Deep Learning Approach for Imputation of Missing Values in Actigraphy Data: Algorithm Development Study.

Area of Science:

• Biomedical Engineering
• Data Science
• Physical Activity Research

Background:

• Actigraphy devices provide objective physical activity data, but missing values are common.
• Traditional imputation methods rely on statistical assumptions, potentially limiting accuracy.
• Deep learning offers a data-driven approach to imputation without prior assumptions.

Purpose of the Study:

• To develop and evaluate a deep learning model for imputing missing values in actigraphy data.
• To compare the performance of the deep learning imputation model against established statistical methods.

Main Methods:

• A denoising convolutional autoencoder was employed to build the deep learning imputation model.
• The model was trained and validated using large-scale datasets (NHANES, KNHANES, KChronic).
• Performance was assessed by comparing partial Root Mean Square Error (RMSE) and partial Mean Absolute Error (MAE) against mean imputation, zero-inflated Poisson regression, and Bayesian regression.

Main Results:

• The zero-inflated denoising convolutional autoencoder achieved lower partial RMSE (839.3) and partial MAE (431.1) compared to other methods.
• Mean imputation yielded partial RMSE of 1053.2 and partial MAE of 545.4.
• Zero-inflated Poisson regression and Bayesian regression also showed higher error metrics than the deep learning approach.

Conclusions:

• The developed deep learning imputation model significantly outperforms traditional methods for handling missing actigraphy data.
• This advanced imputation technique improves the reliability and completeness of physical activity data derived from actigraphy.