Hierarchical heated markov modeling for synthesizing activity data from wearable devices

Area of Science:

• Biomedical Informatics
• Data Science
• Wearable Technology

Background:

• Wearable devices collect continuous physiological and behavioral data for health applications.
• Real-world data presents challenges like privacy concerns, irregular sampling, and activity-dependent recording, hindering data sharing and analysis.
• A synthesis framework is needed to generate realistic, privacy-preserving data for wearable health research.

Purpose of the Study:

• To develop a novel synthesis framework for generating statistically realistic and privacy-preserving synthetic data from wearable devices.
• To address challenges in real-world wearable data, including irregular sampling and activity-dependent recording.
• To enable scalable data generation for precision and population health studies.

Main Methods:

• A Hierarchical Heated Markov Model (HHMM) was proposed, incorporating conditional dependencies and time-varying behavioral patterns.
• The model utilizes a multi-tiered approach: Tier 1 models activity types, Tier 2 samples heart rate and activity duration, and Tier 3 synthesizes additional health variables.
• Activity-conditioned Poisson subsampling was employed to emulate device-driven irregular timestamps, and the model was benchmarked against CTGAN and TVAE.

Main Results:

• On a synthetic dataset, HHMM showed comparable performance to CTGAN and TVAE in preserving inter-record distance distributions (WD(KS): 0.125 vs. 0.130/0.129).
• On a real Fitbit dataset, HHMM demonstrated competitive results, though domain adaptation is suggested (WD(KS): 0.295 vs. 0.293/0.293).
• The HHMM method proved computationally efficient and effective in generating privacy-preserving synthetic data.

Conclusions:

• The HHMM framework offers a viable solution for generating realistic synthetic wearable health data, preserving key statistical properties.
• While effective on synthetic benchmarks, real-world applications may benefit from domain adaptation strategies.
• The method supports scalable, privacy-preserving synthetic data generation, crucial for advancing wearable health research and applications.