance of executive function and memory over ~3 years), though effects may differ by race/ethnicity. Social environment measures (network size, social activity, loneliness) and demographics/clinical comorbidities also predict cognitive decline; APOE and other genetic susceptibilities can improve forecasts but are more invasive to collect. Wearables have been used extensively to assess aging outcomes, but few studies have exploited high-resolution accelerometer features or demonstrated applicability across device locations and manufacturers. Compared to Casanova et al. (2020), which used Random Forests with demographic/clinical/genetic predictors to classify cognitive trajectories without accelerometry, the present work evaluates whether accelerometry pattern features add predictive value beyond conventional predictors.

Study design and cohorts: Two non-overlapping observational cohorts of community-dwelling older adults without dementia were analyzed. (1) Hip accelerometry cohort (FACE Aging study): community sample near the University of Chicago geriatrics practice; inclusion: age ≥65, community-dwelling; exclusions included recent hospitalization/surgery, certain medication changes, oral steroids, beta-blockers, marked hyperglycemia, life expectancy <1 year, history of moderate/advanced dementia or MoCA ≤18. Data collected: baseline survey/physical exam; 7-day free-living hip accelerometry; additional metabolic and DEXA measures (not used); 1-year follow-up survey and exam. Final analytic N=115 with complete clinical data and ≥1 valid wear day (≥10 daytime hours). (2) Wrist accelerometry cohort (NSHAP): nationally representative cohort; a random subset in 2010–2011 wore wrist accelerometers; cognitive reassessed in 2015–2016. Final analytic N=584 with complete clinical data and ≥1 valid wear day; performance metrics reported for N=575 after exclusions in some analyses.

Accelerometer protocols: Hip: Actigraph wGT3X+ worn over right anterior hip with elastic belt for 7 days continuously, 30 Hz sampling; data extracted with ActiLife v6.0; no low-frequency extension filter applied. Wrist: ActiWatch Spectrum worn on non-dominant wrist for 72 consecutive hours, 32 Hz sampling; manufacturer preprocessing (Actiware): maximum absolute value per second, summed over each 15-s epoch; non-wear (identified via galvanic heat sensor) excluded; ≥10 h daytime recording defined a valid day; 24-h interval used for feature generation.

Cognitive outcomes: Hip cohort: MoCA at baseline and 1-year follow-up. Wrist cohort: MOCA-SA (18-item) administered in 2010–11 with estimated MoCA (0–30) via linear prediction; repeated in 2015–16. Cognitive change defined as Δ_MoCA = MoCA_follow-up − MoCA_baseline. Participants with Δ<0 were labeled as cognitive decline (Δ−); others as no decline (Δ+). Class balance: hip 67/48 (Δ−/Δ+), change range −8 to 6 over 1 year; wrist 279/296, change range −14.9 to 14.9 over 5 years.

Covariates: Hip: age, gender, race (Black/African American vs Other), education, monthly income categories, Charlson Comorbidity Index (0–30). Wrist: age (centered), gender, race (White, Black, other) and Hispanic ethnicity, modified Charlson Index (0–16) from self-reported conditions.

Signal processing and features: Vector magnitude r(t)=sqrt(x^2+y^2+z^2) computed for hip data; wrist data provided as vector magnitude at 15-s epochs. Data reshaped to D×T matrices (days × samples/day). Using non-overlapping 1-minute windows, Euclidean Norm Minus One (ENMO), Counts Per Minute (CPM), and Vector Magnitude Count (VMC) were derived: ENMO(t)=mean(max(r−1,0)); CPM(t)=H×ENMO(t) with H=1800 (hip) or 4 (wrist); VMC(t)=mean(|r(t+h)−r(t)|). Two categorical activity measures were defined per participant: C4 from CPM_75 and V4 from VMC_75, categorizing cohort-based distributions into quartiles: inactive, moderately active, active, extremely active. From minute-level ENMO(t) and VMC(t), 98 statistical and harmonic features were extracted, including: mean, median, SD, min/max, quartiles, skewness, kurtosis, entropy, Beta distribution shape (α,β); FFT-based features (top 15 coefficients, FFT entropy); periodogram frequency mean, SD, kurtosis, skewness; RMS amplitude. Total features: 105 (hip) and 104 (wrist); difference due to cohort-specific demographic categories.

