Question

Wearable-derived activity signatures vs disease activity in elderly rheumatoid arthritis (RA): In adults ≥70 years with RA, how strongly do continuous, real‑world wearable metrics (gait speed, step‑count variability, activity fragmentation, circadian rhythm stability, sleep continuity, HRV/PPG if available) associate with clinical disease activity (DAS28‑CRP, SDAI, RAPID3), and can these metrics anticipate RA flares by ≥7 days?

Answer (protocol‑grade plan + expected benchmarks)

1) What to measure (domains & instruments)

Domain	Instrument / Protocol	Primary metrics	Notes
Ambulation	Wrist/hip accelerometer (24/7, 60–100 Hz; 1‑min epochs)	Daily steps; gait speed proxy (cadence×stride length model); peak 30‑min cadence; sit‑to‑stand transitions	Device‑agnostic feature set; non‑wear auto‑detect
Variability & fragmentation	From minute‑level activity	Coefficient of variation (CV) of steps/day; bout distribution α; transition probability (active↔sedentary); activity fragmentation index	Higher fragmentation = worse endurance/flare risk
Circadian rhythm	Actigraphy	Interdaily stability (IS), intradaily variability (IV), relative amplitude (RA) of rest–activity rhythm	Lower IS / higher IV suggests disruption
Sleep	Actigraphy or PPG	Total sleep time, wake after sleep onset (WASO), sleep efficiency	Morning stiffness proxy: low activity in first 90 min post‑wake
Autonomic	PPG (if available)	Nighttime HRV (RMSSD), resting HR trend	Artifacts filtered; use nightly medians
Clinical activity	Monthly clinic	DAS28‑CRP, SDAI; RAPID3 monthly; daily eDiary: pain (0–10), stiffness minutes	Flare defined a priori (e.g., OMERACT + ΔDAS28≥1.2 or ≥0.6 if high baseline)
Inflammation	Labs	CRP, ESR	Drawn monthly; optional home micro‑sampling
Context	Questionnaires & logs	Medication changes, glucocorticoid bursts, infections, weather/seasonality	Time‑varying covariates

2) Study design

• Design: 6‑month prospective, n = 160 adults ≥70 years with RA (community + clinic). Power allows detection of modest lead‑time effects.
• Visits: Baseline + monthly clinics; continuous wearable data; weekly phone/LLM check‑ins to confirm flare symptoms.
• Exclusions: Recent arthroplasty (<6 months), unstable cardiac/neurologic gait disorders, long non‑wear (>20% days).

3) Feature engineering (daily → weekly)

• Compute daily features, then aggregate to rolling 7‑day windows (median, IQR, slope). Create delta‑features vs each participant’s 28‑day baseline.
• Create morning stiffness index: activity counts in first 90 min after wake normalized by day’s median.
• Derive flare candidate time points using eDiary and ΔDAS28 rules; mark ±21‑day windows for modeling.

4) Statistical analysis plan (SAP)

• Cross‑sectional: Mixed‑effects models (random intercepts by person) for same‑week associations of wearable features with DAS28/SDAI/RAPID3. Report standardized β and marginal/conditional R².
• Lead‑time (prediction): Lagged models with features at t−7 and t−14 predicting flare at t (logistic mixed models). Evaluate AUC/PR‑AUC with nested CV by person.
• Temporal precedence: Granger‑style tests (VAR/LSTM with ablation) to assess whether changes in fragmentation/IS/HRV precede ΔDAS28 beyond pain.
• Time‑to‑event: Andersen–Gill (recurrent events) Cox models with time‑varying wearable covariates for risk of flare onset.
• Added value: Compare Wearables‑only, Symptoms‑only (pain/stiffness), and Combined models; compute ∆AUC, NRI, decision‑curve net benefit.
• Drift control: Include weather, daylight hours, weekend/holiday dummies; adjust for medication changes (DMARD/GC).

5) Benchmarks to claim success (pre‑specified)

Property	Target
Association (same‑week)	|β| ≥ 0.25 for fragmentation or IS/IV vs DAS28; partial R² ≥ 0.07
Prediction (t−7 → flare)	AUC ≥ 0.78 (95% CI lower bound ≥ 0.72); PPV ≥ 0.35 at 20% alert rate
Lead time	Median correct‑alert lead ≥ 7 days
Calibration	Calibration slope 0.9–1.1; Brier score improvement vs symptoms‑only
External validity	Stable performance across device brands and gait‑aid users (∆AUC < 0.03)

6) Power & sample size (sketch)

• Assume ~0.5 flares/person over 6 months ⇒ ~80 flares total. With person‑clustered CV, this yields ~80–90% power to detect AUC 0.78 vs 0.70 baseline at α=0.05.
• For mixed‑effects |β|≈0.25 with ICC≈0.5 and ~24 weekly measures/person, n=150–160 is adequate.

7) Expected findings & interpretation

• Fragmentation rises first: Increases in activity fragmentation and intradaily variability (IV) are expected to precede flares by ~7–10 days, even after adjusting for pain.
• Morning stiffness signal: Lower morning‑activity index and reduced peak 30‑min cadence often herald flares.
• Autonomic drift: Nighttime HRV (RMSSD) dips and resting HR rises 3–7 days before clinical flare in those with PPG.
• Clinical utility: A combined model (wearables + symptoms) supports a low‑burden “digital flare alert,” prompting self‑management or early contact for therapy adjustment.

8) Practical alert policy (example)

• Trigger alert if 2 of 3 conditions hold over a 7‑day window: (a) fragmentation index > 1 SD above personal baseline, (b) morning‑activity index < 0.8× baseline, (c) IV increase > 0.5 SD; and no recent steroid burst.
• Suppress alerts for planned travel/illness weeks and within 7 days after intra‑articular steroid.

9) Pseudocode (model skeleton)

# Weekly dataset with person_id, week, DAS28, flare (0/1), features_{t}, features_{t-1}, covariates
# Fit lagged mixed model for t-1 (≈7 days back) predicting flare at t

from sklearn.linear_model import LogisticRegression
from sklearn.metrics import roc_auc_score
import numpy as np

X = weekly[["frag_t-1","IS_t-1","IV_t-1","morningIdx_t-1","peak30_t-1","HRV_t-1","restHR_t-1","pain_t-1","stiffmin_t-1","weather","weekend"]]
y = weekly["flare_t"]

# Person-level cross-validation
aucs = []
for train_ids, test_ids in person_group_kfold(weekly.person_id, n_splits=5):
    model = LogisticRegression(max_iter=200, penalty="l2", class_weight="balanced")
    model.fit(X.iloc[train_ids], y.iloc[train_ids])
    p = model.predict_proba(X.iloc[test_ids])[:,1]
    aucs.append(roc_auc_score(y.iloc[test_ids], p))

print("AUC:", np.mean(aucs))

Abbreviations — DAS28: Disease Activity Score in 28 joints; SDAI: Simplified Disease Activity Index; RAPID3: Routine Assessment of Patient Index Data 3; IS/IV: interdaily stability / intradaily variability; PPG: photoplethysmography; HRV: heart‑rate variability; RMSSD: root‑mean‑square of successive differences; DMARD: disease‑modifying antirheumatic drug.