Binned Data Provide Better Imputation of Missing Time Series Data from Wearables

Area of Science:

• Data Science
• Machine Learning
• Health Informatics

Background:

• Missing values in time-series data are a prevalent challenge.
• Existing imputation methods show variable performance dependent on data characteristics.
• Wearable health trackers generate time-series data prone to missing values.

Purpose of the Study:

• To evaluate the effectiveness of data binning for time-series imputation.
• To compare various imputation algorithms using binned versus whole datasets.
• To identify optimal bin sizes for imputing missing data from wearable health trackers.

Main Methods:

• Imputation of missing data using Expectation Maximization, K-Nearest Neighbor, Iterative Imputer, Random Forest, and Simple Imputer.
• Application of data binning strategies with varying bin sizes (15 min, 30 min, 45 min, 1 h).
• Evaluation of imputation performance using Root Mean Square Error (RMSE) on simulated missing data (15 min and 1 h).

Main Results:

• Data binning significantly improved imputation accuracy compared to using the whole dataset.
• Expectation Maximization demonstrated the best performance with binned data.
• Smaller bin sizes (15 min, 1 h) generally resulted in lower RMSE values.
• Random Forest performance was unaffected by data binning.

Conclusions:

• Data binning is a valuable technique for enhancing time-series imputation, particularly for health tracker data.
• Expectation Maximization combined with data binning offers a robust imputation strategy.
• The proposed method shows potential applicability beyond digital health domains.