Pattern-based clustering of daily weigh-in trajectories using dynamic time warping

Area of Science:

• Behavioral Science
• Data Science
• Health Informatics

Background:

• Smart scales offer novel opportunities for frequent weight monitoring and behavioral pattern analysis.
• Understanding weigh-in frequency patterns can provide insights into weight loss dynamics.
• Previous studies often focused on overall adherence rather than time-invariant patterns.

Purpose of the Study:

• To characterize time-invariant weigh-in patterns using hierarchical clustering with dynamic time warping (DTW).
• To evaluate the performance of DTW against other distance metrics for binary time series clustering.
• To assess the association between identified weigh-in patterns and weight loss outcomes in an adult cohort.

Main Methods:

• Utilized data from an 18-month behavioral weight loss study involving 55 overweight or obese adults.
• Applied hierarchical clustering with dynamic time warping (DTW) to binary time series representing daily weigh-in adherence.
• Conducted simulation studies to compare DTW with Euclidean and Jaccard distances and evaluated cluster validation indices (CVIs) with different linkages.

Main Results:

• Dynamic Time Warping (DTW) demonstrated effective pattern recovery in binary adherence time series compared to Euclidean and Jaccard distances.
• Simulation results informed the application of clustering techniques to real-world weigh-in data.
• Identified distinct weigh-in adherence trajectory patterns among participants.

Conclusions:

• The study successfully characterized individual weigh-in behaviors using DTW-based hierarchical clustering.
• Specific adherence trajectory patterns were found to be significantly associated with overall weight loss.
• Findings highlight the utility of "smart" scales and advanced data analysis for personalized weight management insights.