ウェアラブルデバイスからのアクティビティデータを合成するための階層的加熱マルコフモデリング

Wearable devices enable continuous acquisition of physiological and behavioral signals with broad utility in precision and population health. However, real-world datasets often have privacy constraints, irregular sampling, and activity-dependent recording, complicating data sharing and modeling. We present a synthesis framework that preserves statistical realism by matching marginal distributions and inter-record distances.

We propose a hierarchical heated Markov model (HHMM) that captures conditional dependencies and time-varying behavioral patterns. Tier 1 generates minute-level activity types using hour-specific priors and a heating mechanism to achieve a target ergodic distribution. Tier 2 samples heart rate (via Metropolis-Hastings conditioned on activity) and activity duration from empirical conditional distributions. Tier 3 synthesizes additional variables (e.g., sleep duration, calories) conditioned on Tiers 1-2, with deterministic rules for activity codes and floors climbed. Activity-conditioned Poisson subsampling emulates device-driven irregular timestamps. We benchmark HHMM against CTGAN and TVAE using a synthetic IEEE BHI dataset and validate on a real Fitbit dataset. Fidelity is assessed by comparing distributions of pairwise inter-record distances-within synthetic vs. within reference cohorts-via Wasserstein distances for Kolmogorov-Smirnov (KS), Jensen-Shannon (JSD), and distance-correlation (DC) metrics.

On the BHI dataset, mean WD(KS)/WD(JSD)/WD(DC) were 0.125/0.242/0.327 for HHMM, 0.130/0.245/0.257 for CTGAN, and 0.129/0.246/0.249 for TVAE. On the Fitbit dataset, values were 0.295/0.489/0.233 (HHMM), 0.293/0.488/0.209 (CTGAN), and 0.293/0.588/0.160 (TVAE).

HHMM offers task-specific gains on synthetic benchmarks. Real-world results highlight a need for domain adaptation. The method is computationally efficient and privacy-preserving, supporting scalable synthetic data generation for wearable health research.