Modeling: A binary classifier (CDPred) using XGBoost distinguished Δ− from Δ+. Three models per cohort: (1) CDPred: demographics and clinical covariates only; (2) CDPred-4: CDPred plus C4 and V4; (3) CDPred-4+: CDPred-4 plus all 98 accelerometry harmonic/statistical features. Hold-out sets: 10% (hip) and 15% (wrist). Hyperparameters tuned via 5-fold cross-validation to maximize AUC. Performance reported as accuracy and AUC on hold-out samples; feature importance reported for best-performing models. Ethics approvals obtained (University of Chicago IRB #13-0443); informed consent collected.

• Model performance (hold-out):
  • Hip device (1-year prediction): CDPred accuracy 0.75, AUC 0.74 (0.11); CDPred-4 accuracy 0.78, AUC 0.75 (0.10); CDPred-4+ accuracy 0.84, AUC 0.86 (0.11).
  • Wrist device (5-year prediction): CDPred accuracy 0.66, AUC 0.65 (0.05); CDPred-4 accuracy 0.67, AUC 0.66 (0.05); CDPred-4+ accuracy 0.69, AUC 0.73 (0.05).
• Adding accelerometry features (C4/V4 and 98 harmonic/statistical features) improved prediction over demographics/clinical covariates alone in both cohorts; largest gains observed with CDPred-4+.
• Feature importance: Many accelerometry pattern features ranked above demographic and clinical variables (including age) in importance for predicting decline.
• Effect sizes (selected): Age and baseline MoCA were associated with decline in both cohorts (e.g., Cohen’s D for age: hip 0.457 [0.103, 0.801]; wrist 0.284 [0.119, 0.449]; baseline MoCA: hip 0.572 [0.216, 0.928]; wrist 0.418 [0.252, 0.584]). Gender showed a higher odds of decline in the hip cohort (OR 3.533 [1.223, 10.207]) but not in the wrist cohort (OR 0.865 [0.620, 1.207]).
• Robustness: Predictive improvements were observed despite differences in device location (hip vs wrist), wear durations (7 days vs 72 h), follow-up intervals (1 vs 5 years), and device manufacturers.
• Sample characteristics: Hip cohort N=115 (mean age 73.2; 80.9% female; 81.7% Black); Wrist cohort N=575 (mean age 67.0; 59.1% female; 73.6% White).

Incorporating free-living accelerometry patterns substantially improved forecasting of preclinical cognitive decline compared with models using only demographics and clinical comorbidities. The best-performing models (CDPred-4+) achieved high accuracy for short-term (1-year) prediction with hip-worn devices and moderate accuracy for longer-term (5-year) prediction with wrist-worn devices. Compared to prior machine learning work (e.g., Casanova et al., 2020) that did not include accelerometry, this study shows that accelerometry-derived harmonic and statistical features can outperform many traditional predictors, highlighting the value of noninvasive, remote activity monitoring to augment clinical evaluation. Despite heterogeneity in device location, protocols, and follow-up, accelerometry features remained informative, suggesting promise for broader clinical translation to flag older adults vulnerable to subsequent cognitive decline. Importantly, accelerometry is not diagnostic but reflects patterns of daily movement that correlate with health risk, including cognitive outcomes.

This study demonstrates that 24-h free-living accelerometry, summarized via comprehensive statistical and harmonic features, can forecast preclinical cognitive decline among older adults without dementia, outperforming models based solely on demographics and comorbidities. Hip-worn 7-day data predicted 1-year decline with high accuracy, while wrist-worn 72-h data predicted 5-year decline with moderate accuracy, indicating feasibility across device locations and protocols. These findings support integrating accelerometry-based risk stratification into clinical workflows for aging populations. Future work should aim to further improve performance and generalizability by incorporating additional accelerometry-derived metrics, complementary physiological sensors (e.g., heart rate), richer clinical data (e.g., blood or genetic markers, family history, medications), and harmonized protocols across devices and body locations, as well as evaluating subgroup performance across sociodemographic strata.

• Potential participant overlap between cohorts could not be definitively ruled out, though any overlap is likely minimal given sampling frames (local vs nationally representative) and lack of address data in one cohort.
• Limited sample sizes constrained evaluation of model performance across sociodemographic subgroups; effect sizes by race/ethnicity are provided as a preliminary step.
• Overall predictive accuracy (approximately 70–84%) leaves room for improvement; performance was lower for wrist-based, shorter-duration data with longer follow-up, potentially due to increased motion noise and protocol differences.
• Differences in device location, wear duration, and manufacturer pose challenges for standardization and clinical translation, though open-source processing and cross-protocol feature engineering were used to mitigate these issues.