Data imputation for accelerometer-measured physical activity: the combined approach

Area of Science:

• Biomedical Engineering
• Physical Activity Measurement
• Data Science

Background:

• Accelerometers are popular for assessing physical activity, but missing data due to non-compliance is a significant issue.
• Traditional imputation methods for missing accelerometer data often discard information from partially invalid days.
• This leads to potential biases and reduced accuracy in physical activity level estimations.

Purpose of the Study:

• To introduce a novel imputation approach for missing accelerometer data.
• This method uniquely incorporates data from invalid wear days to improve accuracy.
• The study aims to address limitations of traditional imputation techniques.

Main Methods:

• A combined imputation approach was developed, integrating data from both valid and invalid accelerometer wear days.
• The method was tested on 4069 participants from NHANES 2003-2004 and 2005-2006 data.
• Simulation studies compared the new approach against traditional methods using root mean squared error (RMSE) for accuracy and effect-size estimation.

Main Results:

• The combined imputation approach demonstrated significantly superior performance compared to traditional methods (P < 0.001).
• RMSE for accuracy was reduced by 12.4%–17.3%.
• RMSE for sex-physical activity effect-size estimation saw reductions of 19.8%–32.9%.

Conclusions:

• The novel combined approach significantly outperforms traditional imputation algorithms for accelerometer data.
• This method offers a more accurate way to handle missing data in physical activity research.
• Utilizing data from invalid days enhances the reliability of accelerometer-based physical activity assessments.